This is going to hurt. Part 2: the big three
Three AI labs carry 2.07 trillion dollars of paper value on roughly 72 billion of revenue: 44 Estonias of valuation on one Slovenia of income. The rise broke the memory market. The fall is already visible: open weights three points behind and shrinking toward a gaming card.
Deze pagina is geschreven met hulp van AI.
Three AI companies are worth 2.07 trillion dollars this summer, on roughly 72 billion dollars of combined revenue. That is 44 Estonias of paper value on one Slovenia of income. Part one looked at what the AI wave does to your daily work; this part follows the money, because the question "how f*cked are we?" has a second reading: how f*cked is the industry selling us all this?
The story so far: part one showed the AI wave repricing the tasks inside your job, faster than your gut expects. Now follow the money.
The hurt comes in two acts: first the rise, then the fall.
The rise
The numbers first, because they are so big they stop feeling like numbers. Anthropic closed a 65-billion-dollar Series H in May at a 965-billion-dollar valuation, briefly the most valuable private company on earth, on a revenue run-rate of 47 billion. OpenAI raised 122 billion in March, the largest private round in history, at 852 billion, on roughly 25 billion of run-rate. xAI folded into SpaceX at a 250-billion-dollar mark, with around half a billion of standalone AI revenue. Both leaders filed to go public a week apart in June. And the concentration is the real story: OpenAI and Anthropic together took 217 billion dollars in the first half of 2026, about 43 percent of every startup dollar invested on the planet, in a record half-year of 510 billion. And 2.07 trillion is roughly 260 dollars staked for every human alive, on three companies.
To keep some feel for the scale: Anthropic's revenue run-rate, the genuinely impressive 47 billion, equals the entire GDP of Estonia, and the three labs' combined 72 billion is roughly one Slovenia. Iceland's whole economy is smaller than either. The valuations, meanwhile, sit in a different universe than the revenue.
The scale problem44 Estonias of paper value, on one Slovenia of revenue.
Too big to grasp, on revenue too small to justify it
How big is this bet? Big enough that the usual yardsticks run out. TARP, the fund that stopped the 2008 banking crisis, was authorised at 700 billion dollars and ended up disbursing around 443 billion; the big three's paper value is three whole TARPs. The IMF eventually put the entire world's bank write-downs from the 2008 crisis at roughly four trillion dollars, and this single bet is already worth more than half of that. The 2023 mini-crisis that took Silicon Valley Bank, Signature and First Republic did it with about 549 billion dollars of assets between them, the worst bank-failure year on record; the big three are worth nearly four of those years stacked on top of each other.
These are the numbers we keep for things too big to fail. We are now using them to describe three companies that, between them, still lose money on most of what they sell. That is the shape of a bubble: a claim on the future priced as if the future were already here and already theirs.
And the money increasingly flows in a circle. NVIDIA is lined up behind hundreds of billions of dollars of OpenAI-related financing, money that comes back to it as GPU orders, and Bloomberg now keeps a running map of who in the Microsoft-OpenAI-NVIDIA triangle is paying whom. When the customer's purchases are underwritten by the supplier's guarantees, revenue stops being independent evidence of demand.
The bull case deserves a fair hearing before the verdict. The revenue is real and growing at rates software has never seen, Goldman calls the AI software sell-off overdone, and if enterprise adoption keeps compounding, today's multiples get retro-justified the way Amazon's 1999 multiple eventually was. It is not a stupid bet. It is a bet that only pays if intelligence stays scarce and rented, and scarcity is exactly what the rest of this part erodes.
The bubble has a physical footprint
You do not have to take a position on whether this is a bubble to see what it is doing to the physical world. AI data centres will consume around 70 percent of the world's memory output in 2026, up from a quarter a few years ago. The three manufacturers who control 95 percent of the DRAM market redirected their capacity to AI-grade memory, Micron left the consumer business entirely, and the spot price did this:
Prices keep climbing even as consumers hit their limit. A 32-gigabyte memory kit that cost 90 dollars now costs around 529. Gartner expects the crunch to add 17 percent to PC prices, call it seventy euros on a four-hundred-euro school laptop, and 13 percent to phones; SK Hynix's CEO expects 2027 to be the worst supply year in the industry's history. Read that as an ordinary citizen: your next laptop is more expensive because three companies are building server halls on borrowed conviction. The bubble is not an abstraction in a spreadsheet. It is a line item in every school's device budget.
The fall does not need a crash. It needs a download.
Now the other act, and it is not the one the shorts are betting on. While the big three were raising nation-state money, the open-weight world quietly closed the gap. On the Artificial Analysis index the frontier sits at 60 for Claude Fable 5 and 59 for GPT-5.6; Moonshot's Kimi K3, weights downloadable by anyone, scores 57. A year ago the best open models scored in the low thirties.
