For most people, that means football. Flags. Goals. Bad national anthems sung with full emotional commitment. One uncle who suddenly becomes a tactical expert after two beers.

It began in Uruguay in 1930. Now it arrives as a 48-team, 104-match, three-country spectacle across 16 North American cities.1

In the US, that alone is a strange story.

Not long ago, soccer was mocked as a foreign sport. When America hosted the 1994 World Cup, many still treated it like a temporary curiosity. Roger Bennett wrote in the Wall Street Journal that soccer once ranked below tractor-pulling in a US spectator-sport survey.

So what happened?

How did a sport America used to laugh at become big enough to fill stadiums, streaming platforms, sponsor decks, private-equity memos, and presidential photo ops?

And what exactly are we watching tomorrow: a football tournament, a media product, a political stage, or one of the largest attention machines on earth?

Subscribe now

1) America finally joins the party

The funny part is that the main host country once treated football as a foreign hobby.

When the US hosted the 1994 World Cup, many Americans still saw soccer as strange, slow, and suspiciously European. Roger Bennett wrote in the Wall Street Journal that shortly after he moved from the UK to the US, soccer ranked 67th in a national survey of favourite spectator sports. Tractor-pulling ranked 66th.

That is hard to improve on as a cultural insult.2

But America changed.

The internet made global football easier to follow. Streaming made Premier League matches available every weekend. Video games taught a generation the names of clubs, players, shirts, leagues, and derbies they had never seen in real life. The US women’s national team gave the country winners. Messi in Miami made football feel less foreign. Ted Lasso probably did more for the Premier League than some club marketing departments, which is annoying but true.

Now the tournament arrives in a different America.

The 2026 World Cup has 48 teams and 104 matches. It is the biggest tournament in FIFA history.3

And America is no longer just hosting football. America is starting to reshape it.

2) The money machine

The old World Cup was simple: teams arrived, countries watched, sponsors paid, and FIFA collected.

The new World Cup is more aggressive.

More teams. More matches. More hospitality. More premium tickets. More broadcast inventory. More sponsor slots. More data. More resale fees. More everything.4

Statement of revenue for 2023-2026 (USD million)

The Financial Times reported that FIFA expects close to …:

• $4bn from television rights

• about $1.8bn from sponsorships

• and more than $3bn from ticketing and hospitality

… for this World Cup. Ticketing and hospitality are the big jump: more matches help, but so do higher prices and FIFA’s own resale platform.5

That is the World Cup business model in one picture:

National emotion
      ↓
Global attention
      ↓
TV rights + sponsors + tickets + hospitality
      ↓
FIFA revenue
      ↓
Distributions, reserves, operations, development programmes
      ↓
More political power inside football

Simple game. Complicated machine.

The reason sponsors and broadcasters pay is obvious: almost nobody can still gather the planet at the same time.

The 2022 final between Argentina and France reached close to 1.5bn people globally, according to FIFA. That is more than ‘audience’. That is a rare piece of shared human attention.6

In a fragmented media world, that matters.

Most content is now personal. Your feed is not my feed. Your Netflix is not my Netflix. Your children probably live inside an algorithmic cave you are allowed to finance but not enter.

The World Cup cuts across that.

For a few weeks, people watch the same thing. In bars, offices, living rooms, airports, WhatsApp groups, and slightly suspicious fan zones sponsored by global beer brands.

That is why it is valuable.

3) America is buying the game too

There is a second story running underneath this World Cup.

US capital has moved heavily into football.

The FT reported that Americans now own 117 European clubs, including more than half of the English Premier League, more than a third of Italy’s Serie A, and more than a quarter of France’s Ligue 1.7

The logic is not difficult.

US sports teams have become brutally expensive. European football clubs still look cheaper on a revenue multiple basis. The sport is global. The brands are old. The fan bases are emotional. The stadiums can be upgraded. Commercial operations can be expanded.

To an American investor, football can look like an under-monetised asset class.

To many European fans, that sentence alone is enough to ruin the afternoon.

Because football is not only an asset. It is also heritage, place, memory, class, family, routine, religion without paperwork.

That is the tension.

American money looks at football and sees upside.

Many fans look at American money and see ticket inflation, VIP boxes, multi-club ownership, Super League ghosts, and half-time shows that may one day require their own insurance policy.

Both can be true.

4) The awkward part

You cannot write about FIFA’s money machine without mentioning the smell in the basement.

In 2015, the US Department of Justice unsealed charges against FIFA officials and sports marketing executives involving racketeering, wire fraud, money laundering and corruption. The scandal eventually forced Sepp Blatter out.89

The FT has also raised questions about FIFA’s current financial transparency, including how World Cup revenues are spent and how money flows to FIFA’s 211 member associations.1011

That does not mean the whole tournament is rotten. But it does mean the business model should not be treated like harmless entertainment.

When one organisation controls the biggest sports event on earth, distributes billions across member associations, chooses hosts, sells rights, sets rules, and operates with limited external scrutiny, governance is not a side issue.

It is part of the product.

The World Cup sells joy.

It also sells access, legitimacy, prestige, and political theatre.

This year, that political theatre will be especially visible. The tournament is hosted mainly in the US, under Donald Trump, with cities, immigration politics, ticket prices, and FIFA itself all becoming part of the story.12

But the politics should not swallow the football.

That would give politicians too much credit. A common mistake, usually made by politicians.

5) Why it still works

Here is the strange thing.

People know the World Cup is commercial. They know FIFA is controversial. They know ticket prices are often absurd. They know sponsors are not doing this out of spiritual commitment to corner kicks.

And still they watch.

Because the product is almost unfairly strong.

A World Cup match is simple enough for a child to understand and emotionally large enough to occupy an entire country. It is one of the few remaining formats where a single moment can jump from a stadium to a billion people almost instantly.

A penalty.

A mistake.

A teenager becoming famous.

A goalkeeper becoming a national enemy.

A country discovering, for 90 minutes, that it has always had strong views on defensive midfield structure.

This is why the World Cup survives its own organisers.

FIFA gets to stage and monetise it.

But it does not fully own what makes it powerful.

The power comes from the fans, the players, the flags, the old memories, the impossible goals, and the fact that every country can briefly imagine a better version of itself.

That is the magic. The business is built on top of it.

Our view

The World Cup is the clearest example of what happens when emotion becomes infrastructure. Founders and leaders of any organization can learn from FIFA.

1. Football creates the emotion.

2. Media turns it into reach.

3. Sponsors turn reach into money.

4. FIFA turns money into power.

5. And then, every four years, the cycle starts again.

So tomorrow, watch the football. Enjoy the goals. Complain about the referees.

Question the ticket prices.

And remember: behind the world’s simplest game sits one of the most sophisticated business machines ever built.

Not bad for something America once ranked below tractor-pulling.

🎚️🎚️🎚️🎚️ Producer’s Note

When Snoop Dogg shows up as a shareholder at Swansea, Ryan Reynolds turns Wrexham into a global content asset, and US investors keep buying European clubs, something has clearly shifted.

Football is still football.

But around it, a new industry is forming: sport, media, celebrity, private capital, hospitality, streaming, and global fandom all rolled into one.

So yes, Snoop fits.

Who owns the game now?

Go Switzerland 🇨🇭

Fab 🗽
LinkedIn | Insta | X

StudioAlpha Capital is a Delaware-structured pre-seed venture fund backing AI-native B2B software startups at day zero. Legal counsel: Cooley LLP. Fund administration: AngelList.

Sources

He may even be partly right. China is subsidising, overbuilding, exporting, and putting real pressure on European industry.

But blaming the other side usually says something about the person doing the blaming. We learned this quite early in school. Apparently it remains useful in geopolitics.

So the question is not only whether Macron is right. The question is why Europe now sounds like the victim in a story it helped write.

Subscribe now

1. Europe enters this fight tired

Europe is starting to talk tougher about China. Fine. But Europe is not entering this fight from a position of strength. It is entering it with weak growth, expensive energy, an unresolved war on the continent, higher defence costs, and a political class that has spent years confusing regulation with strategy.

The European Commission expects EU GDP growth to slow to 1.1% in 2026. Germany, after two years of recession and weak growth of 0.2% in 2025, is expected to grow only 0.6% in 2026. That is not a recovery. That is a patient sitting up in bed and being congratulated for movement.23

France has another problem: the size of the state. In 2024, public spending in France stood at 57.2% of GDP. Germany’s general government expenditure reached 49.5% of GDP. These are not small numbers. When the state absorbs around half the economy, and the result is still weak schools, slow trains, broken infrastructure, and endless administrative friction, the issue is no longer “we need more public money”. The issue is output.4

Savoir vivre: Changes in public spending in France, the euro area, and the euro area excluding France

Germany is a good example because the gap between reputation and reality has become hard to ignore. Deutsche Bahn’s long-distance punctuality was 63.4% in H1 2025. The federal audit office said only 69 of 280 planned bridge refurbishments were completed in 2024. KfW’s 2025 Municipal Panel put Germany’s municipal investment backlog at €215.7bn, with schools alone accounting for €67.8bn. This is what makes the China debate uncomfortable. China is not hitting a healthy European system. China is hitting a slow, expensive one.56

Jamie Dimon put the broader decline more brutally. Speaking in Dublin, he said Europe had fallen from 90% of US GDP to 65% over 10 or 15 years. The exact comparison can be debated. The direction is harder to dismiss. IMF work points in the same direction: Europe’s GDP per capita is around 72% of the US level, and much of the gap comes from lower productivity and a fragmented internal market. That is the background before China even enters the picture.78

Plainly said: It looks like things have been going downhill since the Euro monetary union.

2. China is now visible in Europe’s industrial mirror

The simple concern: Europe is becoming more worried about its dependence on China. That concern is justified. The EU’s goods trade deficit with China reached €359.8bn in 2025. In Q1 2026 alone, the deficit reached €98bn, the highest quarterly level since Q3 2022. It is a scorecard.910

For decades, Germany and parts of Europe were the global reference point for cars, machinery, chemicals, specialised components, and industrial systems. Now China is moving into many of the same categories. China is becoming “Germany in the global supply chain” and replacing Europe in areas where German and European manufacturing used to be strong. Yes, that hurts.

The public fight is focused on electric vehicles because cars are visible. German brands are not just companies; they are national furniture. EVs also sit at the intersection of climate policy, jobs, software, batteries, industrial policy, and pride. When Chinese EVs enter Europe at scale, the reaction is therefore not only economic but psychological.

The EU has imposed definitive countervailing duties on battery electric vehicles imported from China, after concluding that China’s BEV (Battery Electric Vehicle) value chain benefits from unfair subsidies and threatens economic injury to EU producers. That may be justified. But cars are only the cleanest version of the problem. China has built a manufacturing system Europe struggles to match: scale, subsidies, supplier depth, domestic competition, export discipline, and increasingly strong engineering. Europe has committees. The gap is visible.11

3. Tariffs can buy time. They do not create competitiveness.

Europe now wants stronger trade defence tools. That is understandable. No serious economy can ignore subsidised competition forever. Letting strategic industries disappear and then discovering dependence during a crisis is negligence. But a tariff only buys time. It does not fix high energy costs. It does not fix weak productivity. It does not fix fragmented capital markets. It does not make European software companies scale faster. It does not make permitting move at Chinese speed. And it does not turn a protected incumbent into a serious competitor.

James Kynge makes the sharper point in his podcast. His fear is that Europe puts up trade barriers against China and ends up protecting the very problems that created the weakness:

• huge inefficiencies,

• high energy costs,

• high labour costs,

• restrictive labour laws,

• restrictive investment laws,

• and very low productivity.12

If Europe blocks Chinese imports, Chinese companies can still move production into Europe. That is already happening. Rhodium Group and MERICS report that Chinese direct investment in Europe rose to €16.8bn in 2025, the highest level since 2018. The automotive sector alone attracted €7.6bn, with 93% focused on the EV supply chain.13

So the tariff question becomes more awkward. If Chinese EVs are imported from China, Europe can tax them. If Chinese EVs are produced in Hungary, Spain, France, or Germany, the story changes. Europe may get factories, jobs, and ribbon-cutting photos. It may not get European champions. It may simply get Chinese industrial capability operating inside Europe’s borders.14

4. Germany is trapped inside the argument

Germany sits in the middle of this problem. It wants Europe to defend industry. It also wants to avoid a trade conflict that damages German companies in China. The transcript captures this directly: many of Germany’s biggest companies still have deep exposure to China, and that makes Berlin cautious.

That is the German trap. Germany helped build Europe’s China exposure. For years, China was a market, supplier, manufacturing base, and profit machine. Now China is also a competitor. That changes the relationship. The same country that bought German cars is now producing electric vehicles that make parts of the German model look old.15

The old German model was built around engineering depth, export discipline, industrial employment, affordable energy, and access to global markets. Several pillars are now under pressure at once. Energy is more expensive. China is more competitive. US industrial policy is more aggressive. Software matters more in cars and industrial systems. AI changes productivity expectations. And Europe’s internal market remains less integrated than it looks from a Brussels slide deck.

This is why Europe’s China policy risks becoming confused. France wants more protection. Germany fears retaliation. Southern Europe wants jobs and investment. Eastern Europe may welcome Chinese factories. Brussels wants strategic autonomy. Everyone agrees Europe needs to become stronger. Nobody wants to pay the transition cost. The minutes will be excellent.16

Pretty tricky.

5. China can hit back

A trade war with China is might be not good idea. China has tools such as brandy, pork, agriculture, critical minerals, and rare earths. That is the right list.

China has already targeted EU brandy in the wake of the EV tariff fight. Reuters reported that China imposed temporary anti-dumping measures on EU brandy in October 2024, hitting French brands including Hennessy and Rémy Martin. In 2025, China also placed initial anti-dumping duties on EU pork imports, deepening the trade tensions that followed the EV dispute.1718

Agriculture matters because Europe’s political economy is not a clean spreadsheet. A French cognac producer, a Spanish pork exporter, a German carmaker, a Hungarian battery plant, and a Dutch port operator do not experience China policy in the same way. China knows that. It can apply pressure sector by sector. Europe then gets what it does best: domestic fights about who should suffer first.

The more serious exposure is critical raw materials. The Council of the European Union says China provides 100% of the EU’s supply of heavy rare earth elements. These materials matter for defence, EVs, wind turbines, electronics, motors, sensors, and industrial systems. That is not a dependency. That is a leash.19

Strategic critical raw materials
(orange)

So before Europe talks itself into a trade war, it should ask one simple question: what happens on day three after China responds? If the answer is another summit, the strategy is not ready.

Really tricky.

6. The AI layer makes this worse

The old industrial fight was about cars, machinery, chemicals, energy, labour cost, and supply chains. The next industrial fight adds AI. This is where Europe’s position looks worst.

According to the Stanford 2026 AI Index, US private AI investment reached $285.9bn in 2025. The US invested 23 times more private capital into AI than China. Stanford also notes that US generative AI investment exceeded the combined total of China and Europe by a wide margin.20

This is the capital map. The US has the hyperscalers, frontier labs, chip access, enterprise software giants, and the largest venture market (91% in Q1/2026 as written in our Quarterly Investor Letter). China has state coordination, manufacturing scale, engineering depth, domestic deployment, and the ability to connect AI with industrial production. Europe has talent, universities, research, and speeches about sovereignty. Talent matters. Speeches less so.

AI is not an isolated tech sector. It becomes part of industrial competitiveness. It changes how cars are designed, how factories are run, how logistics are managed, how software is built, and how companies learn. The next industrial base is software, data, compute, energy, capital, distribution, and speed. Europe is weak on too many of those at once.

Large AI infrastructure announcements are therefore useful but not sufficient. France attracting major AI investment pledges is positive. Europe needs compute, data centres, energy capacity, and sovereign options. But compute does not automatically create category-leading companies. A data centre is infrastructure. It becomes strategically useful only when it supports products, customers, talent density, venture capital, procurement, and fast adoption. Otherwise it is concrete with a press release.21

7. The question Europe should ask

Europe should not be passive about China. China uses state support, industrial strategy, export scale, critical-material leverage, and domestic market power. Europe has every right to defend itself against unfair competition.

But blaming China is the easy version. China did not create Europe’s productivity gap. China did not fragment Europe’s capital markets. China did not make France’s public spending reach 57% of GDP. China did not make Germany slow in software-defined cars. China did not stop Europe from building deeper AI capital markets. China did not stop Europe from implementing the Draghi report. China is exposing the weaknesses Europe built for itself.22

That is the better frame. Trade defence should buy time for reform. It should not become a blanket over weak execution. A serious European response would defend genuinely strategic industries, fix the internal barriers that stop companies from scaling across one real market, and treat AI, software, compute, energy, and industrial policy as one connected competitiveness problem.23

Trade policy interventions in Europe

Without that, tariffs become theatre. The uncomfortable lesson for founders is simple: do not wait for Europe to protect you, finance you, or move fast enough around you. Build for the real world: international capital, fast product cycles, hard technical differentiation, distribution outside your home market, and customers who pay because the product works. That is the part Europe still controls.

Europe is not enough as a default market. That is why we help founders build for the US from day zero. Apply here.

🎚️🎚️🎚️🎚️ Producer’s Note

Europe and China, basically:

Je t’aime… moi non plus. (“I love you… me neither.”)

Love, dependence, suspicion, trade deficits, and just enough drama to keep Brussels busy.

Freedom would be nice.

Best,
Fab 🗽
LinkedIn | Insta | X

StudioAlpha Capital is a Delaware-structured pre-seed venture fund backing AI-native B2B software startups at day zero. Legal counsel: Cooley LLP. Fund administration: AngelList.

Sources

At the same time, markets have kept rising.

That looks wrong.

It may not be.

There is an old investor line about buying when cannons fire. The harder question is what markets are actually buying when the world looks less stable, not more.1

Subscribe now

1) War is not one economic event

The word “war” covers too much.

A regional oil shock, a European land war, a naval blockade, a cyberattack, and a direct conflict between major powers are not the same economic event.

They may all create fear. But they do not hit the economy in the same way.

The FT’s latest chart analysis makes this visible. Since the Iran war began, the world’s biggest listed companies have added about $5.4 trillion in value. But this is not broad market calm. Semiconductor companies with market values above $10 billion gained $3.7 trillion, while airlines, consumer goods companies, carmakers, miners and several defence groups came under pressure.2

For markets, the relevant question is narrower:

Which economic channel is being hit?

Ukraine and Iran sit in different parts of this table.

• Ukraine has become an energy, fiscal, food, defence and reconstruction story.

• Iran is more directly an oil, shipping, inflation and escalation story.

That difference matters.

This also explains why markets can look calm when the world is not calm.

Markets are not designed to measure suffering. They measure claims on future cash flows. A war can be a human disaster and still not immediately break the profit outlook for the largest listed companies.

