Wall Street Built a Perpetual Motion Machine for AI — Nvidia’s May 20 Earnings Will Test Whether Physics Still Applies

Imagine an alchemist who claims to have built a furnace capable of turning a single gold coin into three: he drops the coin in, three coins come out, and he proves the trick by weighing the output. The audience applauds. Nobody asks where the metal came from.

This, more or less, is the story Wall Street has chosen to tell itself about artificial intelligence in the spring of 2026. Nvidia has invested roughly $30 billion in OpenAI and $10 billion in Anthropic. OpenAI and Anthropic use a large slice of that capital — and of every fresh round on top of it — to buy Nvidia GPUs. The chips become record revenue for Nvidia. The record revenue justifies a multi-trillion-dollar market cap. The market cap funds the next round of investments. The next round of investments returns to Nvidia. Three coins out of one.

You can call this a flywheel. Michael Burry calls it «a picture of fraud, not a flywheel», and has put $186.6 million in Nvidia puts and $912.1 million in Palantir puts behind the claim. He may be early. He may be wrong. But on Wednesday, May 20, 2026, after the closing bell, Nvidia will publish its first-quarter fiscal 2027 results, and the market will, for the first time in months, be forced to weigh what is actually inside the furnace.

The closed loop has a name and a number

The contours of the loop are not hidden. Anthropic is in talks to raise as much as $50 billion at a $950 billion valuation — a figure that, if confirmed, would push the Claude maker past OpenAI’s $854 billion mark and place it among the most valuable private companies in history. Google has pledged up to $40 billion in compute and capital, Amazon up to $25 billion. SpaceX and Anthropic recently announced a joint AI infrastructure project anchored on more than 220,000 Nvidia GB300 GPUs in a single data center. Each of those deals reads like a customer order. Each of them is also, in part, an investment by Nvidia in its own demand curve.

For a primer on how this skepticism has been building since 2024, see our long-running thesis that there is no real market for AI yet. The Anthropic numbers have aged that argument badly — annualized revenue tripled from roughly $9 billion at the end of 2025 to over $30 billion this spring, gross margins on inference reportedly above 70%, more than 1,000 enterprise customers spending over $1 million a year. The market is real. The question is who, exactly, is paying for it — and whether the same dollar is being counted twice.

Bertrand Russell’s turkey, in a Santa Clara office park

In 1912 Bertrand Russell described an inductivist turkey: fed at 9 a.m. every morning for a thousand days, the bird concluded with growing confidence that feeding would continue. On day one thousand and one, it was Thanksgiving. The lesson — repeated by Nassim Taleb in italicoThe Black Swan/italico — is that the longer a streak of confirming evidence runs, the more conviction it generates, and the worse the surprise on the day the regime breaks.

The AI complex looks, today, very much like day one thousand. Nvidia posted $68.13 billion in revenue for the fourth quarter of fiscal 2026, up 73% year over year, with the data center segment alone delivering $62.3 billion. The Zacks consensus for the May 20 print sits at $78.75 billion of revenue and $1.77 of earnings per share, with the options market pricing a roughly 8% post-print move in either direction. Morgan Stanley still ranks Nvidia, Microsoft, and Alphabet at the top of its list of AI stocks to own. Every metric points the same way. The turkey has never been more confident.

The Cisco analog, and what it actually means

Defenders of the trade will point out — fairly — that the demand is real. Enterprises are signing seven-figure contracts. Inference loads are real. Training runs are real. None of that was true of pets.com or Webvan in 1999. This is not, the bulls insist, that bubble.

That is, almost word for word, what defenders of Cisco Systems said in late 1999. Cisco was the dominant supplier of internet plumbing, beat earnings consistently, served customers whose demand was unquestionably real, and peaked at $80.06 a share on March 27, 2000. By October 2002, it had fallen 88%. A quarter-century later, the stock only just clawed back to that bubble high. Cisco’s mistake was not selling a hallucinated product. Its mistake was selling a real product at a price that assumed the regime — capex inflation across the entire telecom sector — would never end.

Burry’s argument, stripped of theatrics, is the same: «sometimes the new company is the same company on a pivot». Vendor financing, then and now, accelerates the cycle on the way up and on the way down. When Cisco lent its customers the money to buy Cisco gear, demand looked structural. When the lending stopped, so did the demand.

The catalyst, and the math

Jensen Huang told the Morgan Stanley TMT conference on March 4, 2026 that Nvidia’s investments in OpenAI and Anthropic would likely be «the last» of their kind, since both companies are tracking toward IPOs that will close his window. Read charitably, that is a CEO maturing past speculative bets. Read uncharitably, it is the supplier quietly turning off the financing tap before the customers find new ones — and, perhaps, before April’s hotter-than-expected 3.8% CPI and a 10-year Treasury at 4.59% start to make the cost of borrowed compute genuinely uncomfortable.

The May 20 print is therefore not really a referendum on Nvidia. It is a referendum on whether the loop still closes. A clean beat-and-raise — call it $80 billion in revenue with second-quarter guidance comfortably above $85 billion — keeps the alchemist on stage. A merely-in-line print with cautious commentary on hyperscaler capex, or a single sentence from Huang acknowledging that customer financing terms are tightening, would be enough to put a meter on the furnace.

When the meter goes on

Every closed-loop scheme in financial history — vendor financing in 1999, structured credit in 2007, crypto exchanges lending to their own market makers in 2022 — has worked beautifully until somebody measured the energy entering and leaving the box. The italicophysics/italico of capital, like the physics of heat, is brutally simple: you cannot extract more value from a system than enters it. The appearance otherwise is always, in the end, an accounting artifact.

Wall Street has spent eighteen months treating AI as a perpetual motion machine. On Wednesday at 5 p.m. Eastern, somebody is going to put the meter on. If the reading still says the device works, we keep dancing. If it doesn’t, we will remember — as we have every cycle since the Dutch tulip — that the only thing perpetual about a perpetual motion machine is the embarrassment of having believed in it.