On coding benchmarks the gap is down to a couple of points at a fraction of frontier prices, and the estimate of how far open weights trail has shrunk from six-to-nine months to three-to-five. Fairness requires the counterweight: on human-preference leaderboards the top fifteen are still proprietary, on hard agentic work the gap widens to double digits, and the labs are actively closing the distillation shortcut. The frontier is real. The question is how much of the market needs the frontier, and the next chart answers it.
Eighteen months took the memory footprint of a frontier-class open model from 350 gigabytes to 24, which is precisely the size of the graphics card a teenager buys for gaming. Capability per gigabyte is improving faster than capability itself. Follow the dashed line one more year and premium-tier intelligence runs locally on consumer hardware, at which point "having AI" decouples from "renting AI". That is the moat those two trillion dollars assume, evaporating at a slope you can measure. The fall of the big three does not need a crash on a trading floor. It needs a download button, and the download button already exists.
The next escalation is on the calendar as I write this. Alibaba has promised the open weights of Qwen 3.8 27B for this very week, a model built precisely for the 24-gigabyte consumer-card class; the local-AI crowd greeted the announcement as "absolute nightmare fuel for the closed labs" and "a godsend for the 24GB VRAM class". The satire is already ahead of the hearings: a viral parody testimony has a lab CEO pleading with Congress that "a $900 GPU is now running frontier intelligence completely offline ... no provider in the loop", asking for "a thoughtful framework. For humanity." It is parody, and worth quoting anyway, because the anxiety it lampoons is the same moat problem this chart measures: once intelligence is a possession, the subscription loses its seat at every table, including the safety one.
And below the gaming card sits an even harder floor: silicon itself. AMD's freshly acquired Taalas etched Llama 3.1 8B directly into an ASIC, and it now answers at 17,000 tokens per second; in a side-by-side demo, the builders' own numbers rather than a third-party benchmark, the chip finished in 0.01 seconds what GPT-5.6 needed 0.23 for. The same corner of the internet is already asking for Qwen 3.8 on the same treatment at a casual 7,000 tokens per second. When a model becomes a chip, the marginal cost of intelligence becomes the marginal cost of a component, and nobody has ever paid a monthly subscription for a component.
Europe watches from the stands
And where is Europe in the biggest capital reallocation of the decade? Holding 19 percent of the tickets.
Europe invested 252 billion euros of venture capital against America's 1.33 trillion over 2020 to 2025. The EU has zero venture funds above five billion dollars; the US has more than fifty. In the first half of 2026 European AI startups raised 7 percent of the American total, and the US investors who do show up in European rounds take 69.2 percent of the deal value. Close to 30 percent of European-founded unicorns moved their headquarters abroad, overwhelmingly to the US, often because their American investors required it. Luis Garicano carries the Draghi report's sharpest line: no EU company worth more than 100 billion euros has been founded from scratch in fifty years, while all six American trillion-dollar companies were. And the money we do save flows west; as one European founder put it brutally: "I've divested completely from Europe because the ROI is so bad here." The home team gets its counterweight: 2025 was Europe's third-best venture year on record. Best of the rest is still the rest. And the craft side of the board is not empty: Mistral keeps shipping open weights from Paris and the EuroHPC machines are already public infrastructure. Europe's problem is capital, not capability.
So Europe cannot join the spending race. That may matter less than it looks, because the commodity wave described above is the one part of this story that favours the players without the capital. When frontier-class intelligence runs on a 2,000-euro card, the game stops being "who can borrow 122 billion" and becomes "who can integrate, host and trust the thing", and that is infrastructure, procurement and craft. Europe's chance is to be the best in the world at running what the big three commoditise: open weights on sovereign hardware, in gemeenten, hospitals and MKB, under our own law. The same lesson as the governance post: don't crown, layer.
How f*cked are we, then? This is a dangerous bubble, and its detonator is unusual. It will not be popped by a short-seller or an interest-rate move. It will be popped by a download. The entire two-trillion-dollar valuation rests on intelligence staying scarce and rented; the moment a good-enough model runs free on a card in a bedroom, the revenue thesis underneath it cracks. And a bubble this size does not deflate quietly. It is wired into the memory market, into the hundreds of billions the cloud giants are pouring into data centres, and into the pension funds quietly holding all of it: Dutch funds alone have more than 150 billion euros in tech stock, and ABP has already started trimming because the AI weight grew uncomfortable. Your retirement is in this trade whether you follow it or not. When it goes, it will feel like 2008, and the fuse is a free download that already exists.