At the moment, AI and chip stocks are the larger force at index level. That does not mean the war has no economic cost. It means the cost is showing up unevenly.

This is not broad calm. It is selective pricing.3

2) 🇺🇦 Ukraine: the long economic shadow

Ukraine shows how war becomes a balance sheet problem.

The first market shock in 2022 was only one phase. The longer economic effects came through energy prices, food prices, sanctions, fiscal support, migration, infrastructure damage, defence spending and reconstruction.

The World Bank’s latest assessment, prepared with the Ukrainian government, the European Commission and the United Nations, estimates Ukraine’s reconstruction and recovery needs at almost $588 billion over the next decade. That is nearly three times Ukraine’s estimated nominal GDP for 2025.4

The same assessment estimates direct damage at around $195 billion, with housing, transport and energy among the worst-hit sectors.5

That is the economic structure of a long war.

It is not only destroyed buildings. It is damaged power generation, broken logistics, weaker public finances, labour displacement, higher insurance costs, and years of capital allocation into repair rather than growth.

For Europe, Ukraine also exposed a dependency that had been underestimated for years: cheap imported energy.

Once that assumption broke, the shock moved from geopolitics into electricity prices, industrial margins, household budgets and central bank policy.

War becomes economically powerful when it touches an existing dependency.

3) 🇮🇷 Iran: the oil channel

Iran is different.

The market does not need to price a global war to worry about Iran. It only needs to price oil, shipping, inflation and escalation risk.

Reuters reported that oil prices rose as talks with Iran stalled, while U.S. equities still edged higher because AI and chip strength remained the dominant market force.6

The economic chain is direct:

Iran → oil → inflation → rates → valuations

Higher oil prices can squeeze consumers, airlines, logistics firms and energy-intensive industries. They can also support energy stocks. If oil keeps inflation higher, central banks have less room to cut rates. If rates stay higher, equity valuations become harder to defend.

This is why Iran can matter even for companies with no direct exposure to Iran.

The channel is macro, not local.

4) Trade may reduce war. Decoupling changes the calculation.

Soumaya Keynes’ recent Financial Times column is a useful entry point into the deeper question.

The column revisits Norman Angell, the British writer often mocked for supposedly arguing before the First World War that trade had made war impossible.

That version is too simple. Angell later won the Nobel Peace Prize in 1933. The irony is that he is remembered mostly for being “wrong” before World War I, although his real argument was not that war was impossible, but that modern war had become economically self-defeating.7

Angell’s actual argument was more careful. War was not impossible. War had become economically irrational. Trade had made conflict more expensive and conquest less useful.8

A new NBER working paper by Ling Feng, Qiuyue Huang, Zhiyuan Li and Christopher M. Meissner tests a related idea with modern data.

The paper studies global bilateral data from 1962 to 2014. To address the problem that peace itself may cause trade, the authors use changes in air transport technology as a way to isolate the effect of trade on conflict. Their result: increased trade significantly reduces both the probability and intensity of conflict between countries.9

The paper calls this the “peace dividend” of international trade.

The mechanism is straightforward. When two countries trade heavily, conflict becomes more expensive. Firms lose customers. Consumers lose supply. Governments lose tax revenue. Sanctions become more powerful.

That does not mean trade guarantees peace. It clearly does not.

It also does not mean every dependency is healthy. Some dependencies become strategic vulnerabilities.

But the argument creates a serious tension for today’s world.

Governments want more resilient supply chains. They want fewer critical dependencies on rivals. They want to reduce exposure to countries they no longer trust.

That is understandable.

But removing economic links can also reduce one layer of deterrence.

De-risking may make countries safer in one way and more exposed in another.

5) Geopolitical risk (GPR) still has a cost

The fact that markets can rise during conflict does not mean geopolitical risk is harmless.

Dario Caldara and Matteo Iacoviello’s Geopolitical Risk Index (GPR Index), published in the American Economic Review, measures adverse geopolitical events and related risks using news data. Their work finds that higher geopolitical risk foreshadows lower investment and employment and is associated with larger downside risks.10

This matters because geopolitical risk often works through hesitation.

• Companies delay investment.

• Boards postpone expansion.

• Consumers become more cautious.

• Investors demand higher returns.

• Governments allocate more money to defence and less to other priorities.

The effect is not always immediate. It can accumulate.

This also explains why index levels can mislead.

The S&P 500 and Nasdaq can rise while parts of the real economy absorb new friction. A small number of large AI and chip companies can pull the index higher while airlines, industrial firms, retailers or European manufacturers face a more difficult cost environment.

The index is not the economy.

The FT’s chart analysis shows how unusual the current rebound is. The MSCI World index recovered from its low point after the Iran conflict in 12 trading days. After Russia’s full-scale invasion of Ukraine and the 2022 rate shock, the same recovery took 338 trading days. After Covid, it took 117 trading days.

That does not mean geopolitical risk has disappeared. It means the index has recovered faster than the underlying economic damage has cleared.

6) The market’s current answer

At the moment, the market appears to believe that Iran is an oil and inflation risk, but not yet a full global shock. Ukraine remains a major European and fiscal burden, but it is not currently breaking U.S. corporate earnings. AI-related earnings are strong. The U.S. economy remains resilient. Liquidity is still supportive.

That assessment may prove wrong.

But it is not irrational.

Markets can look calm when the world is not calm because markets are comparing forces. Today, AI earnings and U.S. resilience are still stronger than the priced economic damage from Ukraine and Iran.

If the oil channel becomes persistent, if shipping routes are disrupted for longer, if inflation expectations move again, or if escalation changes the policy outlook, that balance can shift.

Our view

The old saying about buying when cannons fire contains one useful lesson: Do not confuse fear with analysis. The weaker version of the saying turns war into a trading slogan. That is too crude.

The better version asks a colder question: Which economic channel has changed?

For investors, the relevant channels are:

• energy

• inflation

• rates

• trade

• sanctions

• defence

• reconstruction

• confidence.

Our current reading: energy and inflation are the channels to watch most closely. Trade and sanctions are slow-burn risks. Reconstruction is real, but later. Defence remains structurally supported, though not every defence stock benefits automatically.

For the rest of the year, we would expect more sector dispersion than index panic: AI and chips can keep carrying the index, while energy-sensitive, logistics-heavy and consumer-facing businesses feel more pressure underneath.

For founders and CEOs, the same logic applies. The question is not whether the world feels unstable. It does. The question is whether the instability changes costs, customers, suppliers, financing or timing.

Markets may keep rising while wars remain unresolved.

That does not mean the wars do not matter. It means the market has not yet found a transmission channel strong enough to dominate the story.

B O X

For paid readers

‘Pulse’ is where we unpack what is moving.
’Backstage’ is where we turn similar shifts into leadership briefs, decision tools, and operating questions for founders, CEOs and boards.
If you want the practical layer behind essays like this, join Backstage.

🎚️🎚️🎚️🎚️ Producer’s Note

The market is not calm because the world is calm. It is calm because, for now, earnings and AI are louder than oil and war.

That can hold.

Until the channel changes.

Share

Fab ☮️

LinkedIn | Insta | X

StudioAlpha Capital is a Delaware-structured pre-seed venture fund backing AI-native B2B software startups at day zero. Legal counsel: Cooley LLP. Fund administration: AngelList.

Sources

Last week’s All-In Podcast1 opened with the pressure around OpenAI: missed user targets, missed revenue targets, IPO doubts, and a courtroom fight with Elon Musk. David Sacks framed it as a bad week for OpenAI. It exposes the deeper question: what kind of company OpenAI has become.

It is also easy to read badly.

The loud story is the lawsuit. The more important story sits in the operating model. OpenAI is trying to scale one of the most capital-intensive technology businesses ever built while the physical world is struggling to keep up: power, data centres, chips, grid connections, transformers, and the cost of turning model intelligence into paid work.

Can OpenAI scale intelligence fast enough now that the constraint has moved from software into infrastructure?
And for users, founders, and investors, the practical question is just as direct: is ChatGPT still the AI platform to bet on?

Subscribe now

1. The noisy week

The visible facts are simple.

The Wall Street Journal reported that OpenAI missed internal targets for user growth and revenue. The company had expected to reach 1 billion weekly ChatGPT users before the end of 2025, but had not done so several months into 2026. The same report said OpenAI had also missed its 2025 revenue target for ChatGPT, while carrying very large future compute commitments.2

Reuters reported that OpenAI CFO Sarah Friar had raised concerns internally about future computing contracts if revenue growth does not keep pace. OpenAI pushed back publicly. Sam Altman and Sarah Friar said they were aligned on buying as much compute as possible.3

Then came the trial.

Reuters reported that Elon Musk’s lawsuit against OpenAI is now putting the company’s nonprofit-to-profit transition, founder incentives, Microsoft relationship, and internal governance under the microscope. Greg Brockman, OpenAI’s president and co-founder, disclosed a stake estimated near $30 billion and financial ties to Sam Altman. Musk argues that OpenAI betrayed its original mission. OpenAI argues that Musk is trying to regain influence and slow a competitor.4

This gives the week a clean media shape.

OpenAI looks stretched. The CFO worries about compute. The CEO wants to move faster. The old nonprofit story collides with one of the largest private-company valuations in the world. Elon Musk, now building xAI, is back in the room through the court door.

The Wall Street Journal’s Tim Higgins framed the same week as a test of Sam Altman’s Silicon Valley lore: “the founder who made OpenAI look almost inevitable is now facing missed-target scrutiny, IPO pressure, Anthropic momentum, and Elon Musk in court.”5

That is enough for a good media narrative. It is not enough to understand OpenAI.

The missed targets matter. Governance matters. The lawsuit matters, at least for control and perception. But none of these explains the deeper tension. OpenAI is being judged like a software company while its economics increasingly look like infrastructure.

That mismatch is where the story gets interesting.

2. OpenAI is still a product force

Yes. Below is Chapter 2, same structure, sources added only where needed.

The simple bear case says: OpenAI missed targets, therefore demand is weaker than expected.

That is too clean.

OpenAI still has enormous consumer reach. ChatGPT remains one of the most important software products ever launched. Even if the company missed its internal billion-user target, the scale is extraordinary. The problem is that internal targets, public perception, product strength, and infrastructure capacity are now all tangled together.6

OpenAI is also still shipping.

What GPT-5.5 changes

OpenAI released GPT-5.5 on April 23. Compared with Anthropic’s more dramatic Claude launches, it arrived almost quietly: less theatre, more product update. It describes GPT-5.5 as a model for complex work: coding, research, data analysis, documents, spreadsheets, and tool use. The main claim is less “better chatbot” and more “can carry more of the task itself.”7

---

---

The system card says GPT-5.5 understands tasks earlier, asks for less guidance, uses tools more effectively, checks its work, and keeps going until the job is done.8

OpenAI also released GPT-5.5 Instant as the new default ChatGPT model, claiming fewer hallucinations and better factuality in high-stakes areas like law, medicine, and finance.9

The company has been pushing ChatGPT deeper into work: coding, research, documents, spreadsheets, long context, tool use, and agent-style tasks. Codex and the broader developer workflow matter here. Coding has become one of the most commercially important AI use cases because it turns model capability into measurable productivity, budgeted usage, and high-frequency demand.10

That matters for the OpenAI vs Anthropic story.

OpenAI’s edge is that it has been more aggressive in securing dedicated compute capacity ahead of demand. Anthropic is growing fast, but much of its capacity comes through Amazon, Google, and other partners. That may be prudent. It also means Anthropic can be more exposed when compute gets rationed.

You can already see the tension around Claude. Anthropic says recent Claude Code quality issues came from prompt changes, not reduced reasoning, and published a postmortem after user complaints. But the wider market read was clear: when AI products become serious work tools, users notice every throttle, slowdown, quota, and quality wobble immediately. The White House also reportedly raised compute-capacity concerns around Anthropic’s plan to expand Mythos access, alongside national-security concerns. That does not prove Anthropic faked a security argument because of compute. It does show that safety, access, and capacity are now tangled together.1112

OpenAI does not simply “own data centres” like Google/Microsoft. Stargate is partner-centric: Oracle, SoftBank, Microsoft, utilities, builders, finance partners. OpenAI says it is securing/building compute capacity with partners, not doing everything alone.13

Anthropic also has big compute deals, but mostly through hyperscalers: Google/Broadcom TPUs, Amazon, CoreWeave. That means Anthropic is more dependent on partner allocation. Anthropic itself says it signed a Google/Broadcom deal for multiple gigawatts of TPU capacity starting in 2027.14

Over the past year, Anthropic has taken a lot of the enterprise narrative. Claude became especially strong in coding and knowledge work. In our recent Anthropic vs. OpenAI piece, the key point was that Anthropic still trails OpenAI in consumer reach, but the revenue and enterprise gap narrowed much faster than most people expected.15

OpenAI is now fighting on several fronts at once.

Anthropic is strong in enterprise and developer workflows. Google has Gemini, Search, Android, Workspace, Cloud, and its own infrastructure. xAI has Elon Musk, distribution through X, an aggressive compute build-out, and now a reported SpaceX agreement around Cursor. DeepSeek and other open-weight models put pressure on pricing and on the assumption that frontier capability must always remain closed and expensive.16

This is not a simple two-company race anymore.

OpenAI still has the broadest public brand in AI. Anthropic has a focused enterprise machine. Google has unmatched distribution and infrastructure. xAI has a founder who can turn attention, capital, and engineering speed into a weapon. DeepSeek (see our essay last week) has shown that efficiency and open weights can move the market even when the US labs still lead in many areas.

That makes OpenAI’s position more complicated, but not weak.

The company is still one of the central players in the market. It still has distribution, product velocity, brand, developer attention, enterprise demand, and one of the strongest research teams in the world.

The issue is different. OpenAI’s ambition requires the physical world to move at software speed.

The physical world is not very good at that. Annoying habit.

3. The constraint has moved into the physical world

AI used to be discussed mainly as a model race.

• Who has the best benchmark?

• Who has the longest context window?

• Who is ahead in coding?

• Who is cheaper per token?

• Who has the best agent demo?

Those questions still matter. But they no longer carry the whole story.

In our DeepSeek V4 piece, we used a simple line: the model gets the headline, but the stack carries the weight.

That stack now reaches deep into the physical world.

The International Energy Agency says electricity consumption from data centres is rising fast, and AI-focused data centres are growing faster than the overall data-centre market. In its 2026 energy and AI work, the IEA projected that data-centre electricity consumption could roughly double from 2025 to 2030, while AI-focused data-centre electricity consumption could triple over the same period.17

Electricity consumption by data centres, 2020-2035

The Financial Times put the US power crunch in sharper terms. According to its analysis, new data centres may require about 44GW of additional capacity by 2028, while only about 25GW may come online in time. That leaves a gap of roughly 19GW.18

Bloomberg has reported on another bottleneck: electrical equipment. Transformers, switchgear, and other grid components are not glamorous, but without them, data-centre plans do not become operating capacity.19

This is the part of AI that does not fit neatly into the app-store imagination.

You can launch a model update overnight. You cannot launch a grid connection overnight. You can raise capital quickly in a hot market. You cannot manufacture transformers, secure permits, build substations, contract power, and bring a massive data centre online at the same pace.

That matters for OpenAI because compute is not a side cost. It is the production input.

Every better model, longer context window, coding agent, research task, voice interaction, image generation, and enterprise workflow consumes compute. The more useful the product becomes, the more demand it creates. If the system works, usage expands. If usage expands, the need for compute expands with it.

This creates a strange-looking business pattern.

A normal software company tries to scale revenue while marginal costs stay low. OpenAI is scaling a service where the product becomes more valuable as it becomes more compute-hungry. Efficiency improvements help, and they matter a lot. But the aggregate demand curve is still brutal.

This is why the missed-target story can be misleading.

A SaaS company missing user and revenue targets often points to weaker demand, weaker retention, poor pricing, or competition. For OpenAI, those factors may all matter, but the supply side is unusually important. Capacity, pricing, latency, reliability, model routing, inference cost, and access to compute all shape what the company can sell and how profitably it can sell it.

The better OpenAI gets, the more infrastructure it needs.

That is the core tension.

4. What OpenAI has to prove

The company now has to prove that it can turn capital into compute, compute into reliable products, and reliable products into paid work at a pace that justifies its valuation and spending commitments.

The Wall Street Journal reported that OpenAI missed internal user and revenue targets while carrying concerns around very large future compute commitments.20

That is a different kind of company from the one people imagined when ChatGPT first broke into the mainstream.

Back then, the story felt like software: a new interface, a magical chatbot, a consumer product spreading through the world at absurd speed.

Today, OpenAI looks more like a hybrid between a software platform, a research lab, and a massive infrastructure buyer with enormous distribution.

That makes the business harder to read.

If OpenAI underbuilds compute, it risks product quality, latency, reliability, and enterprise momentum. If it overbuilds compute, it risks carrying massive commitments before revenue catches up. If it moves too slowly, Anthropic, Google, xAI, and open-weight models take share. If it moves too aggressively, the CFO problem becomes real.

This is why the lawsuit is a distraction, even if it matters.

The courtroom asks who controls OpenAI’s mission.

The market asks whether OpenAI can justify its valuation. The Financial Times reported that some investors have questioned OpenAI’s $852bn valuation as the company shifts strategy and faces pressure from Anthropic and Google.21

The operating question is more concrete: can OpenAI secure enough compute, at the right cost, and turn it into work customers will pay for every day?

That is where Anthropic matters: Anthropic’s strength is not only that Claude is good. It is that Claude has become useful in high-value workflows, especially coding and enterprise knowledge work. That type of usage is easier to monetise than casual chatbot traffic because it sits closer to budgets, teams, and measurable output.

The Wall Street Journal reported that Anthropic is launching a $1.5bn joint venture with Wall Street firms that aim to sell artificial-intelligence tools to companies, including those backed by private-equity firms. — a clear push deeper into enterprise deployment.22

That is where Google matters: Google can place Gemini into existing surfaces: Search, Android, Workspace, Cloud. It also owns large parts of the infrastructure stack. In Q1, Google said Gemini Enterprise paid monthly active users grew 40% quarter over quarter.23

That is where xAI matters: Elon Musk understands distribution, capital markets, manufacturing pressure, and infrastructure theatre better than most technology founders. Bloomberg reported that SpaceX has an agreement giving it the right to acquire Cursor for $60bn later this year, or pay $10bn for joint work, as Musk pushes into AI coding tools.24

DeepSeek pressures the cost of intelligence from below.
Anthropic pressures OpenAI in enterprise workflows.
Google pressures it through distribution and infrastructure.
xAI pressures it through speed, capital, and coding ambition.

The power grid pressures everyone.

OpenAI sits in the middle of all of it.

Our view: The company is still central to the AI market. But the easy phase of the story is over.

The next test is whether OpenAI can industrialise intelligence.

