<h2>The loop</h2>
Here’s the structure, as plainly as I can put it: Google, Amazon, and Microsoft have collectively spent somewhere north of $800 billion building AI data center infrastructure. The overwhelming majority of that capacity — estimates range from 75% to 85% — exists to serve two customers: OpenAI and Anthropic. Both companies lose money at scale. Both depend on continuous capital infusions from the same hyperscalers building the infrastructure. The hyperscalers then reinvest in the labs to keep them solvent, which keeps the compute bills paid, which justifies more infrastructure spending.
That’s not a market. That’s a self-referencing accounting loop.
Ed Zitron has been making this argument in detail, and the numbers he cites are hard to dismiss. Outside of OpenAI, Anthropic, and the hyperscalers themselves, he can’t identify a single GPU compute customer spending more than $100 million per year. The entire demand story — the one justifying 15.2 gigawatts of data center capacity under construction by 2027 — rests almost entirely on two deeply unprofitable companies.
Anthropics’s own projections show $11 billion in annual losses in both 2026 and 2027, with $86 billion in planned training costs through 2029. The company has raised roughly $58 billion in eight months. When xAI reportedly handed off its entire 300MW Colossus-1 data center to Anthropic, the implication was uncomfortable: even xAI, flush with Elon Musk’s backing, apparently had no use for that much compute.
<h2>OpenAI’s numbers</h2>
The Wall Street Journal reported in early 2026 that OpenAI missed internal targets for both revenue and user growth. ChatGPT fell short of a goal of one billion weekly active users by year-end. CFO Sarah Friar reportedly warned company leadership that revenue may not grow fast enough to cover future computing contracts. Sam Altman and Friar pushed back publicly, calling the report “prime clickbait.”
Markets didn’t buy it. Oracle fell 7.7% on the news. CoreWeave dropped 7.4%. SoftBank lost nearly 10%. Chip stocks declined 2-6% across the board.
The reaction makes sense once you understand the dependency chain. OpenAI has committed to roughly $600 billion in compute spending through 2030. Its current revenue sits around $25 billion annually. To justify those commitments, it needs to reach $280 billion in revenue by 2030 — more than a tenfold increase. Anthropic, meanwhile, has now surpassed OpenAI in annualized revenue at $30 billion, and Google’s Gemini is gaining consumer share. OpenAI’s CFO has reportedly questioned the company’s readiness for a 2026 IPO.
Oracle’s situation illustrates the concentration risk clearly. The stock has dropped nearly 50% from its September high despite 41 of 51 Wall Street analysts rating it a buy. The $300 billion Stargate deal with OpenAI is simultaneously Oracle’s biggest growth story and its biggest liability — OpenAI is both a customer and an investment of the companies funding it, including Oracle itself. The stock now moves primarily on OpenAI news rather than Oracle’s own fundamentals.
<h2>The PE workaround</h2>
Both labs appear to have recognized that traditional enterprise sales cycles aren’t going to get them to the revenue numbers they need. Their response is interesting.
OpenAI has finalized a $10 billion joint venture called The Deployment Company, backed by 19 private equity firms including TPG, Brookfield, Bain Capital, and Advent International. The structure is unusual: OpenAI commits up to $1.5 billion while guaranteeing PE backers a 17.5% annual return over five years. In exchange, the PE firms open their portfolio companies as a captive customer base. OpenAI engineers embed directly inside client organizations across healthcare, logistics, manufacturing, and financial services — essentially the Palantir forward-deployed-engineer model.
Anthropichas a parallel $1.5 billion deal with Blackstone, Goldman Sachs, and Hellman &amp; Friedman.
The logic is straightforward: instead of convincing enterprises one by one, you buy access to entire portfolios at once. The risks are also straightforward: the guaranteed-return structure will attract regulatory attention, large-scale software rollouts inside PE portfolios are notoriously difficult to execute, and OpenAI’s upside is capped if the venture actually works well.
But the deeper question these deals raise is whether they represent genuine demand creation or just another layer of financial engineering on top of the existing loop. PE firms aren’t buying because their portfolio companies are clamoring for AI tools. They’re buying because OpenAI is guaranteeing their return. The end customers — the portfolio companies — are essentially captive.
<h2>What’s actually at stake</h2>
I find myself genuinely uncertain about how this resolves. The bull case is that AI infrastructure demand is real and fungible — that even if OpenAI stumbles, someone else will absorb the capacity. The bear case is that the entire buildout was sized for a demand curve that doesn’t exist outside the circular financing structure.
What’s harder to dismiss is the math. If you can’t identify meaningful compute customers outside the named parties, and the named parties are losing money while being subsidized by the same companies building the infrastructure, then the demand signal is noise. The 15.2 gigawatts under construction by 2027 needs customers. Right now, the honest answer is that nobody can clearly identify who they are.

Collections

Google, Amazon, and Microsoft have collectively spent over $800 billion on AI data center infrastructure, with 75-85% of capacity serving just two customers: OpenAI and Anthropic — both of which are deeply unprofitable and dependent on capital from the same hyperscalers building that infrastructure. OpenAI missed internal revenue and user growth targets, needs a tenfold revenue increase by 2030 to justify compute commitments, and its CFO has questioned IPO readiness. Anthropic projects $11 billion in annual losses through 2027. Both labs are turning to private equity workarounds — OpenAI via a $10B joint venture guaranteeing PE firms 17.5% annual returns, Anthropic via a $1.5B Blackstone deal — to manufacture enterprise demand rather than earn it organically. The core concern: outside the named parties, no meaningful GPU compute customers spending over $100M/year can be identified, raising serious questions about whether the entire infrastructure buildout is sized for demand that doesn't exist beyond circular financing.

OpenAI missed revenue and user growth targets as Anthropic closes the gap

The interesting part isn’t whether OpenAI is “in crisis.” It clearly isn’t. The real question is whether the AI infrastructure trade was priced for a growth curve that’s becoming harder to defend.
$25B revenue is massive, but if reports of ~$600B in compute commitments and a $280B 2030 revenue target are directionally right, that requires near-perfect execution across consumer, enterprise, coding, agents, and API usage.
Anthropic gaining ground matters because coding and enterprise workflows are exactly where durable, high-value AI revenue lives.
My take: this is the AI boom entering its unit-economics phase.
2023–2025 was: “Who has the best model?”
2026 is becoming: “Who can turn usage into profitable, defensible revenue fast enough to justify the infrastructure buildout?”
Curious how others see it: normal hypergrowth turbulence, or the first real stress test of the AI capex trade?