SEC vs Anthropic's IPO: An $895B Compute Promise

Opening Dialogue

Shiwen: On Monday, November 2nd, Anthropic submitted an amended Form S-1 to the SEC, including financial information for the third quarter and plans for a listing in November. However, the dates, price range, and number of shares were still placeholders in the document. On the same day, TechPolicy.Press published an article with the title “The SEC Should Stop Anthropic’s IPO to Protect the Public,” written by Matt Scherer from the Open Markets Institute. This issue has limited discussion in the tech community; the post on HN received only over forty likes and a dozen comments, with most attention focused on financial and policy media. Tell us, what exactly does this article advocate?

Yongliang: In one sentence: The author believes the SEC should require Anthropic to provide audited financial reports for two consecutive quarters, along with full disclosure of expenditure commitments, before allowing the listing. The rule cited is that when a registration statement is “seemingly incomplete or significantly inaccurate,” the SEC has the right to delay or even prevent it. It’s important to note that this is an opinion article; the author is an opinionated researcher, not an official action by the SEC. The SEC hasn’t done anything yet.

Shiwen: Let’s present both sides’ documents. First, look at the numbers in the documents.

Q1: What are the numbers in the S-1?

Shiwen: Which financial figures in the amended S-1 are the most striking?

Yongliang: According to the amended S-1 and reports from CNBC and Axios, third-quarter revenue was approximately $17 billion, with a cumulative total of $33.24 billion over nine months, a year-over-year increase of about 1200%. The growth rate is very strong, but the quarterly trend is slowing down: 132% in the first quarter, 144% in the second quarter, and dropped to 47% in the third quarter. TechPolicy.Press says “summer growth almost stagnated,” and both sources agree.

Shiwen: What about losses? I was shocked by headlines like “Net loss of $104 billion over nine months.”

Yongliang: This number needs to be broken down. A net loss of $23.29 billion in the third quarter, with a total net loss of $104 billion over nine months. But most of it is non-cash losses from changes in the fair value of convertible bonds and derivative liabilities. In simple terms, the more rapidly the valuation rises, the more losses are recorded on the books. The actual operational figure reflects the business: a net operating loss of $567 million in the third quarter, while the second quarter had an operating profit of $974 million. A quarterly shift from profit of $974 million to loss of $567 million is more worth paying attention to than a net loss of over 100 billion. Over the next nine months, calculations and infrastructure expenses increased by $14.4 billion, a rise of 329%.

Shiwen: There are two versions of cash reserves?

Yongliang: Yes, there’s a discrepancy in the figures. The S-1 states $131.8 billion in cash and short-term investments as of September 30th; TechPolicy.Press’ article mentions $96 billion. It might be a difference between “cash” and “cash plus short-term investments,” or it could be due to different time points. I’ve listed them here side by side without ruling out which one is correct. There’s an even more fundamental issue: the article points out that the financial results for 2026 have not been formally audited, and the S-1 itself contains over 90 pages of risk factors.

Q2: What does the commitment of $895 billion mean?

Shiwen: What is that astronomical figure, 895 billion, in the article title?

Yongliang: According to the revised S-1, the combined purchase commitments and lease agreements amount to approximately 895.4 billion US dollars, almost entirely related to computing and infrastructure. The breakdown is as follows: about 447.5 billion will be paid by the end of 2030, with the remaining 447.9 billion spread over 2031 to 2045, with nearly 70 billion due in the next 12 months. TechPolicy.Press added that at least 582 billion is an unfulfillable commitment to five companies: Amazon, AMD, Broadcom, Google, and Microsoft. There’s also a detail: the initial draft stated this amount as around 450 billion, while the amendment directly doubled it to 895 billion. The article speculates that the contract commitments by the end of the year could exceed one trillion, noting that this is the author’s speculation, not the content of the document.

Shiwen: Having managed a team for seventeen years, from an engineering management perspective, what does this commitment mean?

Yongliang: It means there is a forced gap between the scale of commitments and delivery capabilities. Over the years as technical director, the most expensive lesson I’ve learned is that signed procurement commitments do not equal production capacity, and production capacity does not equal revenue. The article provides a reference: a 1GW data center requires approximately 31 billion US dollars, and Anthropic has committed to 15GW, estimating a final need of 20 to 25GW. At this unit price, just 20GW would require an amount of capital expenditure in the six-hundred-billion range. And its revenue? Only 33.2 billion in nine months. The commitments are at the twenty-year scale, while revenue is earned quarter by quarter, relying on financing and cash flow to hold it together. If it holds up, it’s an epic infrastructure project; if not, it’s the largest breach in history.

Shiwen: Does the article also mention pressure from the demand side?

Yongliang: Yes, the author listed several points: multiple studies say that large models have not provided positive ROI for most companies; competition for commercialization has intensified, making it difficult to distinguish between model quality; this summer, Anthropic and OpenAI’s price war continues, and open-source models and on-prem models in different markets are also pricing lower; there are external pressures from communities where data centers are located, supply chain tensions, and geopolitical conflicts that drive up costs. The article even characterized the flagship model Fable 5 as having a “cold” commercial response, which is the author’s statement, I’ve quoted directly.

Q3: Is the demand real? The perceived ROI of corporate AI

Shiwen: Do you have first-hand experience with the argument that “LLM’s have not provided positive ROI for most companies”?

Yongliang: Yes, but let me clarify, this is my experience, not statistics. The article cited a study saying that AI usage in the U.S. workforce dropped from 46% in the summer of 2025 to 33% in the summer of 2026. My own team has also experienced a similar curve: a year ago everyone was excited, with everyone having a subscription; last year we started reducing, only retaining those scenarios that were truly integrated into workflows. In my team of seventy people, only about one-third are still frequently used, which coincidentally is close to that 33% figure. Most pilot projects fail for one reason: the time saved did not turn into money saved because the processes didn’t change, just with an extra step of “letting AI see first”.