Industrialising intelligence means building the physical, financial, and operational system behind the software. It means buying compute before revenue fully proves it. It means betting on demand curves that may be real but uneven. It means managing a product race while also managing infrastructure risk.

It is a new kind of industrial build-out.

🎚️🎚️🎚️🎚️ Producer’s Note

YUNGBLUD fits this one.

OpenAI and Anthropic are still the young blood: barely a few years old, moving fast, loud, imperfect, and already carrying too much expectation.

That is the strange part of this AI moment. These companies still feel like startups, but the systems around them now look industrial: courts, governments, data centres, power contracts, and capital markets.

Young blood with transformer shortages. Less romantic. More expensive.

Share with your yung friends.

Share

Fab 🌓

LinkedIn | Insta | X | Studio⍺

StudioAlpha Capital is a Delaware-structured pre-seed venture fund backing AI-native B2B software startups at day zero. Legal counsel: Cooley LLP. Fund administration: AngelList.

Sources

Sixteen months ago, DeepSeek made the market wobble. Nvidia lost roughly $593bn in market value in one day after investors realised that a small Chinese lab had built a strong AI model at a fraction of the expected cost.1

This week, pressure came from the American side of the trade. The Wall Street Journal reported that OpenAI had missed internal user and revenue targets, while carrying concerns about more than $600bn in future compute commitments. Oracle, CoreWeave, Nvidia and SoftBank fell after the report.2

A few days earlier, DeepSeek had released V4. Bigger context window. Open weights. Aggressive pricing. Huawei in the background.

And this time, DeepSeek is no longer just a new startup story. China’s state, chip policy, domestic AI stack, and even founder freedom are now part of the picture. The Manus case makes that visible: AI talent and AI IP have become strategic assets.

The story got heavier.

New: Backstage Briefs launches next Tuesday

For the past two years, this publication has been our public place for essays on AI, technology, markets, and company-building. We now call these free weekly essays Pulse. Nothing else changes.

Next Tuesday, we add a second format: Backstage — paid decision briefs for founders, CEOs, owners, and senior leaders.

Backstage briefs are short, practical, and built for people who have to turn AI, market, and organisation shifts into concrete moves. Each brief comes with a usable artifact: a checklist, script, memo structure, decision table, or action plan.

• Pulse stays free every Thursday.

• Backstage briefs run every second Tuesday for paid subscribers.

1. The unusual company behind DeepSeek

DeepSeek has an unusual origin story.

It grew out of High-Flyer, a Chinese quantitative hedge fund based in Hangzhou and co-founded by Liang Wenfeng. High-Flyer used machine learning for trading before DeepSeek became a public name in AI.3

That background matters.

A quant fund has a different starting point from a consumer software company. Models are not built for nice demos. They are built to find patterns, test signals, manage uncertainty, and work under pressure.

That seems to have shaped DeepSeek’s character: less theatre, more focus on efficiency, cost, and architecture.

In 2023, High-Flyer announced that it would move beyond investing and put resources into exploring artificial general intelligence. DeepSeek was created later that year.4

The timing matters too. DeepSeek developed while U.S. export controls were limiting China’s access to the most advanced Nvidia chips. That pressure did not stop Chinese labs from working on frontier AI. It made efficiency more important.5

A simple timeline helps:

DeepSeek is still a young company. That makes the story harder to read cleanly.

Some claims around training cost, chip access, and model development sit inside a fog of company statements, competitor reactions, geopolitics, and market theatre. The clean version is probably too clean.

The useful point is simpler: DeepSeek came from a place where algorithms already had to perform, where cost mattered, and where hardware constraints were real. That makes V4 worth looking at closely before we get lost in the usual benchmark parade.

2. China is building its own AI stack

DeepSeek V4 also lands inside a bigger China question.

Jensen Huang, Nvidia’s CEO, recently debated this with Dwarkesh Patel. Huang is not a neutral voice here. Nvidia wants to sell chips to China. Still, his argument is useful because he explains AI as a full stack: energy, chips, software, models, developers, cloud, and applications as already discussed in our Backstage piece “How to invest in AI”. In the interview, he called AI a “five-layer cake.” He also argued that the U.S. should try to win across every layer, including the chip layer.6

Here is the plain version:

Huang’s China numbers are blunt.

He said China manufactures around 60% of the world’s mainstream chips. He said China has 50% of the world’s AI researchers. Later in the interview, he also said 50% of AI developers are in China. He described China as the second-largest computing market in the world, with plenty of energy, plenty of chips, and enough compute to aggregate if it chooses to.7

That is the pro-engagement argument.

If China already has energy, chips, researchers, developers, and a large domestic market, then cutting it off from Nvidia does not stop Chinese AI. It may push Chinese labs to build faster around Huawei and other domestic alternatives.

DeepSeek V4 fits that pattern.

Reuters reported that DeepSeek V4 was adapted for Huawei’s Ascend chips. Huawei said its Ascend 950-based supernode clusters support the V4 series, and that Huawei chips were used for part of V4-Flash’s training. Earlier DeepSeek models, including V3 and R1, were trained on Nvidia chips.8

Huang’s concern is simple: if the strongest open models start running best on non-American hardware, the global AI stack may slowly shift away from Nvidia and CUDA.

He put it sharply in the interview. China is, in his words, the largest contributor to open-source software and open models. Today, much of that still runs on the American tech stack. If that changes, the U.S. could lose influence over the layer where developers actually build (compare “software stack”).

Dwarkesh pushed the harder counterargument.

Compute is not a harmless export good. More compute can mean stronger models, faster training, and more deployed copies of capable systems. In cyber, the difference between one thousand agents and one million agents matters.

He pointed to Anthropic’s Mythos Preview, which Anthropic did not release publicly because of cyber-offensive capability concerns. According to Dwarkesh’s summary in the interview, Mythos found thousands of high-severity vulnerabilities across major operating systems and browsers, including one in OpenBSD that had existed for 27 years. Compare also our Soundcheck piece “How Anthropic Closed the Gap with OpenAI”.9

He also pushed on the raw compute gap.

In simple terms, FLOPs measure raw AI horsepower: how much mathematical work computers can do. Dwarkesh said some estimates put China at roughly one-tenth of U.S. compute on that measure.

He then pointed to chip manufacturing. Advanced chips are usually described by process nodes like 7nm, 3nm, 2nm, or 1.6nm. Smaller usually means a more advanced generation. His argument: China is still closer to 7nm, while Nvidia and its manufacturing partners are moving toward 3nm, 2nm, and eventually 1.6nm.

Huang pushed back.

His point was that chip manufacturing is only one source of progress. Better architecture and better software can create much larger gains than a smaller chip node alone.

The graph “benchmark and efficiency comparison figure”10 says two things:

• Left side: DeepSeek claims V4-Pro performs roughly in the same league as Claude, GPT and Gemini on reasoning, coding, and agent-style tasks. It is a benchmark comparison, so treat it as vendor evidence.

• Right side: the more important part. Compared with DeepSeek V3.2, V4 needs much less compute and memory for long context. At around 1M tokens, V4-Pro uses about 3.7x less compute and 9.5x less KV-cache memory; V4-Flash is even cheaper.

Simple meaning: DeepSeek is not only trying to make the model smart. It is trying to make long-memory AI much cheaper to run.11

He gave Nvidia’s own example: Moore’s Law now improves performance by roughly 25% per year. Between Hopper and Blackwell, the transistor improvement was around 75% over three years. But Blackwell delivered roughly 50x the performance of Hopper.

That gap is the point. The big jump did not come only from smaller transistors. It came from better system design: how chips are connected, how memory is used, how data moves, and how software tells the hardware what to do.

That is an important point for DeepSeek.

DeepSeek is not trying to win only by having more chips. It is trying to win through architecture, efficiency, mixture-of-experts models, better attention mechanisms, and lower memory cost. In plain English: if compute is scarce, you get smarter about using it.

Huang said this directly. Limited compute forces strong researchers to invent better algorithms. He mentioned mixture-of-experts and attention mechanisms as examples of software choices that reduce compute needs.

Dwarkesh also pushed on hardware quality.

He said Huawei’s 910C looks weaker than Nvidia’s H200 on FLOPs, bandwidth, or memory — roughly half to one-third, depending on the comparison. Huang’s answer was practical: China can use more chips, connect them together, and lean on networking. “They use twice as many,” was his rough answer.12

That sounds inelegant. It may also be enough.

Side note: Huang’s argument is worth taking seriously. He built Nvidia. Dwarkesh Patel’s role is different. He has become one of the better-prepared AI interviewers in Silicon Valley, and in this interview he pushed the security side of the China-chip debate harder than most business journalists would. Worth watching.13

AI systems are not only built from the best single chip. They are built from chips, networking, memory, power, software, and engineering discipline. Huawei is a networking company. China is good at manufacturing. Energy matters. Scale matters.

This is why DeepSeek V4 on Huawei matters.

It does not prove that China has caught Nvidia. It shows that Chinese labs are adapting the model layer to the domestic chip layer. That is the early shape of a separate AI stack.

Reuters adds the practical constraint: DeepSeek said V4-Pro can cost up to 12 times more than V4-Flash because high-end compute capacity remains tight. The company also said Pro pricing could fall once Huawei Ascend 950 supernodes are deployed at scale.14

The Manus case shows the same state logic from another angle. Meta reportedly agreed to buy Manus, a Chinese-founded AI-agent startup now based in Singapore, for more than $2bn. China’s National Development and Reform Commission then ordered Meta to unwind the deal on national-security grounds. Reuters also reported that Manus’ co-founders had been summoned by regulators and barred from leaving China.

The message is not subtle: Chinese-origin AI technology, talent, and IP may remain politically sensitive even after a company moves offshore. This is what state involvement looks like in practice. It can support a domestic AI stack. It can also limit founder freedom, foreign capital, and exit paths.15

So the China story is not clean.

Huang says the U.S. should stay ahead, keep the best technology at home first, and still compete globally. Dwarkesh says advanced chips can accelerate dangerous capabilities and should be controlled more aggressively.

Both sides have a point. Annoying, but useful.

DeepSeek V4 sits exactly in the middle of that argument. It is a Chinese model, built under chip pressure, released with open weights, and now adapted to Huawei hardware. That makes it more than a model launch. It is a small preview of a world where AI competition becomes stack against stack.

My view: Huang’s argument is stronger than the political mood around export controls. The U.S. should protect genuinely sensitive capabilities. But trying to hold China back as a broad strategy has a bad incentive structure: it pushes Chinese labs, chipmakers, cloud providers, and developers to build around the U.S. stack faster. That is already visible in China’s localisation drive. The harder, better strategy is to stay ahead, keep developers tied to the best global stack, and compete from strength. Economics first: China will build AI either way. Pretending otherwise is policy theatre with expensive side effects.1617

3. What DeepSeek V4 actually is

A little engine-room tour first. Then we translate the jargon. DeepSeek V4 is technical enough that skipping the machinery would make the story too blurry.

DeepSeek V4 comes in two versions: V4-Pro and V4-Flash.

V4-Pro is the heavy model: 1.6 trillion total parameters, with 49 billion active during inference. V4-Flash is the smaller, cheaper version: 284 billion total parameters, with 13 billion active. Both support a one-million-token context window. DeepSeek says the models are open-sourced, available through its API, and support both thinking and non-thinking modes.18

It means DeepSeek V4 is a mixture-of-experts model.

It has 1.6 trillion parameters in total — the full “library” of knowledge/weights.
But for each answer, it only activates 49 billion parameters — the relevant “experts” for that task.

Simple analogy: a hospital may have 1,600 doctors, but your case only needs 49 specialists in the room (very expensive hospital, yes). This is how the model can be huge in capability, but cheaper to run than activating everything every time.

The Hugging Face model card lists the license as MIT and says the model weights are also licensed under MIT. That matters. It means this is not just another closed API product. Developers can inspect, host, adapt, and build around the model in ways they cannot with most frontier systems from OpenAI, Anthropic, or Google.19

The simple comparison looks like this:

OpenAI’s GPT-5.5 API docs describe it as a frontier model for complex professional work, with configurable reasoning effort, a 1,050,000-token context window, text and image input, and pricing of $5 per million input tokens and $30 per million output tokens.20

Anthropic’s pricing docs list Claude Opus 4.7 at $5 per million input tokens and $25 per million output tokens, and say Opus 4.7, Opus 4.6, and Sonnet 4.6 include the full one-million-token context window at standard pricing. Anthropic also describes Sonnet 4.6 as a major upgrade across coding, computer use, long-context reasoning, agent planning, knowledge work, and design, with pricing from $3/$15 per million tokens.21

So V4 is not arriving in an empty field.

The leading closed labs have also moved to very long context. OpenAI and Anthropic are pushing hard into coding, agents, computer use, enterprise workflows, and tool use. The interesting part is not that DeepSeek has a one-million-token context window. The interesting part is that DeepSeek is trying to make that capability cheaper, more open, and available on a Chinese hardware stack.

4. Plain English: tokens, memory, and why 1M context matters

A one-million-token context window sounds impressive. It also sounds abstract.

In plain English, it means the model can hold a lot more material in view at once: a codebase, a long legal file, dozens of papers, a large company knowledge dump, or the history of a complex agent task.

Plain English: what is a token?
A token is the small unit of text an AI model reads. It can be part of a word, a full word, punctuation, or a space.

OpenAI’s rough rule for English: 1 token is about 4 characters, or about three-quarters of a word. So 1 million tokens is roughly 750,000 words.

In normal terms: enough room for several books, a large legal file, a meaningful part of a codebase, or the working memory of a long agent task.22

“Token” brings us, unavoidably, to Tolkien. The full Lord of the Rings trilogy has about 576,000 words. So 1 million tokens is roughly enough room for the trilogy — plus space left for the appendices, family trees, and whatever else the elves insisted on documenting.

But long context is expensive.

The model has to keep track of what came before. That creates memory and compute costs. In transformer models, a key part of that cost sits in the attention mechanism and in something called the KV cache. You do not need to love the term. You just need to understand the consequence: the longer the context, the more expensive it becomes to keep the conversation alive.

This is where DeepSeek V4 is interesting.

DeepSeek says that, at one-million-token context, V4-Pro uses only 27% of the single-token inference FLOPs and 10% of the KV cache compared with DeepSeek V3.2. That is the kind of detail that matters more than the usual benchmark table. It means DeepSeek is not only trying to make the model smarter. It is trying to make long-context work cheaper to run.23

That matters for agents.

A chatbot can answer a question. An agent has to keep track of goals, files, tools, mistakes, prior steps, and unfinished work. It needs memory. It needs persistence. It needs to carry context across a task without losing the plot after six turns and confidently driving into a wall.

Very human, in that sense.

Long context does not solve this alone. But cheaper long context makes more workflows economically possible. Coding agents can inspect larger repositories. Research agents can work across more documents. Enterprise agents can keep more business context in view. Legal, finance, and operations tools can work with full files instead of chopped-up fragments.

This is why the technical story of V4 is not “bigger model”. It is cheaper working memory.

5. The benchmark circus

DeepSeek’s own materials say V4-Pro leads current open models in several areas and rivals top closed models in reasoning, math, STEM, coding, and agentic work. That is useful, but it should be read as a vendor claim first. Every lab now launches with a benchmark bouquet. Some of the flowers are fresh. Some are plastic.24

The Economist adds a useful wrinkle to the “shrug” story. DeepSeek did not publish a training-cost estimate for V4, unlike the widely discussed $6m figure around its earlier models. That omission matters. The 16-month gap between V4 and its predecessor also suggests that the new model may have required much more compute than the original DeepSeek shock implied. And while V4 looks strong, it does not appear to have leapt ahead of the latest U.S. frontier models. According to The Economist’s reading of DeepSeek’s own tests, V4-Pro only comes close to leading American models from roughly three to six months ago. In AI time, that is not nothing. Meanwhile China’s AI market has become more crowded: Alibaba’s Qwen, ByteDance’s Doubao/Dola, Moonshot and Z.ai are all fighting for attention. The bigger Chinese internet companies are also pushing AI into applications, commerce, and “super app” workflows. Clever models alone are no longer the full story. The money is moving toward the layer where AI touches users, transactions, and daily work.25

Compare with Chapter 2, The graph “benchmark and efficiency comparison figure”: DeepSeek’s own chart mainly supports the claim that V4 is cheaper to run at long context. The Economist’s point is different: V4 may have been expensive to train, and its performance appears closer to leading U.S. models from three to six months ago than to the newest frontier. Cheap inference is not the same as cheap training. Efficient is not the same as leading.

Artificial Analysis says DeepSeek V4-Pro is back among the leading open-weight models, with strong gains versus V3.2 and leading open-weight performance on some agentic work tasks. But it also says V4-Pro and V4-Flash show very high hallucination rates on its omniscience-style evaluation: 94% and 96% respectively when the model does not know the answer.26

That is not a small footnote.

For consumer chat, hallucination is annoying. For enterprise workflows, it is structural risk. A model that answers confidently when it should stop is a problem in law, finance, healthcare, compliance, security, and any workflow where “sounds plausible” is not an acceptable operating standard.

This is where Claude and OpenAI still have an advantage in many enterprise contexts.

Not necessarily because their raw model capability is always better on every benchmark. The gap is broader: safety systems, enterprise trust, procurement familiarity, product integration, compliance posture, tooling, and support. Boring things. Expensive things. Things that make the difference between a demo and deployment.

DeepSeek pressures the price of intelligence. It does not erase the cost of trust.

So where does this leave us?

The first DeepSeek moment was easier to read.

V4 is harder to read.

The market reaction was calmer, but the setting is heavier: Huawei chips, export controls, China’s domestic AI stack, the Manus case, and fresh pressure on the American side of the AI trade as investors start asking harder questions about compute commitments and revenue.

DeepSeek V4 can still be over-read.

It is a preview. Benchmarks are incomplete. Open weights do not make enterprise deployment easy. One million tokens do not guarantee understanding. A model can hold more text in view and still answer with confidence when it should stop.

The legal and political fog also remains. Reuters reported that the U.S. State Department ordered diplomats to warn foreign governments about alleged AI theft by DeepSeek and other Chinese firms. China rejected the accusations as baseless, and DeepSeek has previously said it did not intentionally use synthetic data generated by OpenAI.27

So the useful conclusion should stay modest. DeepSeek V4 does not prove that China has caught up. It also does not prove that the U.S. compute boom is broken.

It shows that the AI race is moving into a more serious phase.

Models still matter. So do chips, energy, software, developers, cloud contracts, pricing, regulation, trust, and distribution. The model gets the headline. The stack carries the weight.

For leaders, this makes the job more demanding. AI is no longer a feature question. It is becoming a strategy, infrastructure, talent, risk, and execution question at the same time. That is also why we are launching Backstage briefs next Tuesday: to take these shifts out of the essay format and translate them into decisions, checklists, memos, and operating moves.

