The Zhitong Finance App learned that “big short” investor Michael Burry (Michael Burry) said that Anthropic's IPO prospectus did not change his bearish view on AI deals; even as technology advances, it could not resolve his concerns about financing, infrastructure spending, and growth.
This contrasts with the enthusiasm of retail investors. On Stocktwits, retail sentiment surrounding Anthropic was “extremely bullish” and the volume of news was “extremely high”; while the overall market was significantly darker: SPDR S&P 500 ETF (SPY) and SPDR Dow Jones Industrial Average ETF (DIA) sentiment were both “bearish” and the volume of news was “normal,” and Invesco QQQ Trust (QQQ) was “extremely bearish” and the volume of news was “high.”
Whether the technology is good or not has nothing to do with the shorting argument
Bury made this statement in response to the prospectus in its Substack chat. “There's a lot of content here, but nothing changed my mind,” he said. “What's happening is far beyond what this S-1 can cover.”
This remark comes from a reader asking whether advances in the Big Language Model (LLM) have changed his views on this technology. Readers wrote that since 2024 and 2025, everyone said LLM would hit a wall or hit the platform stage, “but it really didn't, and it's still improving exponentially”. The narrative is still strong and continues to fuel more construction and investment.
“My opinion in a nutshell is: technology is improving, and this has nothing to do with any shorting arguments,” Barry answered.
He immediately listed the financial and industrial risks underpinning his position: “I'm looking at financing, the leverage in the fuse, the latest chips not being powered up to support demand for older chips, a retaliatory return of DRAM production capacity adapted to HBM, compression on open weight models and the spread of small models (SLM), and not just companies, but the fundamentals behind the transactions they make.”
He also pointed out that there was a mismatch between the funding period and the project being funded, and then went back to the difference between technological progress and growth: “I never said LLM would stop progressing. I think they're already being used more efficiently, and this trend will continue — it's bad for their growth rate.”
Understanding Barry's point of view
Bury's list of risks can be broken down into five, each pointing to the same judgment: the pressure is not on model ability, but on the financial and hardware structure that supports this round of construction.
First, financing and “cover leverage.” The original article did not explain the latter; in a literal sense, it points to leverage in AI infrastructure financing through insurance funds or special purpose carriers; Burry juxtaposed it with overall financing, indicating that he was not concerned about a certain company's balance sheet, but how the funds entered the industry.
Second, the chip cannot be connected to power. The latest-generation chips are unable to meet the requirements for electricity, computer rooms, and grid-connection, so demand for the previous generation chips is “floating” — this not only boosts the book demand for old chips, but also means that the actual pace of deployment of new chips may be overestimated.
Third, the retaliatory regression of memory. The DRAM production capacity he transformed for high-bandwidth storage (HBM) has not disappeared, and will return once the demand structure changes — a direct counterbalance to the recent “when the price increase cycle will peak” debate in the storage chain.
Fourth, model compression and the spread of small models in an open weight ecosystem point to a decrease in the cost of computing power per unit, which in turn reduces the pricing capacity of cutting-edge models.
Fifth, not only the company, but also the fundamentals of the transaction itself — that is, the bundled orders and investments between cloud vendors, chip vendors, and model companies, the quality of which needs to be tested separately.
What does Barry think is missing from the prospectus?
Regarding Anthropic itself, Burry said the prospectus did not reflect recent changes in the developer market. “There were many changes in the developer landscape between the end of the second quarter and the third quarter, and these are not in this prospectus,” he said.
He also questioned the document's treatment of the decline in “tokenmaxxing” (that is, the practice of maximizing AI token consumption). “The overall slowdown in growth was not attributed to a drop in tokenmaxxing, but I didn't expect it to mention that,” Barry said. “The Information reports that after tokenmaxxing was restricted, Meta's Claude Code users dropped from around 60,000 to 30,000.”
This contrasts with the prospectus's own statement. Anthropic clearly stated in the document: “We cannot guarantee that (revenue) will continue to grow,” while stating that it is expected to achieve adjusted operating profit for the second consecutive quarter.