What would change my mind? A year of honest, positive margins on inference sold at market prices, or the open-weight curve above stalling for a year. Watch those two lines, not the headlines.
The one consolation is the same fact from the other side: your workday survives the crash, because the thing that pops the bubble, cheap abundant intelligence, is the thing that keeps working on your own machine afterward. The valuation was the fragile part. The capability is not going anywhere.
Next, part three: the same subscription logic that reprices software has quietly taken almost everything you used to own, and that costs you more than convenience.
Sources
Figures dated 2026 were retrieved on 10 August 2026; the DRAM, open-versus-closed and VRAM trend curves are anchored to cited endpoints, and dashed sections are projections, not published series.
- CNBC. Anthropic tops OpenAI as most valuable AI startup, nears $1 trillion valuation. The $65B Series H at $965B and the $47B revenue run-rate.
- ValueAddVC. OpenAI valuation 2026: $852B, a 34x revenue multiple. The $122B round and the multiple mathematics.
- ValueAddVC. The highest-valued private AI companies in 2026. xAI at $250B inside the SpaceX deal, and both IPO filings a week apart in June.
- World Bank. GDP, current US$. Estonia at $47.03B, Slovenia at roughly $72B and Iceland at roughly $35B, 2025.
- CNBC. AI memory is sold out, causing an unprecedented surge in prices. Data centres consuming the world's memory output; the manufacturers' pivot to HBM.
- Tom's Hardware. Memory price surge begins to cool as consumers hit affordability limit. DRAM and NAND prices climbing through Q3 2026.
- Technology.org. Why your next laptop costs more: inside the AI memory shortage. The consumer fallout: PCs +17 percent, phones +13 percent.
- Interconnects (Nathan Lambert). Kimi K3: the open-weights escalation. The open frontier at 57 against closed at 59 and 60.
- The Rundown. Moonshot's Kimi K3 closes the frontier gap. Near-frontier scores at a fraction of frontier prices.
- AI Weekly. Kimi K3 puts open weights within months of frontier. The gap estimate shrinking from six-to-nine months to three-to-five.
- Latent Space. Qwen 3.8 Max and 27B: new open weights for coding and cowork. Alibaba's weights promised for the week of 10 August 2026, sized for consumer hardware.
- Alok (X). "Absolute nightmare fuel for the closed labs". The local-AI community's reception of Qwen 3.8 27B for the 24GB VRAM class.
- sudoingX (X). The parody testimony. Satire, quoted as satire: no such statement was made to Congress, which is exactly why the joke lands.
- Robert Sterling (X). European sovereign fund counts. Zero EU funds above $5bn against more than fifty in the US.
- AI_4_Healthcare (X). Cumulative VC 2020–2025. €252bn versus €1.33tn, and the 24 versus 42 percent household-equity split.
- TFN (X). European AI funding H1 2026. $23bn against roughly $330bn, and US investors at 69.2 percent of European deal value.
- ContRiftShow (X). Unicorn relocations. Close to 30 percent of European-founded unicorns moved HQ abroad, 2008–2021.
- Mio_Mind (X). US VCs and the HQ requirement. American investors routinely requiring a US headquarters before investing.
- Luis Garicano (X). The Draghi line. No EU company above €100bn founded from scratch in fifty years.
- levelsio (X). The divestment quote. "I've divested completely from Europe because the ROI is so bad here."
- Dealroom (X). Europe's 2025 venture year. The fair counterweight: $63.7bn raised, the third-highest on record.
- US Treasury. Troubled Asset Relief Program (TARP). Authorised at $700bn in 2008, roughly $443bn disbursed.
- IMF. World Economic Outlook, 2009 crisis-era estimates. Global bank write-downs from the 2008 crisis estimated toward $4tn.
- American Banker. 2023 was the biggest year ever for bank failures. SVB, Signature and First Republic, about $549bn of combined assets.
- NPR. NVIDIA is about to spend $750 billion on AI. Critics are calling it a bubble. The vendor-financing wave behind the revenue.
- Bloomberg. AI circular deals: how Microsoft, OpenAI and NVIDIA keep paying each other. The running map of the money circle.
- NOS. Pensioenfonds ABP trekt zich nog verder terug uit de VS. ABP trimming US tech, its NVIDIA stake nearly halved.
- Financieel Management. Zorgen over kwetsbaarheid pensioenfondsen voor AI-zeepbel. Dutch pension funds' 150-billion-plus tech exposure.
- Conduction ConNext. "This is going to hurt" (2026 talk). The slide deck behind this series, including the Crunchbase-based concentration figures and the banking-crisis comparison.