🎚️🎚️🎚️🎚️ Producer’s Note

With 2.06B equivalent units, a Korean-American pop collaboration became the world’s biggest single of 2025. Talent, production, platforms, culture, distribution, and timing moving together. Apparently “just write a good song” was not the full strategy.28

The model gets the headline, but the stack decides how far it travels: chips, software, developers, cloud, pricing, trust, and distribution.

Subscribe now

Fab

LinkedIn | Insta | Twitter | Studio⍺

StudioAlpha Capital is a Delaware-structured pre-seed venture fund backing AI-native B2B software startups at day zero. Legal counsel: Cooley LLP. Fund administration: AngelList.

Sources

7

Dwarkesh Podcast

Jensen Huang – TPU competition, why we should sell chips to China, & Nvidia’s supply chain moat

Listen now

2 months ago · 367 likes · 57 comments · Dwarkesh Patel

9

Dwarkesh Podcast

Jensen Huang – TPU competition, why we should sell chips to China, & Nvidia’s supply chain moat

Listen now

2 months ago · 368 likes · 57 comments · Dwarkesh Patel

12

Dwarkesh Podcast

Jensen Huang – TPU competition, why we should sell chips to China, & Nvidia’s supply chain moat

Listen now

2 months ago · 368 likes · 57 comments · Dwarkesh Patel

13

Dwarkesh Podcast

Jensen Huang – TPU competition, why we should sell chips to China, & Nvidia’s supply chain moat

Listen now

2 months ago · 368 likes · 57 comments · Dwarkesh Patel

If Beatles vs. Stones was the old format for public rivalry, this is the AI version: different style, same pressure.

This blog is about what these two models actually are, what is going on inside the companies building them, and why this moment — what the Economist called the ‘Mythos moment’ — may mark a real shift in how frontier AI is handled.

Subscribe now

1. The models: what Mythos and GPT‑5.4‑Cyber actually do

Start with the basic question: what are these things?

Mythos is Anthropic’s newest model. Anthropic announced it on April 7th and, unusually, also announced that the general public would not be getting access to it. The reason is not that it is buggy or unfinished. The reason is that it is good at finding and exploiting security vulnerabilities in software — operating systems, web browsers, cryptographic libraries, cloud infrastructure — at a level that Anthropic believes would put critical infrastructure at risk if released openly.1

Anthropic says Mythos has already found thousands of high- and critical-severity bugs across major software projects — including one in FreeBSD (a widely used operating system), one in FFmpeg (a video library used everywhere), and one in cloud computing infrastructure that remains unfixed. For context, an older, pre-Mythos version of Claude found nearly 20 percent of all high-severity Firefox bugs fixed in 2025.2

The benchmark number that caught the most attention: on advanced multi-step cybersecurity tests — the ones that require a model to execute dozens of sequential steps to successfully take control of a target machine — Mythos was noticeably ahead of all other models tested, including older Claude versions and leading OpenAI models. On simpler tests, the gap was smaller. But at the hard end, Mythos was in a different category.3

GPT‑5.4‑Cyber is OpenAI’s answer, announced four days later. It is a tuned version of GPT‑5.4 built explicitly for cybersecurity work: penetration testing, vulnerability triage, exploit reproduction, and binary reverse-engineering — which means it can analyse compiled software executables, not just source code, to find what an attacker could do with them.4

On OpenAI’s own internal benchmarks, the GPT‑5 family has improved sharply on cybersecurity tasks — from around 27 percent on CTF-style tests with early GPT‑5 variants, up into the mid-70s with later versions. GPT‑5.4‑Cyber is aimed at the hardest end of that curve — multi-step, multi-system exploit reasoning.5

Neither model is publicly available. Both sit behind gated access programs:

• Anthropic’s is called Project Glasswing — about 50 critical-infrastructure companies: Apple, Google, NVIDIA, JPMorgan, Microsoft, Cisco, CrowdStrike, the Linux Foundation and others. Anthropic is also contributing $100 million in usage credits and direct funding to open-source security projects so those teams can use Mythos to find and fix bugs before attackers do.6

• OpenAI’s is called Trusted Access for Cyber — vetted security vendors, research organisations, and infrastructure companies, applying via contract and audit, with OpenAI explicitly framing this as preparation for “more capable models over the next few months.”7

The pattern from both companies is identical:

More cyber capability, more gates, fewer users.

That is new. Two months ago neither program existed. Something changed.

2. The companies: same origin, very different machines

To understand why both companies ended up in the same place — gated access, dangerous models, Washington meetings — you need to understand where they came from and what they have become.

The split

Both companies trace back to the same room.

Dario Amodei was VP of Research at OpenAI. In early 2021, he left — along with his sister Daniela and roughly 15 other colleagues — because he believed OpenAI was moving too fast and not taking safety seriously enough. They founded Anthropic specifically to build what they believed OpenAI would not: a lab where safety was not a side constraint but the central question.8

So what you are watching with Mythos vs. GPT‑5.4‑Cyber is not just a product race. It is the original disagreement — how fast is too fast, who should control access — playing out in two competing product decisions made in the same week, five years later. That context makes the convergence on gated access all the more striking. They started from different philosophies and ended up in the same place.9

The revenue race: both at $30 billion

Here is the number that most people buried under the Mythos news cycle.

In March 2026, Anthropic’s annualised revenue hit $30 billion — up from $9 billion at the end of 2025, $4 billion in mid-2025, and just $1 billion at the end of 2024. OpenAI crossed $25 billion annualised in early 2026 and has also been reported at roughly $30 billion by end of March.10

Two companies that barely existed five years ago, now both converging on $30 billion annualised. The gap that looked enormous twelve months ago has essentially closed.

It took OpenAI roughly five years to reach $1 billion in revenue. Anthropic reached scale far faster than OpenAI did in its early years, then appears to have accelerated again from roughly $1 billion to around $30 billion annualised within about 15 months.

If that trajectory holds, some investors believe Anthropic could exit 2026 at a far higher run-rate still:

2022: $0

2023: $100M

2024: $1,000M

2025: $10,000M

2026E: $80-100,000M

If Anthropic gets there by end of the year, 10x growth four years in a row would be extraordinary.11

Where the money comes from

The headline revenue numbers now look far closer than most people expected. But the businesses underneath are still very different.

Anthropic is building a concentrated enterprise machine. In February, the company said about 80 percent of its revenue run-rate came from enterprise customers, and that more than 500 customers were already spending over $1 million a year on its products. That is a high-value business software story.12

A big part of that momentum appears to come from coding. Anthropic’s recent growth has been driven in large part by Claude Code, which has become one of the company’s most important commercial products. That matters because coding revenue is different from casual chatbot usage. Once a model becomes part of an engineering workflow, usage can become sticky, budgeted, and deeply embedded.13

OpenAI is built very differently. It monetises at several layers at once: consumer subscriptions, business and enterprise plans, API usage, and now an early advertising business inside ChatGPT’s free tier. In other words, OpenAI is not just selling one product to one type of customer. It is running a broader commercial stack.14

That broader mix gives OpenAI a very different profile from Anthropic. Anthropic is more concentrated, especially around enterprise and developer spend. OpenAI has more routes to monetisation: the paid consumer seat, the enterprise seat, the API call, and potentially the ad impression. Anthropic is narrower, but that narrowness may also be part of its strength. Enterprise revenue is harder to win, but often harder to lose once it is tied into real workflows.15

This is also why raw user numbers can mislead. ChatGPT still dominates mainstream consumer distribution. As we noted in a previous blog a16z’s March 2026 data showed ChatGPT at roughly 900 million weekly active users, while Claude remained much smaller on web and especially on mobile. Those charts are useful, but they mostly tell you about reach. They do not tell you where the dollars come from.16

The more interesting point is that Anthropic has managed to turn a much smaller user base into a revenue base that is now far closer to OpenAI’s than consumer usage alone would suggest. That is what makes the comparison interesting. OpenAI still dominates distribution. Anthropic has built a much more concentrated enterprise and developer business than many people realise.17

Valuation: big numbers, bigger gap

OpenAI’s current valuation sits at approximately $852 billion, making it one of the most valuable private companies ever built. Anthropic closed its Series G in February 2026 at a $380 billion valuation — raising $30 billion in a single round, the second-largest private tech raise in history.18

One investor who has backed both companies told the FT that justifying OpenAI’s valuation now requires assuming a post-IPO value of $1.2 trillion or more — which makes Anthropic’s $380 billion look, for the first time, like the relative bargain in the space.19

Headcount: the most striking number of all

This is where the comparison gets almost uncomfortable.

OpenAI currently has around 4,500 employees and is planning to nearly double to 8,000 by the end of 2026 — a major expansion across product, engineering, research and sales.

Anthropic has somewhere around 1,500 to 3,000 employees, depending on the source — generating $30 billion in annualised revenue.20

Do the rough math: at even the high-end headcount estimate, Anthropic is generating around $10–12 million in revenue per employee. The most efficient traditional enterprise software companies — Salesforce, SAP — sit at around $350,000–400,000 per employee. Anthropic is operating at roughly 30 times that ratio. 21

Investor Brad Gerstner noted that Anthropic added the equivalent of Databricks and Palantir combined in a single month of revenue growth, and said he would not be surprised if Anthropic exits 2026 at $80–100 billion in annualised revenue.22

The people running them

Sam Altman built OpenAI into a mass-market product company faster than almost anyone thought possible. ChatGPT was the fastest consumer product to 100 million users in history. Altman is a maximalist: more users, more products, more partnerships, more compute. The SoftBank round — $40 billion in early 2025, the largest private tech raise ever — is very much in his style.23

Dario Amodei built Anthropic as a research-first, safety-first lab that happened — almost reluctantly — to become a business. His public writing, particularly the 2024 essay “Machines of Loving Grace,” is the most serious attempt by any AI CEO to think through what genuinely powerful AI actually means for humanity. He is the person who decided Mythos was too dangerous to release. That decision is completely in character.24

The personality difference matters because it explains the product difference. Altman ships to scale. Amodei ships to control. Mythos vs. GPT‑5.4‑Cyber is what happens when both approaches arrive at a capability that forces the same conclusion: this one does not go to everyone.

3. Why Washington is now in the room

The Mythos announcement did not stay inside tech Twitter for long.

Within days, U.S. Treasury Secretary Scott Bessent and Fed Chair Jerome Powell jointly summoned the CEOs of the largest American banks for an urgent closed-door briefing. The topic: what AI-enabled hacking means for financial system stability. It was not a routine meeting. Treasury secretaries do not typically convene bank CEOs to discuss software releases. The fact that this one happened tells you how the model’s capabilities were received at the top of the U.S. government.25

Separately, the White House budget office began preparing a path to give federal agencies controlled access to Mythos — while simultaneously acknowledging in internal communications that the same model, mishandled, could sharply increase federal cyber risk. The UK’s Financial Conduct Authority organised a parallel meeting with British banks the same week.26

The BBC reported on a “productive” White House–Anthropic meeting, with officials describing the interaction as collaborative while acknowledging deeper discussions about what kind of regulatory framework should govern models like Mythos going forward.27

This is not the first time the Trump administration has had to engage with Anthropic in an uncomfortable way. Weeks earlier, the Pentagon stepped in after Dario Amodei declined to allow Claude to be used in fully autonomous weapons systems or for mass domestic surveillance. The administration was unnerved — not because it disagreed with Amodei, but because a single private company was making unilateral decisions about technology central to national security.28

The broader political context is shifting fast. Roughly seven in ten Americans now believe AI will hurt job opportunities — up sharply from a year ago. Grassroots opposition to data centres is rising. Sam Altman’s home has been targeted by activists twice in recent weeks. AI has crossed from a technology story into a political one. And Mythos arrived at exactly the moment Washington was looking for a reason to act.29

The emerging policy response — still improvised, still contested — points toward a gated-access model rather than a new regulator: trusted users get early access to the most powerful capabilities; broader commercialisation requires certification; open-source and foreign models remain an unresolved problem. It avoids the slow process of building new institutions. It lets model makers charge premium prices and control compute. It gives governments a lever they can pull without legislation. It also concentrates power in a small inner circle and creates exactly the lobbying opportunities you would expect.30

The honest counterargument to all of this: it is not obvious that gated access actually works. Security researchers have already shown that smaller, older models can reproduce some of the same bugs Mythos found. Open-source models are months behind the frontier, not years. And once a capability exists, keeping it inside fifty companies is harder than it sounds. The labs may be only months ahead of their open-source competitors — which means the window for “controlled diffusion” is shorter than anyone in Washington seems to assume.31

NVIDIA’s Jensen Huang is reading the same Mythos moment very differently from Dario Amodei. In his recent long-form interview with Dwarkesh Patel, Huang calls Mythos a “step change” that should trigger more US–China dialogue on AI safety, not panic or paralysis, and warns that China already has the energy, chips, and researchers to train Mythos‑class systems despite export controls. He pushes back hard on what he calls “doomer” narratives that scare the public, arguing instead that AI should be treated as critical infrastructure — like electricity or the internet — that governments must secure and guide, but not freeze.32

🎚️🎚️🎚️🎚️ Producer’s Note

Anthropic and OpenAI are pushing each other hard. One leans more into enterprise discipline and controlled rollout. The other leans more into scale, distribution, and consumer reach. Different style, same effect: both are improving faster because neither is alone.

That is not unusual. The Beatles and the Rolling Stones were not copies of each other, and the rivalry was never as clean as the press liked to pretend. But competition helped sharpen both bands. Same here. Rivals are often bad for your nerves and good for your output.

The interesting point is not who is “winning” this week. It is that serious competition is now producing two very different AI companies at very high speed. Annoying for management teams, no doubt. Useful for the rest of us.

Pressure, inconveniently, tends to improve performance.

Share

Fab 🎸
LinkedIn | Insta | Twitter | Studio⍺

StudioAlpha Capital is a Delaware-structured pre-seed venture fund backing AI-native B2B software startups at day zero. Legal counsel: Cooley LLP. Fund administration: AngelList.

Sources

The noise around this is rather something. Europe says Trump is turning NATO into a tool for American wars of choice. Trump says Europe still wants protection without real burden-sharing. Both sides are performing. Both sides also have a point. So before joining the shouting, it is worth asking a simpler question: what, exactly, would break if America really stepped back?2

Subscribe now

1. Why NATO was created

America had already paid heavily for European collapse. World War I cost the United States about $32 billion at the time — roughly $693 billion in 2026 dollars — and 116,516 U.S. military deaths. World War II cost about $330 billion in 1945 dollars — roughly $6 trillion today — and more than 400,000 U.S. military deaths worldwide. Those are total U.S. war costs and deaths, not Europe-only.3

NATO was built because Washington had already learned, twice, that European instability does not stay in Europe. After two world wars — WWII alone killed roughly 36.5 million Europeans, 19 million of them civilians — the United States concluded that a stable Europe was a core U.S. interest, not a favour to others. That is why NATO tied North America to Europe in 1949. NATO now includes 32 countries across Europe and North America.4

That starting point matters.

Because a lot of today’s debate sounds as if NATO were just a loose political friendship that can be re-priced every few years. It was never that. It was a hard strategic arrangement: America helps secure Europe, and Europe remains anchored inside a U.S.-led security order.

That bargain then became so normal that many people stopped seeing it as a bargain at all.5

2. The old bargain: American hard power, European underinvestment — and European underpreparation

Trump is right on the money.

For decades, many European allies benefited from U.S. hard power while spending too little on their own defence. NATO’s own numbers are clear enough. In 2014, only 3 allies met the 2% of GDP guideline. By 2024, 22 did. By 2025, for the first time, all allies were at or above 2%. That is real progress. NATO’s 2025 data also show that Europe and Canada increased defence spending by 20% year-on-year. The United States still carried the largest share by far, but its share of total NATO defence spending edged down slightly — from about 64% in 2024 to roughly 62% in 2025 — as the rest of the alliance paid more. Europe was late, and it was late for a long time.6

Defence expenditure

(billion US dollars, based on 2021 prices and exchange rates)

But money is only half the story.

Europe did not just underinvest. In key countries, it also underprepared. The harder problem is not writing a bigger cheque. It is turning money into combat power: troops, stocks, air defence, artillery, logistics, command structures, and forces that are actually ready when the weather is bad and the situation is worse.7

Germany is the clearest example.

Reuters reported in February 2025 that the Bundeswehr was only about 50% battle-ready, with major gaps in air defence, artillery, manpower, and procurement. In May 2025 Reuters also reported that NATO wanted Germany to add seven more brigades under new capability targets, even though the Bundeswehr was already short around 20,000 troops. That is not a small management issue. That is a warning sign for the largest economy in Europe.8

And sometimes the warning signs are embarrassingly small.

In 2022, the parliamentary commissioner Eva Högl said German troops deployed in Lithuania lacked adequate cold- and wet-weather protection, including thick jackets and suitable underwear. That may sound like a minor story. It is not. If a military cannot get basics right on NATO’s eastern flank, then bigger speeches about strategic autonomy deserve at least one raised eyebrow.9

That is the deeper problem in Europe.

The continent did not just spend too little for too long. It also allowed readiness, procurement discipline, and military seriousness to decay at the same time. Buying a few more players does not automatically give you a team ready for a final. Europe’s defence problem is not only financial. It is operational.10

Germany is not the whole story. But as Europe’s largest economy and a central NATO player, it is the most revealing case. If the locomotive has weak brakes, people are allowed to look at the locomotive. The picture is not uniform. Poland is rearming much faster, and frontline states such as Finland and the Baltic countries generally show a sharper threat perception than much of Western Europe. But sharper threat perception is not the same as proven wartime readiness. That test remains largely unrun without the United States in the lead.11

EU Member State defence expenditure, 2024, in % of GDP

3. What NATO does — and does not — require

This is where the debate usually gets sloppy.

NATO is a collective-defence alliance. Its core logic is Euro-Atlantic defence. It is not a standing obligation for allies to join every U.S. military action outside that core area. That distinction matters now because the current clash is not only about spending. It is also about whether Washington is trying to stretch NATO into something broader and more transactional.12

The strongest corrective fact here is simple.

Article 5 has been invoked only once: after 9/11, for the United States. European allies did support America after that attack. NATO support included access, overflight, AWACS patrols, and then a long military commitment in Afghanistan. So the claim that Europe simply “wasn’t there for us” does not hold up historically.13

At the same time, Europe cannot use that fact to dodge the other problem.

It can be true that Europe showed up after 9/11.
And it can also be true that Europe underinvested in defence for too long.

Those two facts sit together. That is exactly why this topic is worth cleaning up.14

Even inside Europe’s own defence structures, the distinction is clear. The EU can train, support, move, and stabilise. It can also run limited missions such as Aspides in the Red Sea. But even senior EU military voices do not describe that as a substitute for NATO’s job in a major European war. The lane is narrower — and everyone serious knows it.15

4. The perception problem: Europe on America, America on Europe

Part of this story is strategic.
Part of it is psychological.
But only the first part is easy to measure.

