AI IPOs

Anthropic has taken the first formal step toward an IPO, but investors still do not have a public prospectus or offering price. The company announced on June 1 that it confidentially submitted a draft Form S-1 to the SEC. It explicitly said the share count and price had not been set. The detailed numbers now circulating come from Reuters’ review of that nonpublic draft, not from an issuer filing available for independent inspection. They are important, but provisional.

The investor question is straightforward: can demand for Claude grow fast enough, and at attractive enough economics, to support the computing capacity Anthropic is committing to buy? The reported draft gives evidence on both sides. Revenue has risen sharply, while operating losses and long-duration infrastructure commitments have also grown. A public S-1 will be needed to test the economics properly.

Our prediction

By 2028: At least one of Anthropic and OpenAI will reach the public market, but the central investor debate will shift from revenue growth to the cost and rigidity of compute commitments. Public filings will force closer comparison of cash generation, customer retention, dilution, and governance; a large private valuation will not by itself set the IPO price.

What would confirm it: A public S-1 and offering with audited economics that make infrastructure obligations and shareholder rights central to valuation.

What would prove us wrong: Neither company lists by the end of 2028, or public filings show that compute commitments are readily covered by durable operating cash flow and have little bearing on valuation.

What is public, and what is reported

Anthropic also said in May that its annualized revenue run rate had crossed $47 billion earlier that month. That is a measure of the pace at that point in time, not booked revenue for a completed year. Comparing it directly with 2025 revenue can show acceleration, but it cannot establish full-year 2026 sales, retention, margins, or cash generation.

Why the compute commitments matter

Reuters reports that the confidential draft describes at least $111.1 billion of long-term obligations with Google, $110 billion with Amazon, $31.4 billion with Microsoft, and approximately $161.2 billion of Broadcom-related equipment leases. These agreements span different periods, principally seven to ten years. The same report describes an xAI arrangement for up to $84.5 billion of capacity that is largely cancelable with notice, and an AMD arrangement involving potential equity investment and compute supply. Those latter arrangements should not be treated as equally firm obligations or simply added to the $518 billion total.

Securing scarce compute capacity could protect Anthropic’s ability to serve customers and train new models. It also fixes a large part of its cost base ahead of demand. If usage, pricing, or model economics disappoint, commitments payable regardless of usage would magnify the downside. The public prospectus needs to show the payment schedule, minimum commitments, lease accounting, financing, and how much capacity is already contracted to paying customers.

There is a second dependency. Amazon, Google, and Microsoft are simultaneously infrastructure suppliers, distributors, investors or counterparties, and AI competitors. Reuters says the reported draft flags these overlapping roles as a risk. That makes the durability and terms of access to compute as important as the headline size of the contracts.

What could make the investment case work—or fail

The positive case is that Claude becomes a durable work platform. Anthropic’s May financing announcement points to rapid enterprise adoption and a much higher run rate than its 2025 reported revenue. If paying usage fills reserved capacity and costs per useful task fall, today’s infrastructure spend could support a much larger business.

The risk case is that commitments grow faster than profitable demand. Reuters reports that nearly a quarter of 2025 revenue came from two customers and that some large customers lacked long-term spending commitments. Customer concentration and cancellable demand sit uneasily beside largely fixed compute obligations. Competition could also push model prices down before those obligations expire.

Governance deserves its own review. Reuters reports that the draft proposes a Founder LLC with a Class F share carrying 50.1% voting power over specified matters. This is a reported proposal, not a final public charter. Investors will need to compare the precise voting rights, board powers, and public-benefit obligations with the economics of the shares offered.

Reuters also reports a possible valuation above $2 trillion. That is a reported expectation, not an announced target range or a price investors can presently evaluate. No ticker, exchange, listing date, share count, dilution schedule, or use-of-proceeds table has been publicly established by Anthropic’s IPO announcement.

The next document that changes this analysis

The first public issuer S-1 on SEC EDGAR should replace secondhand figures in this note. We would examine its audited statements and cash flows, gross margin and compute-cost classification, 2026 interim revenue, customer and channel concentration, commitment maturities, related-party arrangements, share classes, dilution, and IPO proceeds. We would also compare every amendment with the initial public filing rather than treating an old draft as final. SEC guidance says the registration statement and earlier nonpublic drafts generally become public at least 15 days before a road show, or before effectiveness if there is no road show.

Current view: Anthropic’s demand signal is unusually strong, but the evidence is not yet sufficient to value its public shares. The most consequential mismatch to investigate is the reported scale of mostly fixed, long-term compute obligations against still-unproven public-market economics. This is a research watch item, not a buy or sell recommendation.