An $8 Billion Loss and a $2 Trillion Ask, as the Federal Door on Anthropic Faces Its September 30 Deadline

The leaked IPO prospectus that Reuters published on September 28, 2026 puts Anthropic's 2025 operating loss at $8.06 billion on revenue of nearly $4.6 billion, and the company's timing could hardly be worse. Five days earlier, on September 25, 2026, the D.C. Circuit in Anthropic PBC v. U.S. Department of War, No. 26-1049, upheld the government's FASCSA supply chain risk designation, and the Department of War's own compliance deadline to terminate all Anthropic related activities falls on September 30, 2026. A company preparing to ask public investors to value it above $2 trillion is entering its roadshow with one of its growth channels, the federal one, being closed on schedule.

Start with the numbers, because they are the first audited shaped look at any frontier lab's books. Revenue rose roughly twelvefold in 2025 to nearly $4.6 billion, according to the prospectus reviewed by Reuters. The operating loss widened to $8.06 billion from $2.98 billion a year earlier. Spending on computing and infrastructure ran $7.33 billion in 2025, triple 2024's level, and the company has committed to $518 billion in future cloud, computing and infrastructure obligations, per the filing. Cash, cash equivalents and short term investments stood at $20.28 billion at the end of December 2025. Two customers accounted for close to a quarter of 2025 revenue, and the prospectus concedes that many of the largest customers are not locked into long term contracts. For the biggest deal sitting behind that obligation stack, see <a href="/research/anthropic-akamai-11-6-billion-cpu-turn-warrant">Anthropic&#x27;s $11.6 Billion Akamai Deal Is the CPU Turn, Priced With Equity</a>.

Those are remarkable numbers by any historical standard. A twelvefold revenue year justifies a rich multiple on its own, but the cost structure tells you the pricing question is really a compute question. Anthropic's $518 billion of committed future obligations is larger than the annual output of most mid sized economies, and it is contractual, not aspirational. Revenue has to compound at rates the industry has never sustained for the company to grow into that fixed cost base. Investors who buy at the rumored valuation are underwriting both the growth and the obligation stack, and the prospectus is now public evidence of how lopsided that stack is.

Now layer in the legal development, which changes the composition of the revenue base rather than its headline size. The FASCSA order lets the Department of War bar Anthropic products from defense contract performance. The majority, in an opinion by Judge Katsas joined by Judge Rao with Judge Henderson dissenting, held that model training itself, because it embeds restrictions on what Claude will do, can constitute manipulating the design and operation of a covered product under 41 U.S.C. 4713. In plain terms, the court accepted that safety guardrails, the very feature that defines Anthropic's brand, can be treated as unwanted functionality when the buyer is the Pentagon. Holland & Knight's September 29 alert for government contractors notes the practical scope: without written direction from a contracting officer, most contractors face no affirmative duty to strip Anthropic from their commercial environments, but the enforcement trend and the September 30 termination deadline point one way. The capex cycle driving that fixed cost base is the same one in <a href="/research/ai-capex-power-bottleneck">Five AI Deals in One Day, and Not One Was About GPUs</a>.

There is a confusing wrinkle that matters for anyone tracking this. A different court already ruled the other way. On August 27, 2026, the Northern District of California invalidated the parallel designation under 10 U.S.C. 3252, finding it likely retaliatory against Anthropic's public AI safety advocacy and arbitrary and capricious under the Administrative Procedure Act. The D.C. Circuit did not reject that reasoning; it distinguished it, because FASCSA's definition of supply chain risk covers any person, not just adversaries. The result is a split record: one statutory theory struck down, a second upheld, with an en banc rehearing petition possible since the judgment was stayed seven days. For a public market investor, the relevant fact is that the most restrictive theory survived appellate review while the company's own disclosure says its government business faces termination deadlines.

What does the government channel actually weigh against a $4.6 billion revenue base? Probably modestly at first. The blocked business is concentrated in defense contract performance, and the Northern District's earlier injunction preserved the status quo on the broader federal directives. But the signal costs more than the direct revenue. The designation makes Anthropic the first US frontier lab treated in procurement documents like a suspect vendor, a category normally reserved for companies tied to adversarial governments, as Defense One's September 25 report observed. Enterprise buyers in regulated industries read procurement signals. If the company's own prospectus flags customer concentration, two customers near a quarter of revenue, and short contract durations, an adverse procurement finding compounds the same risk from a different direction: the revenue is concentrated and one concentrated buyer segment is legally pushing it out.

The prospectus itself is candid in ways that should reframe how this IPO is read. Reuters and subsequent coverage noted the filing warns that autonomous systems could behave in unexpected ways, create security problems, and even refers to existential risks to humanity. A company that lists civilizational downside among its risk factors, while it solicits a $2 trillion valuation, is asking investors to price a probability distribution no comparable public company has ever disclosed. That is not disqualifying. It is unusual, and it means the standard comparables, the Nasdaq mega caps that the listing will sit beside, are the wrong benchmark for evaluating whether the price is sane.

The bull case writes itself from the same page. Twelvefold revenue growth, a product set in Claude Code and the enterprise tier that created a compute shortage for the company itself, $20.28 billion of cash to fund the burn, and strategic investors in Amazon and Google who supply the infrastructure. If the November listing prices below the whispered $2 trillion, the risk and reward start to look more symmetric than the current pre IPO market, where traders are pricing above $2 trillion according to pre IPO market coverage this week. The same frontier-lab funding math shows up in <a href="/research/softbank-pays-9-75-percent-to-close-openai-bet">SoftBank Pays 9.75% to Close Its OpenAI Bet</a>.

The bear case is the document itself. An $8.06 billion operating loss on $4.6 billion of revenue means gross economics are not yet visible. The $518 billion commitment stack means the next several years of cash flow are spoken for before a single discretionary dollar. Customer concentration and contract duration disclosures remove the soothing assumption that enterprise AI revenue is sticky. And now the court record shows a federal buyer with statutory authority, appellate backing, and a deadline that lands on September 30.

My read: the prospectus leak, whether intentional or not, does IPO investors a service. It converts a valuation debate into an underwriting question with stated inputs. The D.C. Circuit ruling and the September 30 termination deadline strip away the last comforting assumption, that government friction would resolve before the roadshow. It did not. Anyone underwriting this listing should model the federal channel at zero and ask whether the remaining growth justifies the price. On the disclosed numbers, that is a harder question than the pre IPO market's $2 trillion answer suggests.

This content is for informational purposes only and does not constitute financial advice. Past performance is not indicative of future results. Consult a qualified financial advisor before making investment decisions.