On the European side, anti-American sentiment is real, but it is not one simple continental emotion. Pew’s 2025 data show that views of the United States worsened in many countries, and majorities in several European countries viewed the U.S. negatively. But the pattern is uneven. Poland and Hungary, for example, remained much more favourable than countries such as Germany, France, Spain or Sweden.16

That matters because “Europe is anti-American” is too crude.

A better line is that Europe often depends on America, resents America, admires America, and distrusts America — sometimes at the same time. Pew’s 2025 “ally or threat” work captures that contradiction well: in some countries the U.S. is still seen as a key ally; in others it is simultaneously seen as a major threat.17

On the American side, the picture is different.

The clean survey signal is not that Americans broadly dislike Europe. It is that many Americans still view key European allies positively while also believing those allies should do more for their own defence. That is a more precise and more useful line than trying to force a blanket claim of anti-European sentiment. The baseline is still allied. The patience is thinner.18

About half of Americans say Europeans should increase their defense spending

So the better reading is bilateral ambivalence.

Europe is less trusting of U.S. leadership.
America is less willing to subsidise European strategic comfort.

That is a very different problem from simple mutual dislike.19

5. The scenarios

The real question is no longer only whether Europe should spend more. It is whether Europe can replace what actually matters if the U.S. steps back: leadership, enablers, credibility, and ultimately deterrence. That is why the scenarios matter. The risk is not only a formal U.S. exit from NATO, but a slower hollowing-out of the alliance long before any paperwork changes.20

The same structural weaknesses run through all three scenarios: weaker growth, weaker innovation, weaker state capacity, and Europe’s continued dependence on U.S. leadership and enablers. What changes from Scenario 1 to 3 is not the nature of the problem, but how hard and how fast it hits.

Scenario 1 — NATO survives, but becomes more transactional

In this scenario, NATO stays formally intact.

There is no dramatic U.S. exit. No grand funeral. No one tears up the treaty on live television. But the alliance becomes more conditional, more rhetorical, and less instinctive. Europe keeps spending more. The U.S. keeps demanding more. Trust degrades. Coordination becomes slower and more political.21

This is probably the base case.

It fits the current evidence best because it does not require a legal rupture. It only requires a change in behaviour and expectations. That is already visible. Reuters has reported U.S. threats over NATO in the context of Iran, and Finland’s Alexander Stubb has explicitly spoken about a “more European NATO” taking shape.22

What would trigger it?

Not one single event.
More likely a continued pattern:23

• U.S. pressure

• European pushback

• higher European defence spending

• weaker political trust

• more explicit bargaining over who does what

What would it mean?

Ranges of selected European-operated or planned cruise missiles

NATO remains alive, but the relationship becomes more transactional and less automatic. That would still matter a lot for Europe. A more transactional alliance means more pressure on defence budgets, procurement, cyber, energy resilience, logistics, and dual-use infrastructure. The alliance would continue, but the comfort level would fall.24

Scenario 2 — de facto hollowing-out

This is the most important scenario to take seriously.

In this version, the U.S. stays inside NATO formally, but the practical guarantee weakens. Europe can no longer assume the same level of automatic American backing. The treaty remains. The credibility erodes. That may matter more than formal withdrawal, because deterrence depends on what adversaries believe, not only on what lawyers can cite.25

This is where the expert work becomes useful.

Brookings argued that NATO could be damaged badly even without formal U.S. withdrawal if an administration reduced troop presence, weakened Article 5 credibility, or treated the alliance as expendable in practice. Brookings also noted that European militaries were largely built to plug into U.S.-led coalition structures, not to operate fully independently at scale.26

IISS takes the same issue from the capability side.

Its 2025 assessment of a Europe defending itself without the United States concludes that Europe could reduce dependence, but doing so would require large, slow, and expensive replacement of U.S. capabilities. In other words: possible over time, but not cleanly, cheaply, or fast.27

There is another problem here.

Economic base: Europe would be trying to replace American military weight from a weaker economic base, not a stronger one. Jamie Dimon put it bluntly in 2025: Europe is losing ground. He said Europe had gone from about 90% of U.S. GDP to roughly 65% over the previous 10 to 15 years. Reuters, citing IMF work, reported in late 2024 that Europe’s GDP per capita at purchasing power parity was about 72% of the U.S., with most of the gap explained by weaker productivity growth. The IMF said again in March 2026 that Europe’s productivity gap with the U.S. has widened in recent years. So this is not just a defence problem. It is a growth problem too.28

Innovation: There is another weakness here, and it is not only military. Europe’s problem is also innovation. A major driver of innovation comes from startups. The continent still struggles to turn research into companies, companies into scale-ups, and scale-ups into global winners. The IMF says Europe’s productivity gap with the U.S. is partly a failure to develop innovative startups into “superstar” firms. The OECD says fragmented capital markets, internal barriers, and regulatory burden are still holding back innovation and scale. The Commission’s own biotech strategy makes the point brutally: over the last six years, 66 of 67 EU biotech companies that went public chose non-EU exchanges. So if Europe had to replace more of America’s military role, it would be trying to do so with a weaker innovation engine than the U.S., not a stronger one.29

Military enablers: RUSI makes this sharper still. Its recent analysis points to the real problem not as a vague “loss of America,” but as the loss of specific enablers: command-and-control, intelligence, surveillance, transport, air defence, deep-strike support, and other high-end capabilities Europe still struggles to replace quickly. That is what hollowing-out would actually look like in practice.30

Money helps. It is not enough: That is why bigger budgets alone do not settle the argument. Money helps. It does not automatically create leadership, command structures, readiness, manpower, industrial capacity, or real warfighting power. Europe’s problem is not only that it spent too little for too long. It is that it also let military seriousness, procurement discipline, and execution decay at the same time.31

This is also where a European counterpressure voice matters.

Credibility: François Heisbourg, one of the most seasoned French strategic analysts, warned this week that what is new is not merely another row over burden-sharing, but the open U.S. discussion of NATO as expendable. That does not prove collapse. But it does support the argument that the credibility problem is now out in the open.32

What would trigger this scenario?

A continued formal U.S. membership combined with:33

• reduced U.S. troop presence

• weaker public commitment to Article 5

• more conditional military support

• European planning that starts assuming American help may arrive later, in smaller form, or not at all

What would it mean?

This is where markets may move before treaties do. If Europe starts planning around a less reliable America, the question is no longer only how much more it will spend. The question is whether it can convert weaker growth, slower procurement, and thinner state capacity into actual deterrence. That is a much harder test.

Europe would not just be filling a budget hole. It would be trying to replace the alliance’s dominant organiser, key enablers, and much of the credibility that made deterrence work in the first place.34

Scenario 3 — formal or near-formal U.S. disengagement

This is the dramatic case.

It is still harder legally than many headlines suggest. Reuters reports that U.S. law passed in 2023 says a president cannot withdraw the United States from NATO without either a two-thirds Senate vote or another act of Congress. So a clean one-man exit is not straightforward.35

But politically, the unthinkable is now thinkable.

That is the point. Not that formal withdrawal is imminent, but that it has moved from fringe speculation into serious scenario planning. Reuters, WSJ, and current European commentary all point to that shift.36

Here ECFR is useful.

Its recent work argues that Europe should plan to replace critical U.S. contributions to NATO plans and build structures that can operate with or without American support. That is not the language of a continent expecting the old bargain to continue untouched. It is the language of contingency planning.37

Europe would also be facing this shift against an opponent that is still operating like a warfighting state.

Reuters reported in February 2026 that Russia’s fossil-fuel export revenues over the previous 12 months were still about €193 billion. Lower than before the war, yes. But still large enough to keep financing a serious military effort. So the problem is not only that Europe could lose the U.S. backstop. It could lose it while facing an adversary that still has war experience, military mass, and meaningful energy cashflow.38

What would trigger this scenario?

A much sharper political rupture:39

• explicit White House move against NATO

• deeper Congressional confrontation

• accelerated U.S. force reductions

• a wider collapse in transatlantic trust after repeated strategic clashes

What would it mean?

Europe would be pushed into a much larger rearmament and industrial mobilisation cycle. Defence budgets would rise further. Fiscal trade-offs would get harder. Procurement would stop being a niche issue and become a central political arena.

But money alone would not solve the problem.

As noted in Scenario 2, Europe would be trying to rearm from a weaker economic base than the U.S. In this scenario, that weakness would matter even more because the final American backstop would be disappearing at the same time. So this would not just be a military test. It would also be a growth and competitiveness test.40

And Europe would be trying to do all this from a weak execution base.

The military problem would not arrive on a clean balance sheet. It would arrive on top of older maintenance, procurement, and state-capacity problems that parts of Europe have tolerated for years. The FT’s reporting on Germany’s F126 frigate project reads like a cautionary tale for European rearmament: software failures, delays, cost overruns, paper-based approvals, English-language submissions rejected, and a programme now at risk of leaving a capability gap just as NATO wants more Baltic defence. Rearmament is easy as slogan. It is much harder as engineering, procurement, and execution.41

The same goes beyond defence procurement. Reuters reported in 2025 that Germany was falling behind on bridge renovation goals, and in 2026 that Deutsche Bahn’s rail overhaul could take a decade and cost around €150 billion. Italy’s Genoa bridge collapse in 2018 was a different case, but it points to the same broader truth: infrastructure decay is not just an inconvenience. It is a warning about maintenance, execution, and the state’s ability to keep complex systems working over time.42

So in this scenario, Europe would not just be filling a military gap. It would be trying to rebuild deterrence while also fixing weaker growth, weaker innovation, weaker infrastructure, and weaker state capacity at the same time. That is why formal U.S. disengagement would be so much more than a diplomatic crisis. It would be a stress test of whether Europe can still act at the level it likes to talk about.43

🎚️🎚️🎚️🎚️ Producer’s Note

For years, Europe treated hard power like bad taste.
Too blunt. Too militaristic. Too unfair. Better to talk values, regulation, transition plans, and strategic autonomy in elegant documents with soft lighting.

That worked — until reality asked for delivery.

Now Trump is openly questioning NATO. Russia is still financing war. Europe is rearming, but also discovering that bigger budgets do not magically produce leadership, procurement discipline, missile defence, or bridges that hold.

What is actually left if the American steps back?

The answer is not nothing.
It is just a lot less impressive than the speeches suggested.

Yesterday, FT reported that hedge funds make record bets against European stocks.44 For founders, operators, and investors, that matters.
Because once deterrence, energy, logistics, procurement, and state capacity move from background noise to front-page risk, capital does not get to stay apolitical either.

The soundtrack is not European self-congratulation. It is closer to Sigue Sigue Sputnik: loud, synthetic, slightly toxic, and built on the suspicion that the future may be arriving in worse shape than advertised.

Share

Fab 🤞🏽
LinkedIn | Insta | Twitter | Studio⍺

Sources

I had to rewind. Helium and Hormuz?
I knew helium was not just for balloons and Mickey Mouse voices. But I had not filed it under geopolitics, chokepoints, and oil markets having a nervous breakdown.

The deeper I looked into the Strait of Hormuz, the less this felt like an oil story and the more it looked like a narrow piece of water quietly holding together a surprising amount of modern life.

Easy to judge from a distance. Harder if you are the one in the chair. Quick anonymous poll:

Twenty-one miles.

That is enough to connect gas bills, chip fabs, MRI machines, fertilizers, shipping routes and the usual grand political statements that sound firmer from a podium than they do on a map.

So I tried to build the map.

Not the headline version. The real one: what moves through Hormuz, who can interrupt it, why Europe is more exposed than it likes to admit, and why these situations tend to attract a level of certainty the underlying system has not actually earned.

That is where this starts.

Subscribe now

1) A Body of Water the Width of Your Commute

The Strait of Hormuz separates Iran to the north from Oman’s Musandam Peninsula and the United Arab Emirates to the south. At its narrowest point, the strait measures approximately 21 nautical miles (roughly 39 km) — almost precisely the same width as the Strait of Dover, where the English Channel is narrowest.

But the strait’s actual navigable space is far tighter than its total width suggests. Under the International Maritime Organization’s Traffic Separation Scheme, commercial shipping is funneled into three narrow corridors: a two-mile-wide inbound lane, a two-mile buffer zone, and a two-mile-wide outbound lane. The total shipping corridor is roughly six nautical miles across. Through this compressed space, an estimated 60–80 oil tankers pass every day, alongside LNG carriers, container ships, and bulk freighters.2

Iran sits directly along the northern side of this corridor. Qeshm Island and a chain of Iranian-controlled positions give Tehran physical oversight of nearly all transit traffic. On the southern side, Oman’s Musandam Peninsula juts northward to create the narrowest point.3

Several disputed islands complicate the picture further. Iran seized Greater Tunb and Lesser Tunb in 1971 and occupies Abu Musa under a contested arrangement with Sharjah. These islands now host Iranian military installations and sit close to major shipping lanes.4

That is the physical setup.

A narrow waterway. A tiny shipping corridor. One side heavily militarized. And a large share of the global economy passing through it every day.

The strategic logic is older. Much older.

2) The Names Change. The Logic Doesn’t.

The Persian Gulf has been a trade corridor for at least 5,000 years. Long before oil, goods moved through these waters between Mesopotamia, India and beyond. By the medieval period, Hormuz had become one of the richest trading ports on earth, controlling traffic between the Gulf, India and China.5

Portugal understood this early. It attacked Hormuz in 1507, completed its conquest in 1515, and controlled the island for more than a century by taxing trade through the strait. In 1622, a combined Persian and English force drove the Portuguese out. Control shifted. The chokepoint did not.6

Britain became the dominant outside power in the Gulf in the 19th century, first to protect trade routes to India, later to secure a growing oil economy. When Britain withdrew “East of Suez” in the late 1960s and early 1970s, the modern Gulf states emerged. But again, the underlying logic stayed the same: whoever can shape access to Hormuz shapes far more than a local waterway.7

That logic turned military in the modern era. During the Iran-Iraq War, the so-called Tanker War saw both sides attack commercial shipping in and around the Gulf. Hundreds of ships were hit. The United States responded with Operation Earnest Will, reflagging Kuwaiti tankers and escorting them through the region.89

The lesson was simple. Once trade through Hormuz is threatened, outside powers return very quickly.

The US Fifth Fleet, based in Bahrain since 1995, is the latest version of that pattern. The Portuguese held Hormuz for a century. The British dominated the Gulf for roughly 150 years. The Americans have been there for decades.10

The brand changes. The franchise doesn’t.

This is not a history lesson. It is a structural observation. And it leads to the real question: what actually flows through Hormuz?

3) Not Just Oil: What Actually Flows Through Hormuz

Everyone knows about oil. Almost nobody knows about the rest. That’s the real risk.

Oil 🛢️: Approximately 20–21 million barrels per day of crude oil, condensate, and refined petroleum products flow through the strait — roughly 20–21% of global petroleum liquids consumption. Put differently, about one-third of all seaborne-traded crude oil passes through these six-mile-wide shipping lanes. No other chokepoint comes close. The Strait of Malacca handles roughly 16 mb/d. The Suez Canal and SUMED pipeline together carry 5–7 mb/d. Bab el-Mandeb — where Houthi attacks forced massive shipping reroutes in 2024–2025, adding 10–14 days to Europe-Asia voyages — handles 6–7 mb/d. The Panama Canal carries less than 1 mb/d of oil. Hormuz handles roughly three to four times the oil volume of any other single chokepoint and is, in the short term, irreplaceable.11

The dependency varies dramatically by country. Japan is the most exposed major economy: an estimated 80–90% of its crude oil imports transit Hormuz. South Korea follows at 70–80%, India at 60–65%, and China at 40–45% — though China’s absolute volumes are enormous, at roughly 4–5 mb/d. European dependency stands lower at 20–30% of crude imports, but is rising for LNG. The United States, now the world’s largest oil producer at over 13 mb/d thanks to the shale revolution, imports relatively little Gulf crude directly — but oil is a globally priced commodity, and any Hormuz disruption spikes prices for everyone.12

Existing pipeline bypasses offer cold comfort. The UAE’s Habshan-Fujairah Pipeline (1.5 mb/d capacity), Saudi Arabia’s East-West Petroline (5 mb/d to the Red Sea, itself vulnerable to Houthi interdiction via Bab el-Mandeb), the intermittently operational Kirkuk-Ceyhan pipeline through Turkey (1–1.6 mb/d), and Iran’s new Jask terminal (0.3–1.0 mb/d) together offer a theoretical maximum bypass capacity of roughly 7–9 mb/d — less than half of total strait flows. There is no viable full alternative to Hormuz.13

LNG 🔥: Qatar — one of the world’s top two LNG exporters alongside the United States — ships virtually all of its 80–85 million tonnes per annum through Hormuz. The UAE adds another 6 million tonnes. That’s roughly 20% of global LNG trade. With Qatar’s North Field East and South expansion projects targeting 126 Mtpa by approximately 2027–2028, Hormuz’s share of global gas trade is structurally increasing.14

Helium 🎈: This is the one that stopped me. Qatar produces approximately 25–30% of the world’s helium supply as a byproduct of LNG processing at the massive Ras Laffan Industrial City, operating two major liquefaction plants (Helium 1, commissioned 2005, and Helium 2, commissioned 2013) with a third under development. All of this helium — shipped in specialized cryogenic ISO containers at temperatures near –269°C — must transit the Strait of Hormuz.15

Helium is a non-renewable, non-substitutable resource for several critical industries. In semiconductor fabrication, it serves as a cooling gas for EUV lithography machines (the ASML systems at the heart of cutting-edge chip production), a carrier gas in chemical vapor deposition, and a leak detection medium in the ultra-clean vacuum environments that chip fabs require. In healthcare, liquid helium cools the superconducting magnets in roughly 50,000 MRI machines worldwide to four degrees above absolute zero. In scientific research, it enables everything from CERN’s Large Hadron Collider to the dilution refrigerators that cool quantum computers to millikelvin temperatures.16

The world has already experienced four helium shortages since 2006 — the most recent, “Helium Shortage 4.0” in 2022–2023, saw prices spike when Russia’s Amur Gas Processing Plant suffered delays, Qatari plants underwent maintenance, and the US Federal Helium Reserve continued its wind-down. A Hormuz closure would remove roughly a quarter of global supply instantaneously, with cascading consequences: semiconductor production slowdowns within weeks, MRI diagnostic capacity declining as helium boils off from machines, and quantum computing and particle physics research halting.17

Fertilizers 🌱: Qatar’s QAFCO is one of the world’s largest single-site urea producers at approximately 6 million tonnes per year, and Gulf states collectively supply a significant share of global urea and ammonia. This isn’t abstract: it’s food.18

Petrochemicals 🧪 and metals 🔩: Saudi Arabia’s SABIC alone generates over $40 billion in annual revenue from polyethylene, polypropylene, and specialty chemicals. The UAE and Bahrain together produce roughly 4 million tonnes of primary aluminum annually — about 5–6% of global output — through Emirates Global Aluminium and Aluminium Bahrain. Even sulfur, a critical input for fertilizers and battery production, flows through the strait in vast quantities as a byproduct of Gulf oil and gas processing.19

Here’s the uncomfortable summary. A Hormuz closure doesn’t just spike your fuel price. It threatens food production (fertilizers), manufacturing (petrochemicals, aluminum), healthcare (helium for MRI), and the entire semiconductor supply chain — simultaneously.

