Goldman Sachs's 2025 Tracking Trillions report is an imposing document. It projects that global AI infrastructure investment could reach $7.6 trillion by 2031, driven by hyperscaler capex, data center buildout, and the insatiable compute demands of model training and inference. It acknowledges, almost as a footnote, that AI is currently costing the majority of businesses more than it saves them. The market, largely, chose to focus on the $7.6 trillion.
Ed Zitron spent a week trying to find the enterprise customers who would account for that demand. He hosts the Better Offline podcast and writes extensively about technology's relationship with money and truth. He looked specifically for evidence of non-AI-company enterprises renting the kind of large-scale GPU clusters that would justify the buildout projections. He found, with a few exceptions, almost nothing.
"The demand does not exist," he said on a recent episode. "I've looked everywhere."
The Illusion of Demand
Zitron's argument begins with a structural observation: the AI industry's apparent demand signal was manufactured, not organic. When OpenAI and Anthropic were founded, they were never required to be efficient, sustainable, or profitable. Microsoft built OpenAI's infrastructure at subsidized rates. Amazon and Google did the same for Anthropic. These arrangements meant that in the early years, OpenAI and Anthropic were simultaneously the biggest AI companies and effectively their own largest customers — paying back money that had already been loaned to them.
The rest of the market watched this and drew a conclusion: if the leading AI labs need that much compute, everyone else will too. Data center builders, Nvidia, cloud providers, and investors all calibrated their projections to a world where thousands of enterprises would eventually need the kind of infrastructure that only OpenAI and Anthropic actually needed — and only needed because they were subsidized to build it.
"Every time you look for the supposed gold mine, you find someone just handing someone money instead of the gold coming out."
Ed Zitron, Better Offline podcastWhen DeepSeek launched in January 2025, the market treated it as a story about China. Zitron reads it differently. DeepSeek trained a frontier-competitive model using roughly 50,000 H100s — a fraction of what OpenAI and Anthropic were using. The message wasn't "China can do this cheaply." The message was "nobody needed that many GPUs to begin with." The compute that justified $300 billion in infrastructure investment turned out to be, in significant part, waste.
Following the Money Backward
Nvidia shipped roughly six million GPUs last year. If you try to account for where they went, the math gets uncomfortable quickly. The verified large-scale customers are a very short list: OpenAI, Anthropic, xAI (which built the Colossus 1 cluster and then rented its entirety to Anthropic, which itself is telling), and the major hyperscalers building out their own model training capabilities. Beyond that list, Zitron found "a couple hundred, couple thousand GPUs at most" among even relatively large AI companies like Perplexity.
The neocloud sector — companies like Iron, Nebius, Cipher Mining, and Terowolf — was supposed to represent the next tier of compute demand. Many are former crypto operations that repositioned as AI infrastructure plays. Several have required hyperscaler bailouts to survive. These are not the organic enterprise customers that the $7.6 trillion projection requires.
| Entity | Role in the System | Financial Reality |
|---|---|---|
| OpenAI | Primary AI lab; largest apparent compute consumer | No path to profitability; $108B+ in committed funding required |
| Anthropic | Primary AI lab; rented entire xAI Colossus data center | Massive revenue growth matched by even larger compute costs |
| Microsoft / Google / Amazon | Hyperscalers financing the labs; also building internal AI | Justified capex via equity stakes; not yet seeing enterprise AI payoff |
| Nvidia | GPU supplier; the only entity consistently profiting | Revenue growing; dependent on this cycle continuing |
| Neoclouds (Nebius, Cipher Mining, etc.) | Supposed next tier of compute demand | Multiple required hyperscaler backstops to survive |
The Circular Logic of Hyperscaler Capitalism
Zitron traces the structure of the AI bubble back to 2019. When Nvidia acquired Mellanox — an Israeli networking company that makes the high-speed InfiniBand connections required for large-scale AI clusters — Satya Nadella and Jensen Huang announced an unspecified partnership shortly after. OpenAI received a billion dollars in compute credits. The relationship between Microsoft, Nvidia, and OpenAI wasn't an accident. It was, Zitron argues, a designed circular flow: Microsoft feeds compute to OpenAI, OpenAI generates apparent demand for Nvidia, Nvidia's revenue growth justifies Microsoft's capex, Microsoft's AI story justifies its stock price.
"Everyone's involved in the scam," he said. "Big shout out to [reporter] Daario, who was basically the only person who talked about this."
The hyperscalers' thesis was that they were raising the AI labs as children who would eventually leave the house and keep paying rent — that OpenAI and Anthropic would grow into self-sustaining profitable enterprises that would keep buying compute at scale. That moment has not arrived. OpenAI and Anthropic have each raised hundreds of billions in committed funding and still have no clear path to profitability. The children are still living at home, and the rent they're paying is coming from the same parents who gave them the down payment.
What Would Actually Break This
Zitron's view is that the market is waiting for a signal that cannot be ignored, and that the demand not being there is not, by itself, sufficient. "The market's no longer listening to subtlety or even reality," he said. He has a specific thesis about what would constitute an undeniable signal: three simultaneous events he calls his horsemen — a data center under construction being cancelled, a planned data center being cancelled before breaking ground, and a fully operational data center closing. Together, he believes, those three events would force a repricing of the entire narrative.
None of those three events has happened yet. Data center construction is proceeding. Nvidia's revenue is growing. The $7.6 trillion projection is being treated as a floor, not a ceiling.
How to Hold This
The honest position here is that both things can be true simultaneously. AI is genuinely transformative. The technology is real. The applications are real. The jobs being displaced and created are real — the WEF's 92M/170M projection is not a fiction. The people using these tools to do things that weren't possible two years ago are not experiencing a hallucination.
But the financial structure wrapped around that genuine technology has the shape of a bubble. The companies at the center of it are losing money at extraordinary scale. The demand that justifies the infrastructure buildout is heavily concentrated in entities that are themselves kept alive by the hyperscalers funding the buildout. And the analyst projections are calibrated to a demand curve that exists primarily on paper.
Zitron has been wrong about specific timelines before. He's been wrong about specific events before. What he has been right about is the structural pattern: the circularity, the lack of enterprise customers, the fact that the model labs have no business model that produces profit at any realistic scale. The question isn't whether those structural facts are real. They are. The question is how long that gap between structural reality and market narrative can persist — and what closes it.