Anthropic is weighing whether to reserve 15% to 20% of the shares in its initial public offering for retail investors, according to people familiar with the matter. JPMorgan is expected to lead the US retail portion of the offering. The company’s S-1 prospectus does not commit to a specific retail allocation; it says only that the banks may set aside shares for their online brokerage customers. The reported plan would put Anthropic in the same range as SpaceX, which set aside about 20% of its June listing for retail investors after first planning 30%.

The retail allocation is more than a populist gesture. In large technology IPOs, banks and issuers use retail demand to broaden the shareholder base and to support the deal’s pricing. One market research analyst observed that individual money is less sensitive to price, and a book that leans on it lets banks price the deal higher than institutions alone would support. That dynamic matters for Anthropic because the company is seeking a valuation that could approach $2tn, a figure that would make it one of the most valuable companies ever to list in the US.

Warren sets a 3 November deadline

Senator Elizabeth Warren has told Anthropic to answer four pages of questions about its IPO filing by 3 November. The Massachusetts Democrat sent the letter to chief executive Dario Amodei on Tuesday. She asked how Anthropic can ensure its technology will not become an existential threat to humanity. She also asked about its ties to Amazon and Broadcom and how it will justify a valuation that could approach $2tn. Anthropic did not immediately respond to a request for comment.

Warren wrote that the risks demand greater scrutiny before they become intertwined with the public capital markets. The letter adds a political dimension to a listing that was already unusual for its scale and for the sensitivity of the industry. Anthropic is not a conventional software company. It is a frontier AI developer whose products are trained on enormous compute clusters and whose safety claims are central to its public identity. A public offering would give outside shareholders a claim on both the commercial upside and the safety risks described in the filing.

The timing is notable. The letter arrived as Anthropic was preparing to explain its business to public investors, and it asks for answers within days. It also places the company’s relationships with Amazon and Broadcom under direct scrutiny. Amazon is both an investor and a cloud partner; Broadcom is a semiconductor and infrastructure supplier. Those ties are central to Anthropic’s ability to secure compute, but they also raise questions about dependency, related-party arrangements, and the terms on which the company buys capacity.

The risks in the filing

The filing notes concerns, inside Anthropic and outside it, about where advanced AI could lead. It sits alongside the warning that Anthropic could become a political target. Sufficiently advanced AI systems, the filing says, could, if not developed, deployed, and operated with appropriate safeguards, oversight, and security measures, pose catastrophic or existential risks to humanity. That language is not a side note. It is part of the risk profile that public investors are being asked to accept.

The document also notes that proposed US laws contemplate the nationalisation or mandatory shutdown of frontier AI systems. It adds that the rules Anthropic itself argues for could slow its growth. Government contracts make up about 1% of its revenue. The combination is awkward: Anthropic advocates for regulation that it says is necessary for safety, while acknowledging that the same regulatory environment could constrain its expansion, invite political intervention, or force changes to its operations. For public-market investors, that creates a policy risk that is difficult to model alongside ordinary commercial risks.

The broader debate over AI regulation gives the filing context. Lawmakers and regulators have been debating how to oversee frontier models, whether to require licensing, and how to handle national-security concerns. Some proposals would give governments emergency powers over advanced systems. Anthropic has positioned itself as a safety-focused developer, but its own risk factors acknowledge that a public company could face pressure from politicians, regulators, and activists who want it to slow down or change course.

The prize: knowledge work and a $2tn valuation

The same document sizes the prize. Anthropic says knowledge work was worth more than $40tn globally in 2025, which tops SpaceX’s $28.5tn market estimate. The company is arguing that AI can capture a meaningful share of that economic activity by automating and augmenting tasks across law, finance, software, research, and administration. That is the bull case for a valuation that could approach $2tn: not a niche software market, but a slice of global white-collar work.

The scale of that ambition helps explain why the retail allocation is being discussed. A listing of this size would require enormous demand from institutions, and retail participation could add another layer of demand. It could also create a more visible shareholder base and generate public attention. But retail-heavy allocations can cut both ways. Individual investors may be less price-sensitive in the short term, but they can also be more volatile after listing, especially if the stock trades far above or below its offer price.

More compute deals

Anthropic has signed a series of compute agreements that underpin its growth plans. It has reported deals with Google, Amazon, Microsoft, and Broadcom. It also agreed to spend up to $120.5bn with SpaceXAI, the SpaceX unit formerly known as xAI. That reaches about $3.1bn a month from December. After early 2027, either side can cancel on 90 days’ notice.

• Up to $120.5bn with SpaceXAI, about $3.1bn a month from December, cancellable after early 2027 with 90 days’ notice.
• Up to $44.6bn with Nscale.
• About $41.3bn with Lambda.
• $125bn to $150bn for more TPU chip capacity from 2028 to 2031, outside its commitments table.
• Since 30 June, about $245bn in minimum payments across signed deals.

The total commitment figure is still blank. Reports have suggested that Anthropic could spend $750bn on compute. The numbers illustrate the central economic bet of the frontier AI industry: companies must secure massive amounts of computing power years in advance, even though the revenue that will justify that spending is uncertain. Long-term contracts can lock in supply and pricing, but they also create fixed obligations that public investors will scrutinise. The cancellation terms matter because they determine how much of the commitment is truly binding if demand slows.

Margins and who pays

Cost of revenue, mostly compute that serves customers, was $2.38bn in the first half, against $16.3bn of revenue. That is a gross margin of about 85%, a figure that has been called Anthropic’s key number. It leaves out training. Total compute spending was $8.95bn. The distinction matters: serving customers can be highly profitable at scale, while training frontier models is a separate and expensive race. A company can show strong gross margins on inference and still burn large amounts of cash on training and infrastructure.

The filing is cautious about future growth. It says the company does not expect revenue to continue growing at comparable rates, and it cannot assure investors that revenue will continue to grow at all. Margins may also shrink as new compute arrives before the revenue it supports. Models released in the six months to 30 September made up over 80% of third-quarter revenue. That shows how quickly the product cycle turns over, and how dependent the business is on shipping new models that customers adopt.

Sales through Amazon, Google, and Microsoft’s cloud marketplaces were 48% of first-half revenue, up from 11% in 2023. Google has an option to buy up to $10bn of stock at the IPO price. Anthropic values the convertible notes Amazon holds at $108.6bn; Amazon puts them at $97.9bn. The gap in valuation is a reminder that the company’s most important partners are also counterparties with their own interests. Cloud marketplaces give Anthropic distribution, but they also concentrate its revenue and give its partners leverage.

When insiders can sell

Some shares can be sold from the first trading day. Co-founders and holders of more than 0.5% must wait until after third-quarter 2027 results, or 365 days. Most others are free after first-quarter 2027 results, or 180 days. Lock-up terms are closely watched in IPOs because they determine when a wave of new supply can hit the market. A long lock-up for founders and large holders can signal confidence, but it also delays the moment when insiders can convert paper wealth into cash.

Critics and investors are already debating the filing. One prominent critic wrote that the company described in the S-1 no longer exists, pointing to the blank third-quarter figures. One investment research firm called Anthropic a much, much better business than it had expected, but said it would not yet buy. Those reactions capture the tension in the deal: the financial metrics are strong enough to attract attention, but the risk factors, the scale of the compute commitments, and the unanswered questions from Washington leave plenty of room for caution.

Source: TNW | Anthropic News