The strait is not an oil story. It’s a single point of failure for the modern economy.

4) The IRGC: A Party Army, Not a State Army

The force that matters most on Iran’s side of the Strait of Hormuz is not the regular Iranian military. It is the Islamic Revolutionary Guard Corps, or IRGC.

That distinction matters.

The IRGC is not a normal state army in the Western sense. It was built to protect the Islamic Revolution and the regime behind it. In practice, it functions as a parallel power structure inside Iran, with its own military branches, intelligence apparatus, and deep influence across the economy. It is a structural observation drawn from scholars including Afshon Ostovar (Vanguard of the Imam, Oxford University Press), who describes the IRGC as a "praetorian guard," and corroborated by RAND Corporation research estimating IRGC control of roughly one-third of Iran's economy through construction, oil and gas, telecoms, banking, and import-export networks.20

For this blog, the important point is simpler: the force sitting on Iran’s side of Hormuz is built for coercion, asymmetry, and disruption.

In the strait, that role falls mainly to the IRGC Navy, which is separate from Iran’s regular navy. Its doctrine is built around speed, concealment, and saturation: fast attack craft, naval mines, mini-submarines, and mobile anti-ship missiles spread across the coastline and nearby islands. It is not designed to defeat the US Navy in a conventional war. It is designed to make the strait dangerous, expensive, and difficult to keep open.21

It has shown repeatedly that this is not theory. In 2019, Iran seized the British-flagged tanker Stena Impero in the Strait of Hormuz. In 2023, it seized the Advantage Sweet in the Gulf of Oman and the Niovi in the Strait of Hormuz. In 2024, the IRGC seized the MSC Aries after boarding it from a helicopter and diverting it into Iranian waters.

The pattern is clear enough: Tehran has both the capability and the willingness to interfere with commercial shipping when tensions rise.22

The broader political context matters too. In April 2019, the United States designated the IRGC as a Foreign Terrorist Organization — the first time it had applied that label to part of another state’s military. In January 2023, the European Parliament called for the IRGC to be added to the EU terrorist list. That position later hardened: on 19 February 2026, the EU Council formally designated the IRGC as a terrorist organization.23

And then there is the domestic record. The same organization built to defend the regime externally has also played a central role in repression internally, including the violent suppression of protests after the death of Mahsa Amini in 2022 and the far larger crackdown that followed in 2026.24

If Hormuz is the chokepoint, the IRGC is one of the hands on the valve.25

5) Europe Traded Russian Dependency for Hormuz Exposure

After Russia’s invasion of Ukraine, the EU scrambled to replace Russian pipeline gas — which had supplied roughly 40% of EU gas in 2021. The replacement came substantially from US LNG and Qatari LNG — all of which transits the Strait of Hormuz.26

QatarEnergy signed long-term LNG supply deals with Germany and other European buyers including Shell, helping lock more Qatari gas into Europe’s energy mix. Europe’s LNG dependence has risen sharply since the Russia shock, and new Qatari volumes are part of that shift. At the same time, almost all LNG exports from Qatar and the UAE still have to pass through Hormuz.27

Europe didn’t diversify. It moved the risk.

Meanwhile, the United States became a net energy exporter around 2019 and now produces over 13 mb/d — the world’s largest output. This creates fundamentally divergent strategic interests. The US can absorb a Hormuz disruption far better than Europe or Asia. Yet even the US isn’t immune: oil is globally priced, and gasoline has already risen ~27% since January.28

Europe’s emergency buffer is thin. The EU mandates 90 days of strategic oil reserves. Its naval response — EMASOH, a French-led surveillance mission launched in 2020 — is a monitoring operation, not a combat fleet. It cannot clear mines. It cannot escort convoys. It is wholly inadequate for what is happening right now.29

And yet. In the weeks after the conflict escalated, both German Chancellor Friedrich Merz and French President Emmanuel Macron responded to the situation with the same phrase: “This is not our war.”

Let’s test that claim against what we just learned.

Europe’s homes are heated by gas that transits Hormuz. Its hospitals depend on helium that transits Hormuz. Its chip supply chains depend on helium that transits Hormuz. Its food production depends on fertilizers that transit Hormuz. Its LNG contracts are locked in for 27 years — through Hormuz.

The commodity flows don’t care about political declarations.

6) Five Ways to Think About the War — and Why the Market Still Thinks It Ends Soon

By this point, the basic map is clear.

Hormuz is not background. It is not just an oil route. It is a narrow systems chokepoint sitting inside a war that now forces a much harder question: if you accept that Iran cannot be allowed to become a nuclear-armed theocracy, what exactly are the least bad options?30

That is where the opinions split.

John Mearsheimer — R. Wendell Harrison Distinguished Service Professor of Political Science at the University of Chicago — sees the war through power, geography and escalation. His core argument is blunt: Iran holds more of the cards than Washington wants to admit. The reason is not ideology. It is structure. Iran sits next to the chokepoint. It can raise the cost of war faster than the United States can lower it. In his recent interviews, Mearsheimer’s line is that the US and Israel walked into a war of attrition without a clear off-ramp, and that the deeper the conflict goes, the more leverage shifts to Tehran through Hormuz, the Red Sea, and the global energy system — between them threatening roughly 32% of global oil and gas transit. In that frame, the war is not irrational. It has a logic. It is just a bad logic for Washington.31

Furthermore he noted that none of Trump’s predecessors took the bait when Israel tried to trap the US into war with Iran — Biden refused twice in 2024 — and that Trump is the first president who fell into it, starting with the “one and done” bombing of three nuclear targets on June 22, 2025, which was anything but one and done.32

Jeffrey Sachs — economist at Columbia University, author, and one of America’s most persistent critics of interventionist foreign policy — comes at the same crisis from a very different direction. Where Mearsheimer sees structural rivalry, Sachs sees policy failure, illegality and hegemonic overreach. His argument is that the war was avoidable, that it was chosen, and that it risks becoming economically catastrophic because the people running it are improvising inside an extremely tightly coupled system. Sachs is not subtle about it. He calls the war a war of choice, a breach of the UN Charter, and a path toward a much wider regional or even global disaster. His emphasis is less on balance-of-power logic and more on the collapse of disciplined statecraft.33

Then there is the harder counterpoint — the one many Europeans prefer not to say out loud. A number of mainstream US and Israeli security thinkers argue that however dangerous this war is, a nuclear-armed IRGC state could be worse. That is the Dennis Ross — William Davidson Distinguished Fellow — from the Washington Institute line in broad terms. It is not a chest-thumping “war is good” argument. It is a preventive argument. If diplomacy cannot permanently block Iran’s path to a weapon, then force, coercion, or regime weakening may become the least bad option. Ross’s own phrasing is more careful than the television version. He focuses on blocking all paths to a weapon, preserving leverage, and using military pressure as part of a wider strategy rather than as theatre. Dana Stroul adds the missing caution: direct strikes carry serious escalation risks, but previous presidents weighed those risks precisely because the nuclear threshold question is real. In plain English: this camp thinks the costs of action are high, but the costs of inaction may be higher.34

That matters because this is the strongest serious case for the war. Not that escalation is fun. Not that Trump is a strategic genius. Simply that a regime like this, sitting next to Hormuz, with a bomb, would be in a different category altogether. At that point the toll booth problem becomes much worse. Deterrence becomes murkier. Regional blackmail gets cheaper. And every shipping company, insurer, energy buyer and government planner has to price a permanent nuclear shadow into the system. That is not a fringe concern. It is the core concern.35

That view is no longer limited to Washington think tanks. Parts of the Gulf now appear to believe that a wounded but still intact Iranian regime may be more dangerous than a more decisive military outcome. This morning, the Wall Street Journal reports that the UAE is willing to support military action to force Hormuz open. That tells you how far regional thinking has shifted. In this view, the risk is not only escalation. It is leaving the regime bloodied, angry, and still in control of Hormuz, missiles, drones, and nuclear leverage.3637

That does not make the hard-line view correct. But it does make it harder to dismiss as American hawkishness from a safe distance.

And then there is Wall Street’s view.

The WSJ’s James Mackintosh laid out why the stock market has not treated this war like the end of civilisation. His argument has three parts. First, markets have often absorbed wars better than people expect; geopolitics can be brutal without necessarily breaking earnings. Second, profit forecasts have held up better than feared, especially with energy stronger and parts of corporate America still fairly resilient. Third, the AI trade is still alive, which matters because large-cap US equities are now heavily shaped by a handful of companies whose stories are bigger than one regional war. In other words, the market is not saying the war is harmless. It is saying the war will probably be contained, shortened, or politically forced back into a tolerable range. That is a bet, not a fact. But it is the bet currently in the price.38

Put differently, Chapter 6 is not really about picking your favourite pundit.

It is about learning to separate frames.

• Mearsheimer says: geography and escalation favour Iran.

• Sachs says: policy failure and hegemonic overreach created this mess.

• Ross and the preventive camp say: the mess may still be preferable to a nuclear Iran.

• The WSJ market view says: all of this is ugly, but probably not permanent.

Each frame captures something real. None is complete on its own.
But none of them gets the final vote. Trump does.

🎚️🎚️🎚️🎚️ Producer’s Note

Twenty-one miles of water.

That is all it takes for oil, gas, helium, fertilizers, chips (not potato chips) and half the usual political theatre to become very practical, very quickly.

A week ago, I would not have guessed helium would end up in this story either. That was partly the point of writing it. These systems are more connected than they look from a safe distance, and usually more complex than the loudest people in the room make them sound.

Time is part of the system too. Sometimes one hour matters. In Hormuz, so do 5,000 years. What looks stable in one moment can turn very quickly in the next.

This week’s track: beabadoobee & The Marías — All I Did Was Dream Of You.
Soft song. Car explosion. Lithuanian winter. Fair combination.

Happy Easter.

Share

Fab 🏂
LinkedIn | Insta | Twitter | Studio⍺

Sources

31

The Chris Hedges Report

Why America is Losing the War With Iran (w/ John Mearsheimer) | The Chris Hedges Report

Read more

3 months ago · 702 likes · 41 comments · Chris Hedges

That alone is reason enough to stop and look more closely.

And the timing is unusually good. Over the past few weeks, McKinsey, Andreessen Horowitz, Deloitte, Pew and others have published genuinely useful new data. So instead of repeating AI opinions, we can finally compare what is actually happening — across consumer adoption, enterprise rollout, agents, and markets.

Subscribe now

1) Users moved first

A few days ago, I listened to a very good a16z conversation around their latest “Top 100 Gen AI Apps” report. What makes that report interesting is not just the rankings. It is what the rankings reveal about behaviour. Users are already moving. Companies are still deciding how worried they should be.3

We have seen this movie before.

Back in the Beecom days, we sold a surprising number of Jira licences not through some grand CIO-led digital strategy, but through engineers and team leads who simply wanted to run better teams and paid for the licence out of their coffee box. They bought the tool because it solved a problem. If they had waited for central approval, half of them would still be waiting. The institution came later. First came the behaviour.

That is why the current AI moment is so interesting. We are seeing the same structural pattern again, only faster and with higher stakes. Employees are already using AI to do better work. Some do it with approved tools. Some do it with whatever works. A few do it with tools they absolutely should not be installing on a business machine full of sensitive data (aka Open Claw 🐞). The user moves first. Governance limps behind.

That is also why consumer AI matters even for B2B readers. Consumer usage is not a distraction from enterprise. It is often the leading indicator. It shows what people will do the moment the technology becomes good enough, cheap enough, and easy enough. Then enterprise has to react.

That is what makes the current AI moment so important. The real question is no longer only whether the models are capable. For many practical tasks, they already are. The bigger gap is between bottom-up adoption and top-down organisational readiness.

According to a16z’s March 2026 report, ChatGPT remains far and away the largest consumer AI product. On the web it is 2.7 times larger than Gemini; on mobile it is 2.5 times larger. ChatGPT has added 500 million weekly active users in the last year and now reaches 900 million weekly users, which a16z describes as more than 10% of the global population using it every week. That is not a niche experiment. That is a mass behaviour.4

In a16z’s Top 100 Gen AI Apps discussion, Olivia Moore noted that ChatGPT is “almost 30 times bigger than Claude on web and almost 80 times bigger than Claude on mobile,” underlining how far mainstream consumer usage still exceeds the noise around Claude in tech discourse.5

That said, consumer usage is only half the picture. OpenAI owns consumer. Anthropic is much closer in enterprise.

If you look at mainstream distribution, OpenAI is still in a different league. But if you look at business revenue, especially in coding and enterprise, Anthropic is much closer than these consumer charts suggest.

Reuters reported that OpenAI topped $25 billion in annualized revenue at the end of February 2026, up from $21.4 billion at the end of 2025. Anthropic, meanwhile, was reported at a $14 billion revenue run rate in February. That still leaves OpenAI ahead overall, but far less dramatically than the consumer-usage charts would imply.6

What makes Anthropic especially interesting is where that revenue comes from. The Financial Times reported that roughly 80% of Anthropic’s revenue comes from enterprise customers, that it has more than 500 customers spending over $1 million annually, and that Claude Code alone has moved above a $2.5 billion run rate. In other words: OpenAI still dominates consumer mindshare, but Anthropic has built a far more serious enterprise and coding business than the public conversation sometimes reflects.7

OpenAI still leads clearly on annualized revenue run-rate. But this chart shows why Anthropic should not be dismissed as a niche player just because Claude looks small in mainstream consumer usage. OpenAI remains the dominant consumer platform. Anthropic, however, has built a much stronger enterprise and coding business than the public conversation often reflects. That is the cleaner reading of the market right now: OpenAI owns mass distribution, while Anthropic has closed far more ground on the business side than the consumer charts alone would suggest.8

That also helps explain the latest strategic shift at OpenAI. The Wall Street Journal reported that OpenAI is cutting back side projects and refocusing on coding and enterprise productivity, with Anthropic’s traction serving internally as a wake-up call. So the cleaner reading of the market is this: OpenAI still owns mainstream consumer distribution, but Anthropic has closed much more ground on the business side than the consumer charts alone would suggest.9

So the summary is:

• ChatGPT is 2.7x bigger than Gemini on web

• ChatGPT is 2.5x bigger than Gemini on mobile

• ChatGPT is almost 30x bigger than Claude on web

• ChatGPT is almost 80x bigger than Claude on mobile

• OpenAI is only 1.78x bigger than Anthropic on annualized revenue run-rate.

2) AI is already big, but uneven

What matters even more than the raw scale is the shape of the market. AI is no longer just a chatbot sitting in a tab. It is splintering into different forms: chat assistants, creative tools, desktop apps, browsers, embedded copilots, and increasingly agents. The a16z report also makes a useful point that the old distinction between “AI-native” products and “normal software” is getting weaker. Some mainstream apps are now so AI-enabled that the cleaner question is not whether they are AI companies, but whether AI has become central to the product experience.10

That is a better way to think about where we are. The “is AI real?” debate is over. The real question now is where it is spreading first, how fast, and what is still holding it back.

The answer, in my view, is this: AI capability is no longer the main bottleneck for a large set of mainstream use cases. The bottleneck is organisational adoption.

That claim needs care. It would be sloppy to say the models are perfect or that capability no longer matters. In some domains it still very much does. But for a lot of everyday knowledge work, the bigger gap is now elsewhere: workflow design, permissions, trust, identity, security, and whether the company has done the boring but necessary work of integrating AI into how people actually operate. McKinsey put this unusually clearly just a few days ago: many organisations are struggling to scale generative AI and agentic AI, but the adoption challenges are “not technical — they are experiential.” That is a very sharp formulation, and I think it is right.11

If you want the enterprise version in numbers, Deloitte provides a useful reality check. Its 2026 report says workforce access to sanctioned AI tools has risen from fewer than 40% to around 60% in one year. That is real progress. At the same time, only 25% of organisations say 40% or more of their AI pilots have moved into production.

Put plainly: plenty of companies are experimenting; far fewer are truly scaling. Deloitte also says 85% of companies expect to customise autonomous agents to fit their business, yet governance maturity around these systems remains much lower. Enterprise ambition is running ahead of enterprise execution.12

3) Users always move before institutions

This is where the consumer side becomes useful again. Users adopt what helps them.

Pew’s latest March 2026 summary says the share of U.S. workers who report that at least some of their work is done with AI rose from 16% in 2024 to 21% in September 2025. That is still early. A majority, 65%, still say they do not use AI much or at all in their job. But the direction is obvious. Workplace use is growing even before most organisations have fully redesigned the underlying processes around it.13

If you want the sharper and more uncomfortable version, Microsoft’s latest Cyber Pulse report says 29% of employees use unsanctioned AI agents for work tasks. That number matters because it captures the real dynamic inside companies. The relevant question is not whether the board signed off on an AI strategy deck. The relevant question is whether people are already pulling AI into their workflows because it makes them faster, more capable, or simply less miserable. In many cases, they clearly are.14

In several recent conversations, people told me how useful OpenClaw had become in daily work. They were not talking like security engineers or developers, but simply like users who had found leverage. That part I understand very well. The problem was that the tool had been installed in work environments with access to sensitive data. In that setting, it is not just a productivity shortcut. It is a governance issue.

This is exactly where the current AI moment gets interesting. Workers are not using these tools because they want to break rules. They are using them because they want to do a better job. Often they do not fully understand the security model, the data exposure, or the downstream risk. They simply experience that the tool helps.

Microsoft’s own guidance on OpenClaw makes the risk plain. In February, Microsoft warned that self-hosted agents like OpenClaw should be treated as untrusted code execution with durable credentials and untrusted inputs. The company recommends isolation, dedicated non-privileged credentials, and non-sensitive data only. In other words: this is not something you casually install on a normal enterprise workstation and hope for the best.15

4) Why enterprise is still behind

So where does this leave the “state of AI” question?

It leaves us with a widening gap between bottom-up adoption and top-down readiness.

On one side, consumers and employees are already living in the next interface. They use AI for research, writing, coding, image generation, summarisation, search, transcription, and increasingly for actions rather than just answers. a16z’s report shows this behaviour broadening not only across products but also across geographies. Its per-capita adoption map puts:

1. Singapore at the top

2. followed by Hong Kong

3. the UAE

4. and South Korea,

While the United States ranks 20th. That is a reminder that AI adoption is not simply a function of who built the biggest models. It is also shaped by culture, trust, work structure, and how open a country is to new technology.16

On the other side, organisations are still sorting out who is allowed to use what, where data can flow, which systems can be connected, what kind of memory is acceptable, and how much autonomy a tool should have before someone signs their name under the risk. That is not irrational caution. In regulated sectors, it is often necessary caution. But it does create a structural lag.