The prospectus itself: promises high growth, huge non-cash losses, and huge computing power
Anthropic submitted S-1 registration documents to the US Securities and Exchange Commission on October 26 (Monday). It plans to list on the NASDAQ Global Select Market under the code “ANT”. The document does not determine the number of shares to be issued and the price range. Goldman Sachs and Morgan Stanley lead 24 underwriters.
Financially, revenue for the first half of 2026 was $16.286 billion, compared to $1,163 million in the same period last year; revenue from 2023 to 2025 was $41 million, 386 million, and $4.585 billion, respectively. Net loss for the same period was US$80.748 billion, of which approximately US$72.388 billion came from changes in the fair value of financial instruments such as convertible notes and derivative liabilities, which were non-cash accounting losses; operating losses narrowed from US$2.35 billion in the same period last year to US$693 million, adjusted operating profit of US$1.62 billion, and net cash flow from operating activities corrected to approximately US$3.9 billion. As of June 30, cash, cash equivalents and short-term investments totaled approximately US$131.65 billion. The second quarter had quarterly revenue of US$11.56 billion, and the company recorded its first quarter operating profit of US$974 million.
The cost side is the key. Computing power and infrastructure spending in the first half of the year was about US$8.945 billion (about US$2.3 billion in the same period last year); the company has promised to spend about US$518 billion on cloud services and computing power in the future, including the agreement with Amazon extended to April 2036, amounting to at least US$110 billion, and the agreement with Google until July 2033, at least US$11.5 billion. In terms of customer structure, in the past 12 months up to September 30, there were more than 11,900 customers paying more than $100,000 and more than 1,900 over $1 million; the net revenue retention rate was about 900%. The concentration was equally impressive: the two unnamed customers contributed about 23% of total revenue in the first half of the year.
At the valuation level, some investors have given valuations as high as 2 trillion US dollars, and some investors prefer about 1.5 trillion US dollars; if it exceeds 2 trillion US dollars, it will surpass the scale of about 1.77 trillion US dollars when SpaceX went public in June this year, making it the largest IPO in history. The prospectus uses a large amount of space to suggest risks, including that advanced AI may pose a “disastrous or existential risk” to humans, and that liability for extreme damage may be beyond the scope of any insurance.
The governance structure is also noteworthy: public investors can buy shares, but they cannot dominate the company — one Class F share held by the seven co-founders holds 50.1% of shareholders' voting rights, and a trust elects a majority of the board of directors; a charity fund will hold 5% of the company's equity with unvoting shares, and each co-founder promises to donate 80% of the individual's shares for public welfare purposes. Additionally, Amazon holds $6 billion convertible into non-voting shares, capped at 12% of the company's notes — the financing arrangement and shareholding structure are two sides of the same thing here.
This is the point of Barry's “unrelated to technology” statement: revenue has increased tenfold in ten months, and operating profit can be achieved in a single quarter, but what supports it is a promise of future computing power of more than 500 billion US dollars, procurement agreements deeply tied to major cloud vendors, and a valuation that has yet to be determined by the market.
In other words, this prospectus puts the differences in AI transactions on the same table: revenue curves, customer retention, and first quarterly operating profit; on the other, committed expenses, customer concentration, and financing structures. Bury is betting on the latter — he has never denied that the model is getting better; he believes that getting better itself is depressing the growth rate of this business.
Bury's historical position
This statement is a continuation of his existing views. Earlier on X, Bury said that Anthropic's valuation would be enough to buy 78 profitable S&P 500 companies, including Domino, Clorox, and Lululemon; he also wrote, “For the benefit of humanity, the market should fall sharply to prevent OpenAI and Anthropic's IPOs.” In September of this year, he said that he had brought forward the AI downward timeline and used put options instead of direct short positions.
“If spending stops or slows down, everything will fall apart,” he said at the time.
Anthropic raised $65 billion in May at a valuation of $965 billion. Bury compared this valuation to the largest inflation-adjusted pre-IPO valuation between 1990 and 2000 — the $119 billion United Package, which had been in existence for 92 years, with a price-earnings ratio of 26 times, a net interest rate of about 8.6%, and a market-sales ratio of 2.4 times.