Shiwen: Is Anthropic’s revenue structure healthy?

Yongliang: According to S-1, about 80% of revenue comes from pay-per-use API and model calls, and 20% is from subscriptions. The advantage of pay-per-use is that it increases with customer business, but the downside is that it decreases with business fluctuations, with no fixed minimum contract. Concentration is improving, but there are still concerns: in the first half of 2026, about 20% of revenue came from just two customers; in the third quarter, the top five customers accounted for about 21%, thankfully none exceeded 10%. The number of customers is substantial, with over 11,900 customers spending more than $100,000 annually and nearly 1,900 customers spending over $1 million. The net revenue retention rate is about 900%, note that this reflects an increase in existing customers, not an increase in the number of customers.

Shiwen: I don’t quite understand that part about the cloud market, where 48% of revenue comes from AWS, Google Cloud, and Azure?

Yongliang: This is what the article refers to as the circular trading loop. These three cloud providers are also Anthropic’s computing power suppliers, investors, and competitors. Money circulates like this: cloud providers invest in Anthropic, Anthropic uses this money to purchase computing power from them, customers buy Anthropic models through the cloud market, and revenue returns. Each step exists, but combined, it’s difficult for outsiders to distinguish which parts are end-user demand and which are internal ecosystem transactions. This is what I think needs the most auditor signatures throughout S-1.

Q4: “Retailers Taking Over” and Governance Structure

Shiwen: The article title says “protecting the public,” but what are the specific concerns?

Yongliang: The author’s main concern is that once listed, shares will enter 401(k) plans, pension funds, and index funds. If the AI bubble bursts, “workers, retirees, and retailers will be the ones taking over when the bills come due.” He also believes that U.S. stocks are historically overvalued and concentrated among a few interconnected tech giants, and adding another non-profit giant would amplify the risk of contagion. In a discussion thread on HN, someone asked this question: Will pension funds “automatically” buy Anthropic because the S&P 500 is included? Someone corrected that index inclusion usually happens periodically, not buying immediately after listing.

Shiwen: Are there any opposing views?

Yongliang: Yes, another user in the HN comments has a very typical stance: Let the market handle it. If it crashes after listing, we’ll get a lesson worth billions of dollars. This view actually makes sense; the SEC’s historical role is to ensure sufficient disclosure, not to judge whether something is valuable for investors. Stopping a company that disclosed over 90 pages of risk factors sets a precedent for “regulatory valuation by others.” My compromise view is that sufficient disclosure and credible disclosure are two different things; unaudited financial reports claiming sufficient disclosure lack confidence.

Shiwen: What about governance structure? If retailers buy shares, does that matter?

Yongliang: Not really. Anthropic is a public-benefit company with dual ownership: an institution called Long-Term Benefit Trust (including former Federal Reserve Chairman Bernanke) will appoint most of the board; the founders hold Class F shares through Founder LLC, with 50.1% voting rights; public shareholders get one Class A share per share, with no control rights. Buying its shares is buying a proxy to “believe in the trust and founders,” not a right to participate in governance.

Shiwen: What signals are there from valuation and issuance sides?

Yongliang: According to reports, existing investors expect the valuation to be between $1.5 trillion and $1.75 trillion, which is lower than the over $2 trillion discussed in August. Private placement valuations have reportedly reached $965 billion. The issuance side is very active. According to Reuters, Japan’s investor quota is up to $3.5 billion, exceeding the $2.2 billion for SpaceX’s IPO in June. Steve Eisman, the prototype of “The Big Short,” also discussed this IPO in a program on October 31st. I haven’t verified his exact words, so I won’t quote them here.

Q5: The position of employees, stockholders, and ordinary people

Shiwen: What does this IPO mean for Anthropic’s own employees?

Yongliang: Forbes mentioned an often overlooked point: the IPO timing may add another year of tax burden for employees. I have experienced two company listings before. Here’s my advice based on experience, not statistics: When the tax treatment of equity incentives is misaligned between attribution and realization, employees may face a situation where “the book value increases while the cash tax bill arrives first.” RSUs are taxed at market value, but the shares may still be in lockup. My simple advice for ordinary engineers is: Attributate and sell part of it to cover taxes. Never bind your wealth and employer’s stock, even if you deeply love the company.

Shiwen: What about retail investors considering buying shares or already purchasing?

Yongliang: Three things. First, separate promises from deliveries: $895 billion is a contract signed, not completed production capacity, nor money received. Only nearly $70 billion due in the next 12 months is a real constraint. Second, watch the audits: The requirement for “two consecutive quarters of audited financial reports” is a very low threshold. If you go in without even this, what you’re buying isn’t stocks, it’s narrative. Third, look at your investment channels: If your exposure comes from index funds and pensions “passively,” you haven’t even exercised your option. This is exactly the scenario the author fears, and also the part opponents deliberately skip when saying “the market will resolve it.”

Shiwen: So your overall judgment: Should the SEC intervene?

Yongliang: I don’t think the SEC should judge the value of investments for investors. But “audited financial reports” and “complete disclosure of promises” belong to disclosure quality, which is precisely the SEC’s role. The threshold the author advocates is, from an engineering perspective, a pre-launch acceptance standard: You can say the product will change the world, but the acceptance report must be signed first. $895 billion in promises, twenty-year leases, one-quarter operational fluctuations, combined with an unaudited financial report—this combination would fail any technical review. As for whether it’s a bubble, let the market answer; but for the market to answer, it must first get credible information.

---

Author: Yongliang