And that lag is now becoming the central story of enterprise AI.

What Good AI Implementation Actually Requires: A useful recent McKinsey piece makes the enterprise bottleneck more concrete. Their argument is that adoption often stalls not because the models are too weak, but because the experience around them is still badly designed. They group the problem into four design principles: clarity, continuity, depth, and human–AI collaboration. That is a helpful framework, because it translates “AI adoption is hard” into more practical terms.

First, lead with clarity.
People trust AI more when they can see how it got to an answer, what assumptions it made, and where uncertainty still sits. If the system behaves like a black box, users either overtrust it or ignore it. Neither is useful. Especially in business settings, AI has to show its work well enough that people can challenge it, refine it, and act on it with confidence.

Second, design for continuity.
Most work is not a one-shot prompt. It happens across days, people, meetings, and systems. Yet many AI tools still behave as if every interaction starts from zero. That is a poor fit for real organisations. Context has to carry forward. If the system cannot remember what happened before, connect steps across a workflow, or preserve the thread of a project, it remains a clever toy rather than a serious work tool.

Third, build for depth.
The real opportunity is not a slightly better answer in a chat box. It is AI that supports multistep, domain-specific workflows from end to end. That means gathering inputs, applying logic, testing alternatives, refining outputs, and helping move work forward across several steps. In other words, not just responding faster, but participating in the actual shape of the work.

Fourth, design for human–AI collaboration.
The future is not humans fixing machine output after the fact like tired editors cleaning up after an intern. The better model is collaboration: AI contributes structure, speed, and range; humans bring judgment, context, and accountability. The system should make that partnership easier, not more awkward. When that works well, people are not just “using AI.” They are working with it.

This is a more useful way to think about enterprise lag than simply saying companies are slow or employees resist change. In many cases, the tools fail because the surrounding experience is still fragmented, stateless, shallow, or badly integrated into how people actually work.17

5) The real platform war is moving up the stack

While companies are still catching up operationally, the market is already shifting strategically. The AI race is moving up the stack — away from raw model quality alone and toward context, memory, identity, and distribution.

For the last two years, public discussion has been dominated by model performance. Which model is best? Which benchmark moved? Which lab is ahead this week? That still matters, but it matters less than it did. The more interesting battle now is over context, memory, identity, and distribution.

a16z hints at this when it talks about compounding advantages from context and app-store style dynamics. The idea is simple: once an AI product has your history, preferences, files, identity, and workflows, switching away becomes harder. It is no longer just a better chatbot. It starts becoming a layer you operate through.18

Google’s January launch of Personal Intelligence is a concrete signal of this direction. Gemini can now, if users opt in, connect to Gmail, Photos, YouTube, and Search to produce more personalised answers based on their own data and behaviour. That is a very different product logic from a stateless assistant waiting for the next prompt. It is the beginning of something much stickier.19

OpenAI’s move into testing ads inside ChatGPT points in the same broad direction, though from a different angle. Ads themselves are not the interesting part. The interesting part is that ChatGPT is increasingly behaving like a scaled consumer platform with persistent engagement, multiple monetisation layers, and a growing claim on user attention and identity.

OpenAI says the ad test began in the U.S. on February 9 for logged-in adult users on Free and Go tiers, while paid business and enterprise plans remain ad-free. That is not just a pricing tweak. It is a sign that the consumer AI interface is becoming durable enough to monetise like a platform.20

That is why I suspect the next winners in AI will not be decided only by who has the smartest raw model in isolation. They will be decided by who becomes the default layer for user context, memory, identity, and action. This is still an inference, not a settled fact. But it is where the evidence points.

The same logic applies to agents.

6) Agents are arriving faster than governance

Agents are real now, but it is still easy to talk nonsense about them. Every second product is suddenly “agentic,” which usually means it can do two things instead of one and now comes with a slightly more expensive pricing page. So it helps to strip the drama out of it. The important shift is not the label. The important shift is that more AI products are moving from generating outputs to executing tasks across tools and systems.

That is a meaningful change.

Deloitte says close to three-quarters of companies plan to deploy agentic AI within two years. Yet only 21% report mature governance for those systems. That is a proper mismatch. Companies want the upside of action, but most are still building the controls.21

Anthropic’s March 2026 labour-market research adds one useful nuance here. The company argues that AI remains far below its theoretical capability in actual work usage. Real-world adoption is still only a fraction of what the technology could plausibly support. That matters because it keeps us honest. Yes, the models are improving quickly. No, deployment is not automatic. There is still a long road between capability and widespread productive use.22

So the state of AI right now is not best described as “AGI is near” or “enterprises are adopting fast” or “agents are taking over.” All of those are too loose.

A better description is this:

Consumers and employees have already crossed the line. For many use cases, they treat AI as a practical layer for work and life. Organisations are still trying to catch up with policy, workflow redesign, trust, identity, and security. The next two years will be shaped by that gap.

That should sound familiar, because structurally it is not new at all. First the user finds leverage. Then the team standardises it. Then management notices. Then legal and security arrive with a long face and a bigger checklist. Then, eventually, the institution absorbs the behaviour and calls it strategy.

That is how a lot of software adoption really works. AI is not breaking that pattern. It is accelerating it.

The difference this time is that the leverage is larger, the consumer scale is already enormous, the tools are becoming more stateful, and the downside of getting the rollout wrong is far more serious than buying the wrong ticketing system.

That is why the “state of AI” in March 2026 looks less like a clean enterprise rollout and more like a race between bottom-up usage and top-down readiness.

The models are ready for more than most companies are currently allowing. The users know it. The boards are beginning to realise it. The winners will be the organisations that close that gap fastest — without being reckless about governance, identity, memory, and security.23

7) Why Investors Still Don’t Know What to Make of AI

If companies are still struggling to understand how AI will reshape their workflows, it should not surprise anyone that investors are struggling too.

A sharp Economist piece this week makes exactly that point: markets are often bad at pricing technological revolutions in real time. When change is slow and linear, investors can extrapolate. When change is structural, messy, and uneven, they start guessing. AI is creating exactly that kind of confusion.

You can see it in today’s market already. Goldman Sachs has built one index of firms seen as most exposed to AI disruption and another of firms expected to benefit most from AI-driven productivity. Over the past year, the “at risk” basket fell sharply. But the “beneficiaries” basket also fell. In other words, the market is not simply rewarding the winners and punishing the losers. In many cases, it still does not know which is which.

History suggests that this is normal.

When Edison’s electric-light revolution began in the late 19th century, it looked like the future had arrived and gas companies were finished. Serious money piled in behind the new technology, including J.P. Morgan. But the story turned out to be more complicated. Electric light remained expensive for years, and gas companies adapted by expanding into other uses. In London, gas firms found a large new market in cooking. What looked at first like a clean technological replacement turned into a more uneven and surprising transition.

That pattern matters now. It is tempting to talk about AI as if the market should already know who is doomed and who will dominate. But technological shifts rarely work that neatly. Some incumbents will be hit harder than investors expect. Others will adapt better than expected. Some of today’s supposed winners may turn out to be overhyped. Some of today’s laggards may find a different position in the value chain.

Business history is full of these strange second acts. Western Union famously missed the telephone, then built a durable business in money transfers. American Express began as a freight-forwarder. Samsung once sold dried fish. Reinvention is not unusual. It is one of the main reasons technological revolutions are so hard to price while they are happening.

What makes AI different is not that reinvention becomes impossible, but that the tempo may be far less forgiving. The World Economic Forum expects job disruption to affect 22% of jobs by 2030, with 170 million new roles created and 92 million displaced, while 39% of workers’ core skills are expected to change by then. That does not mean the economy runs out of work. It does mean the rewiring of work may happen faster than many large organisations, boards, and labour systems are used to handling.24

That is why AI is both everything and nothing in the market at the same time. For some firms, it is already a real strategic threat. For others, it is still more noise than economics. And for many, the outcome will depend less on abstract AI capability than on something more ordinary: whether they actually deploy it well, improve their product, and redesign how work gets done.

That brings us back to the central point of this article. The state of AI in 2026 is not defined only by what the models can do. It is defined by who can turn capability into operating reality. Users have already moved. Many organisations are still catching up. Investors, for now, are mostly guessing which companies will close that gap — and which will not.

Outro

If I had to reduce this whole piece to one simple observation, it would be this: users have already moved, and most institutions are still negotiating with reality.

That does not mean every AI-native startup wins. It does not mean every incumbent loses. Business history is more uneven than that. But it does mean the pace of change is picking up, and the comfort of moving slowly is becoming a real strategic risk.

I do not think the next decade will eliminate work. I do think it will rewrite a great deal of it — faster than many boards, managers, and public institutions are prepared for.

That is why this moment matters. Not because every claim made by the AI industry is true. Quite a few are not. But because the underlying shift is real, the adoption patterns are visible, and the cost of waiting for perfect clarity is going up.

🎚️🎚️🎚️🎚️ Producer’s Note

Sorley got there faster than most board committees: stop acting.

Decent soundtrack: The challenge now is not whether the models can do the work, but whether organisations can stop admiring the problem and start implementing.

Let your board know.

Share

Fab 🛑
LinkedIn | Insta | Twitter | Studio⍺

Sources

But the story refused to sit still. Every day there was a new reaction, a new take. If you publish too early, you’re basically writing a diary of your own confusion.

So we waited until the dust settled a bit. And then: Citadel published a rebuttal, and The Economist weighed in. When institutions bother to respond, it usually means something real is being debated.

On Feb 22, 2026, a small research shop published a long, fictional “memo from the future” called “The 2028 Global Intelligence Crisis.” It was explicitly framed as a scenario, not a prediction. (Citrini Research)1

Within days, it was being treated like a market catalyst — partly because it landed in the middle of an already nervous debate about AI “winners vs losers.” The result: a lot of headlines, a lot of hot takes, and (finally) a few serious rebuttals.

Ironically, the report describes a crash in 2028 — and still managed to trigger a real “AI scare trade” in 2026. Software and other perceived AI-losers sold off hard, while the market tried to reprice who actually wins when agents get real.

This is an after-action review of the last ~two weeks — and it matters even if you don’t own a single stock.

Because the real question isn’t “will markets crash in 2028?” It’s whether AI agents start changing how companies hire, how consumers buy, and how fast routine office work gets compressed.

We’ll break down:

• what Citrini actually wrote,

• what serious critics agreed with — and where they pushed back,

• who Citrini and Alap Shah are (and what incentives they have), and

• what this means for founders, investors, employees — and the next generation choosing careers.

Subscribe now

A quick timeline of the ‘AI scare trade’ (Feb 22 → now)

Feb 22 — Citrini publishes The 2028 Global Intelligence Crisis (scenario format, dated June 2028).2
Feb 23–24 — the memo spreads fast; markets wobble; the “AI winners vs losers” debate turns into a trade.3
Feb 24 — Citadel publishes a rebuttal (counter-memo with data).4
Feb 25 — The Economist weighs in on why the economics don’t add up.5
Feb 26 — B Capital frames it as “not 2028, but a real signal,” claims viral amplification, etc.6
Late Feb / early Mar — Commentary proliferates: “this is valuable as a stress test,” “this is attention-economy research,” “this is talking-your-book.” (More on that below.)

If you want a mental model: week 1 was narrative shock, week 2 was counter-analysis.

To understand why this moved markets, you need to see the report the way traders saw it: not as a prediction, but as a clean mechanism with a timeline attached.

1) What the Citrini report actually says (in one clean storyline)

The setup: a “macro memo” dated June 30, 2028

Citrini’s piece7 is written as a fictional post-mortem from June 30, 2028. The opening is deliberately punchy: unemployment prints 10.2% and the S&P is down 38% from its October 2026 highs. The narrator says the market is numb — a print like that would have triggered a circuit breaker two years earlier.

Then they rewind the tape. In their story, late 2026 is the “everything looks great” phase: S&P ~8000, Nasdaq > 30k, layoffs begin, margins expand, earnings beat, stocks rally — and companies funnel profits into AI compute. Nominal GDP still prints fine. Productivity looks like it’s booming because “agents don’t sleep.”

This framing matters: the report isn’t saying “AI fails.” It’s saying AI succeeds — and that success breaks the labour-market equilibrium.

Step 1: Agents aren’t chatbots — they’re labour

The big assumption is agency. Not “AI writes text,” but “AI does work without being babysat.” Citrini leans on the idea that model capability is moving from minutes → hours → days → multi-week autonomy, which would make agents a credible substitute for a lot of knowledge work.

Why they push this: if an agent can complete multi-day or multi-week tasks reliably, you stop thinking of it as “software.” You start thinking of it as a worker you can buy by the seat.

Step 2: “Friction” goes to zero, and pricing power follows

Citrini’s central micro-claim is about friction: many businesses make money because humans find certain tasks tedious (shopping around, switching providers, reading fine print, re-negotiating contracts, re-balancing accounts). A lot of that friction isn’t “necessary” — it’s the by-product of clunky processes and mediocre UX (sometimes accidental, sometimes convenient).

When agents do those tasks continuously, friction collapses. In their story, that’s when business models built on “intermediation” start to wobble — not because consumers stop consuming, but because fees and margins get arbitraged away by machines.

This “agents don’t care about bad UX” angle isn’t new. In a private session a few years ago, Olaf Carlson-Wee from Polychain8 made a similar point: machines don’t get frustrated by menus, forms, onboarding, or switching. They just execute. Citrini packaged it well — but the underlying insight has been in the air for a while.

They give specific examples in the report narrative (intermediation across consumer transactions; payments; and marketplaces).

Step 3: The corporate response is rational — and that’s the problem

Citrini explicitly says the individual company response is rational: boards demand action, stocks are under pressure, and the easiest lever is to cut headcount while using AI tools to maintain output.

Their point is not “companies are evil.” It’s that when every company does the rational thing at once, you get a collective outcome nobody wants.

They describe a world where “software was only the opening act” — the reflexive loop escapes SaaS and spreads to any sector with a large white-collar cost base.

Step 4: “👻 Ghost GDP” + the “no natural brake” loop

This is the report’s most memetic section. They claim headline output can still look fine while household demand weakens, and they label that “👻 Ghost GDP” — output in the national accounts that “doesn’t circulate through the real economy.”

Then comes the mechanism they repeat:

• AI capability improves →

• companies need fewer workers →

• white-collar layoffs increase →

• displaced workers spend less →

• firms face margin pressure and invest more in AI to protect margins →

• AI capability improves again.

And they call this a feedback loop with “no natural brake” — unlike a normal recession where the cycle eventually self-corrects.

Step 5: Displaced white-collar labour floods the “real economy”

Citrini’s next step is the labour reallocation story. People don’t just disappear. They downshift. High-skill / high-wage workers move into lower-barrier service and gig jobs, increasing labour supply there and compressing wages.

They include a specific anecdote (Salesforce PM → Uber driver) to make the point intuitive, then generalise: once overqualified labour hits the service economy, the wage pressure spreads beyond tech.

Step 6: Contagion channels — credit, mortgages, and policy

After “real economy” effects, they move into finance.

6.1 Private credit + life insurers

They argue private credit is full of deals priced off assumptions like “ARR stays recurring,” and that a disruption-driven default cycle hits a system that hasn’t had a real stress event in years.

6.2 Mortgages: 2008 vs 2028 (their framing)

They then claim the mortgage problem isn’t “bad loans,” it’s “good loans written against an income future that later becomes impaired.” They highlight stress showing up first in high-income tech/finance metros.

6.3 The “three drivers” model

At the end, Citrini wraps the story into three interacting loops: real economy → financial system → policy response, with a timeline bar that shows the crisis evolving from disruption → contagion → fiscal/policy stress.

The punchline of Chapter 1

Even if you disagree with the macro spiral, you can see why the piece travelled:

• It frames AI as a cost-collapse event (agents as labour).

• It points at a plausible micro-mechanism: agents reduce switching costs, which reprices intermediation and weakens pricing power.

• It explains why this might not behave like a normal cycle: the “engine” improves every quarter.

That’s the report in its own words and diagrams. In chapter 3 is where we ask: what did serious critics accept, and what did they reject — and why?

2) Who is Citrini (and why many people hadn’t heard of them)

Citrini is not sell-side. It’s not “Goldman Research.”

One reason this surprised people: Citrini isn’t a big bank desk. It’s closer to an independent research brand distributed through Substack and social amplification.

Bloomberg describes Citrini’s founder as a 33-year-old who published the scenario on the firm’s Substack and was surprised by the market reaction. (Bloomberg.com)

WSJ characterised Citrini as a founder-led research effort publishing macro/thematic stock research under the Citrini name. (The Wall Street Journal)

James van Geelen (Citrini founder): background

Fortune reports van Geelen is a former Los Angeles paramedic, with degrees in biology and psychology, and built a reputation around “second-order thinking.”9

MarketWatch/Morningstar coverage paints him as a thematic stock-picker figure who got attention for big themes (AI, GLP-1s) and runs Citrini Research as a newsletter-style product. 10

Why that matters:
He’s not presented as a macro economist or an AI lab founder. His “edge” (if you buy it) is packaging second-order implications in an investable narrative.

Alap Shah (co-author): background

Shah is described by Bloomberg as the chief investment officer at Lotus Technology Management, and the person calling for policy responses like an AI windfall tax. 11

He’s also CEO of an AI company called Littlebird (per Littlebird’s own leadership page). 12

And he co-founded Sentieo, which was acquired by AlphaSense in 2022 (acquisition announcement).13

Why that matters:
Shah has direct experience in both investment framing and AI-driven workflow tooling. He’s still not a frontier model lab founder — but he isn’t random either.

“Talking their book”

This is where scepticism is healthy.

MarketWatch reports Shah disclosed his firm had short positions in companies they believe will be disrupted by AI, while also holding longs (e.g., semiconductors) that benefit from AI.14

So yes: the “book in the background” concern is legitimate — not necessarily malicious, but relevant context for weighting the narrative.

3) What the press and serious analysts said (good + bad)

Most coverage of Citrini was commentary: “interesting scenario,” “spooked markets,” “here’s the vibe.” A few outlets actually tried to pressure-test the economics and the diffusion story. And one player (Citadel Securities) basically published a counter-memo with data and charts.

This chapter has two parts:

1. a quick map of where the debate landed, and

2. a deeper look at the Citadel rebuttal (because that’s the closest thing to a competing model).

3.1 The quick map: who agreed with what (and why)

Where people mostly agree with Citrini

1. Markets are pricing dispersion now — not “AI = good for everyone.”
   A lot of the coverage treats the report as a catalyst for a shift already underway: AI is powerful, but it won’t benefit every layer of the stack equally.15

2. “Friction businesses” are exposed if agents reduce switching costs.
   Even sceptics tend to concede the micro-mechanism: if agents make it painless to compare, switch, and negotiate, then some intermediation margins and some weak-moat software pricing power get pressured.16

Where critics disagree sharply

1. The economics: “👻 Ghost GDP” as a clean macro spiral is the weak link.
   The Economist’s critique is basically: the scenario reads well, but the economics are shaky. The core issue: output, income, and demand are linked by accounting identities and behavioural responses; you can’t just assert “GDP stays up, demand collapses” without a lot of extra assumptions.17

2. Diffusion speed: capability can scale faster than adoption.
   This is where Citadel goes hardest: AI can improve quickly, but real deployment across firms and sectors is slowed by integration costs, regulation, physical constraints (compute/energy/data centres), and organisational change.18

The middle-ground take
“Not 2028 as base case, but the business model pressure is real.”
B Capital’s post basically says the timeline is unlikely, but the investment implications are already investable: pricing pressure, changing moats, and the risk that markets overreact to narrative objects when positioning is crowded.19

Another example in the same lane: iCapital argues the “creative destruction” framing is exaggerated relative to current evidence of job displacement.20

3.2 Deep dive: Citadel’s rebuttal (the “2026 Global Intelligence Crisis”)

Citadel’s response matters because it’s not a two-paragraph dismissal. It’s a structured counter-argument with data, a diffusion model, and a macro accounting critique.

Citadel’s framing is simple: the macro community struggles to forecast near-term payrolls reliably, so treating a Substack scenario as a high-certainty forward path is… ambitious. Their core claim is that the “imminent displacement” story depends almost entirely on the speed of diffusion — and the data doesn’t show a non-linear adoption jump yet.21

1) Start with the data: the labour market doesn’t look like a cliff

They open with a cluster of “reality check” datapoints: unemployment at 4.28%, AI capex framed as 2% of GDP, AI-adjacent commodities up, thousands of US data centres planned, and software engineer postings up 11% YoY.

Their implication: if the story is “AI is already replacing white-collar labour at scale,” you’d expect the forward-looking labour indicators to be rolling over harder — especially in tech. Their chart argues the opposite.

2) The key question isn’t “do you use AI?” — it’s “how intensely do you use it at work?”

Citadel then goes after the adoption narrative using survey data from the St Louis Fed’s Real Time Population Survey. They say most first-order “AI adoption” charts are binary (use or not), but displacement risk depends on intensity: daily use for work should be rising sharply if we’re at the start of a rapid substitution wave. They claim the daily-use lines look “unexpectedly stable,” which to them is evidence against imminent labour displacement.

3) Their main conceptual point: “recursive technology” ≠ “recursive adoption”

This is the cleanest part of the rebuttal. Citadel argues people are conflating:

• the recursive potential of the technology (models improve, tools improve), with

• recursive economic deployment (firms adopting and integrating it at exponential speed).

They point out diffusion typically follows an S-curve: early adoption is slow and expensive; it accelerates as infrastructure and complements arrive; then it saturates as marginal adopters get less value. They argue markets tend to extrapolate the acceleration phase linearly, but real deployment hits friction: integration cost, regulation, and diminishing returns.

Then they add a “hard constraint” layer: even if models improve, large-scale substitution requires semiconductor capacity, data centres, and energy. If compute demand spikes, compute gets more expensive; at some point the marginal cost of compute can exceed the marginal cost of human labour for certain tasks — which becomes a natural boundary on substitution.

4) “Productivity shocks are supply shocks” (and why they think Citrini breaks accounting identities)

Citadel argues AI automation is fundamentally a productivity shock. Historically, productivity shocks tend to be positive supply shocks: they lower marginal costs, expand potential output, and raise real income over time. Their key objection to Citrini is the claim that productivity rises while aggregate demand collapses, yet output keeps rising — which they say violates basic national accounting unless you assume away offsets (investment, transfers, net exports, etc.).

This is basically the “👻 Ghost GDP” critique in accounting language: if real output rises, something on the demand side has to rise too — consumption, investment, government spending, or net exports — unless you add a lot of extra assumptions about persistent redistribution failure and investment absorption failure.

5) New business formation as a counterforce

Citadel then points to business formation data as a reason the “AI only destroys” framing is incomplete. If the cost of doing cognitive work falls, it can also lower the barrier to starting companies, creating new categories of demand and new task bundles.

6) Substitution elasticity + “the economy is full of tasks that are costly to automate”

Citadel’s macro variable is substitution elasticity: how easily can firms substitute AI capital for labour at stable cost? They argue that even with powerful AI, the economy has a vast set of tasks that are physical, relational, regulatory, supervisory — plus trust/liability barriers in cognitive work — which makes full substitution unlikely. Historically, tech shifts task composition more than it eliminates labour entirely.

They also argue that if displacement became severe, democratic governments would likely respond with fiscal/regulatory measures that blunt the worst-case outcome (which lowers effective substitution elasticity).

7) “AI can also create jobs” — via the build-out itself

Citadel adds a practical, near-term counterpoint: regardless of what happens to office work, the AI build-out is driving construction and related hiring. This doesn’t refute long-run displacement risk, but it supports their “diffusion is slower and messier” argument: the first-order impact may be capex-heavy build-out, not instant labour collapse.

8) Their closing metaphor: Keynes and the “15-hour work week”

Citadel ends with a historical analogy: Keynes predicted productivity growth would collapse the work week. He was directionally right on productivity, wrong on labour withdrawal. People didn’t stop working; they consumed more, demanded higher quality, and invented new categories of spending. Their punchline: productivity gains don’t automatically cause demand collapse; they expand the consumption frontier.

Why this rebuttal mattered in the media cycle

Citrini triggered a debate. Citadel tried to pin it to measurable realities: adoption intensity, diffusion speed, compute constraints, and accounting identities. You can disagree with Citadel’s conclusion without dismissing the value of the exercise — but it’s the most serious “counter-model” published in response.

Thank you, Citadel.

4) The part that matters beyond markets: white-collar workers and our kids

This is where “what it means for founders/investors” is too narrow.

The phrase “white collar” sounds precise, but it isn’t. Depending on the country, it can be interpreted as “almost everyone in services” — which is basically the whole economy. That framing creates drama, but it doesn’t create clarity.

So before we talk about “white-collar collapse,” we need a better mental model. The right parallel from history is not “a sector died.” It’s: productivity shifted so much that the share of people needed in one area collapsed over time. It’s structural change.

Box: From farms to desks — big labour shifts are normal

The last time we had a “this changes everything” productivity shock, it was agriculture.

• 1800 (US): about 83% of the labour force was engaged in agricultural production.

• 1820 (US): about 71.8% of “gainful workers” were still in agricultural pursuits.

• Pre-industrial economies (general): it’s normal to see 80–90% of the workforce in farming in very low-productivity settings.

• 2023 (Switzerland): services account for ~77% of jobs, and employment in agriculture is below 50,000.

• 2023 (worldwide): agriculture (incl. forestry & fishing) still employs 916 million people, 26.1% of total employment (Europe ~5%).

• What this analogy gets right:
  When productivity in a sector explodes, the share of people working there can collapse over time. That’s not a moral judgement. It’s structural change.

What this analogy does not prove:
“White collar” is not a single sector like farming.

It’s a bundle of roles spread across the services economy, including:

• healthcare and care delivery (not just administration)

• education and training

• government and regulation

• finance, risk, compliance

• sales, customer success, procurement

• operations, logistics, real-world coordination

• software, media, and “pure information work”

So the right question isn’t “will white collar fall from 80% to 4%?”
The better question is: which white-collar tasks become cheap and automated (routine information-handling), and which roles expand (human-facing work, accountability for outcomes, regulated/physical-world complexity)?

The report’s central fear is white-collar displacement happening faster than new roles appear. Even if you think their 2028 meltdown is overstated, the near-term anxiety is real: many people feel the ladder into knowledge work is wobbling.

There’s mainstream reporting that high-signal parts of the white-collar market are already uncomfortable — for example, MBAs at top schools struggling to land roles.22

And WSJ has separately discussed tougher entry conditions for graduates.23

Layer on what companies are actually doing: large firms continue to cut support roles while talking up efficiency and AI-enabled productivity (for example, financial sector layoffs reported by AP).24

So what should a smart 18–25 year-old do with this?

Here’s the non-dramatic answer: don’t pick a degree based on whether “AI will take jobs.” Pick based on whether you can become the kind of person who uses AI to do meaningful work.

A practical “study strategy” that survives most AI futures:

• Build one deep domain
  Healthcare, law, finance, industrial systems, security, energy, supply chain. Domains with rules, constraints, and real-world consequences tend to resist pure automation.

• Get fluent in the tools (not just “AI awareness”)
  Not “prompting”. Actual workflow: data, spreadsheets, APIs, automation, basic coding, model-assisted research.

• Add one edge that’s hard to commoditise
  Communication, negotiation, leadership, taste, product judgement, relationship-building. These are boring to list and annoying to practise, which is exactly why they remain valuable.

  

  ---

• Aim for roles that “own outcomes,” not roles that “move information”
  If your job is mainly moving information between systems and people, agents will pressure it. If your job owns a result (revenue, risk, product delivery, compliance outcomes), tools become leverage rather than replacement.

That’s the part of the Citrini memo I’d keep: it’s a forcing function to audit which white-collar tasks are just “friction work.”

What to watch next

Citrini is best read as a stress test, not a forecast. The sharp signal is micro: agents reduce switching costs, which pressures business models built on friction and inertia. The weak link is the clean macro spiral — diffusion, constraints, and policy responses make the real path messier than the memo suggests.

So the useful output isn’t a prediction. It’s a checklist: where does your job, your company, or your country rely on friction that software can make cheap?

If you want a simple dashboard for the next 6–12 months:

• White-collar hiring (especially tech-adjacent and “information worker” categories)

• Renewal pricing and seat compression in SaaS

• Evidence of real agent-driven switching (finance, insurance, travel)

• Whether “AI investment” shows up as measurable substitution or mostly as tooling layer adoption

And yes: keep watching the “press vs rebuttals” pattern. The most useful signal in this episode wasn’t the memo — it was that serious institutions bothered to respond to it.25

🎚️🎚️🎚️🎚️ Producer’s Note - “Rolling Stone”

Jim Legxacy shot “aggressive” on a Blackberry and it still lands; Citrini’s report was that kind of raw, viral take — and Citadel showed up as the mix engineer to separate signal from noise.

Thanks for sharing our blog.

Share

Fab 👻
LinkedIn | Insta | Twitter | Studio⍺

Sources

1

Citrini Research

THE 2028 GLOBAL INTELLIGENCE CRISIS

Preface…

Read more

4 months ago · 8119 likes · 91 comments · Citrini and Alap Shah

2

Citrini Research

THE 2028 GLOBAL INTELLIGENCE CRISIS

Preface…

Read more

4 months ago · 8119 likes · 91 comments · Citrini and Alap Shah

7

Citrini Research

THE 2028 GLOBAL INTELLIGENCE CRISIS

Preface…

Read more

4 months ago · 8120 likes · 91 comments · Citrini and Alap Shah

After Anthropic announced a security-focused AI capability, several public cybersecurity names sold off (CrowdStrike, Datadog, Zscaler, etc.) — not because they missed earnings, but because investors instantly priced in a new reality: the model providers are moving into the application layer.1

And then, one day later, came the second signal — even more revealing: Anthropic softened its own safety commitments, explicitly pointing to competitive pressure. In plain English: “we can’t afford to be the only ones slowing down.”2

Every founder has this thought at 2 a.m.:

“What if we build this thing… and then Google / Microsoft / Apple / OpenAI / Anthropic just ships it as a feature?”

It’s not paranoia. It’s a rational fear — and it has moved up-market. It’s no longer just early-stage teams asking this question. It’s later-stage category leaders, and even AI-native companies building on top of the same platforms that might later compete with them.

In the 90s it was the OS and the browser. In the 2000s it was search and marketplaces. In the 2010s it was mobile and app stores. Now it’s the AI assistant becoming a new “interface layer” — and when the interface layer shifts, entire software categories get repriced.

We’ve seen this movie before. The names change. The mechanics don’t.

So instead of hand-waving it away — or turning it into a moral debate — this post does the boring but useful thing: it breaks down how this actually plays out in the real world. When incumbents copy and kill. When copying backfires. When the startup pivots and wins. When the outcome is an acquisition instead of a war.

Same story every decade. New interface layer. Old powers. Fresh founders.

Subscribe now

Part I - The Map

1. Why this topic is back (and why AI changes the stakes)

Before we get into mechanisms and outcomes, it helps to remember: this game has roles.

The classic pattern is old: platforms expand. Microsoft did it. Google did it. Apple did it. And now the AI labs are doing it too.

What’s different this time is speed and surface area.

In the old cycles, “moving downstream” took years. New products had to be coded, shipped, distributed, adopted. Today, model providers can ship a new capability, wrap it in an assistant UI, and distribute it instantly to millions of users already sitting inside their product.

Model providers are no longer “just” infrastructure. They’re actively turning the model layer into an application layer:

• OpenAI launched the GPT Store — basically an app marketplace inside ChatGPT.3

• Google is building Gemini directly into Workspace apps (Docs, Gmail, Meet, etc.). That’s distribution at the point of work — where habits already live.4

• Anthropic pushed “Artifacts” and then “apps inside Claude” — turning the chat UI into a lightweight application runtime.5

This matters because it changes the competitive geometry.

The fear is not “Big Tech copies startups.” That’s old news.

The fear is this:

When ChatGPT / Claude / Gemini become the default front door for knowledge work, they can absorb huge parts of the software stack.

In previous eras, the “interface layer” was the thing users opened first:

• Windows and the browser in the 90s

• search and marketplaces in the 2000s

• mobile and app stores in the 2010s.

• Now the AI assistant is becoming that layer — the place where work starts, and increasingly where work happens.

And once the assistant6 becomes the front door, a lot of software stops being something you “open” and becomes something the assistant “calls” in the background.

• Instead of opening Jira to write a ticket, you tell the assistant: “Create a ticket, assign it to X, summarize the call notes, set due date Friday.”

• Instead of opening a legal drafting tool: “Draft an NDA based on our template, highlight unusual clauses, export to Word.”

• Instead of opening a security dashboard: “What’s the riskiest alert this week, what changed, what should we fix first?”

For users, that’s great. It’s less clicking, less searching, more outcomes.

For software companies, it changes everything.

Because if the assistant owns the front door, it can bundle the most common workflows as built-ins. Entire categories risk becoming “just a prompt” or “just a built-in tool.” Not because startups are bad — but because the platform owns distribution and can make a feature feel “free” inside an existing subscription.

That’s the modern version of bundling.

And it’s why LPs ask the question again — not as a philosophical concern, but as a very practical one:

If the platform keeps moving downstream, where can startups still build durable value — and where are they simply building features that will be integrated away?

If you want a modern AI-era example that captures this fear in one clean story: Jasper.

In 2022, Jasper raised $125M at a $1.5B valuation, riding the first wave of LLM-enabled marketing copy.7

Then OpenAI released ChatGPT to the public on November 30, 2022 — and overnight a huge chunk of “AI writing” went from paid product to free-ish default.

It wasn’t even malicious. It was the platform doing what platforms do: integrate the most common use case.

That Jasper moment is the perfect bridge into the first category we need to analyze: when incumbents copy (or bundle) and the startup collapses — not because the team is stupid, but because the product was shaped like a feature.

That’s Chapter A1.

2. The 5 mechanisms incumbents use

People say “they copied us.” That’s not precise enough.

There are five recurring mechanisms incumbents use to win downstream. And each one produces different outcomes.

Mechanism 1 - Bundling / tying (default distribution wins): If you control the OS, the browser, the app store, the enterprise suite — you can ship a feature to millions overnight.

This is the oldest weapon in the book.

Mechanism 2 - Self-preferencing (rank yourself first): If you control a marketplace, a feed, a search engine, a directory — you can put your product at the top and quietly starve others of oxygen.

Mechanism 3 - Platform rule changes (API, policy, permissions): If you depend on the platform, they can change the rules: access, rate limits, permissions, background activity, pricing, store policies.

No drama. Just a “developer update”.

Mechanism 4 - Cross-subsidy / price compression (“free” beats “best”): Incumbents can bundle the feature into an existing contract, price it at zero, and force the market into a margin collapse.

Mechanism 5 - Data advantage (they sit on the workflow): When the incumbent already owns the workflow and telemetry, it can build the feature faster and improve it faster.

This is especially relevant in AI, where feedback loops matter.

3. The outcomes

Here’s the key: the mechanism doesn’t automatically determine the outcome.

In practice, we see six outcomes:

• ⛔️ (A1) Copied / bundled → startup collapses

• ✅ (A2) Copied / bundled → startup pivots and survives (sometimes thrives)

• ✅ (B) Copied → the incumbent buys it (company or capability)

• ✅ (C) Copy attempt fails → startup wins anyway

• ✅ (D) Regulation shifts the game → category re-opens

• ✅ (E) Coexistence → “good enough” for the mass market, specialists thrive at the edge

Most founders only see A1. That’s why they panic.

But the real world is messy — and the other branches happen constantly.

Part II - The Evidence

4. (A1) Copied / bundled → startup collapses | When incumbents actually kill startups ⛔️

Let’s not sugarcoat it.

Sometimes a platform ships your product as a checkbox and you’re done.

This tends to happen when:

• the startup is feature-shaped (not a workflow owner),

• switching costs are low,

• distribution is controlled by the incumbent,

• and the incumbent can ship “good enough”.

Case: Microsoft vs Netscape (bundling + OS leverage)

The Internet Explorer era is the canonical story of tying and distribution power.

The U.S. v. Microsoft findings of fact describe the market dynamics around Windows’ dominance and Microsoft’s conduct around browsers and threats to the OS monopoly.8

You don’t need to argue ideology here. The lesson is mechanical:

If your competitor owns the default distribution layer, product quality alone won’t save you.

Case: “Sherlocking” — Apple shipping your app as a built-in feature

This is so common it has a verb.

Apple shipped Sherlock 3 with functionality that developers argued closely matched Karelia’s Watson, and “being Sherlocked” became a term in the Mac dev world.9

This is A1 in its purest form: a platform feature makes a third-party product economically irrelevant.

Case: GO Corp / PenPoint vs Microsoft (the “early pen computing” war)

GO Computer’s PenPoint is often cited as an early example of platform pressure in a category Microsoft wanted to control.

There’s real case material around GO’s antitrust claims related to Microsoft’s conduct and PenPoint technology.10

Be careful with wording (because history gets mythologized), but the pattern is again mechanical:

When the platform owner decides your category is strategic, they can drag the center of gravity into the platform and starve independent players.