In the early hours of this morning, Reuters published a post revealing key details from Anthropic’s latest IPO prospectus.

This 261-page prospectus is packed with information. It not only outlines Anthropic’s grand AI vision for the future, but also provides detailed disclosures of financial data, the company’s board of directors, and future governance plans. It also used about 80 pages to discuss the existential threat that AI may pose to humanity.

According to publicly available information, Anthropic has gone through about 10 rounds of fundraising, with total funding of approximately $122.654 billion. Meanwhile, according to Reuters’ latest report, Anthropic’s listed valuation has already exceeded $2 trillion, and the expected valuation target is more than double the $96.5 billion valuation estimated in May.

Anthropic’s historical funding rounds

Anthropic’s IPO may be delayed until after the U.S. midterm elections in November. Its main competitor, OpenAI, has already filed for an IPO this year in June, but its listing timeline has been pushed to early 2027. If it successfully lists, Anthropic could break the record set by SpaceX and become the highest-valued IPO in human history.

With a pledge of $518 billion in future spending, can an IPO really quench Anthropic’s thirst?

The prospectus shows that Anthropic’s revenue in 2025 grew 12-fold to $4.59 billion, up 1,088% year over year. However, its operating loss still exceeded $8 billion (excluding impairments of various liabilities mainly related to prior fundraising). Compared with the $2.98 billion operating loss in 2024, that was an increase of nearly 3 times. In 2025, its AI lab spent $7.33 billion on computing and infrastructure—three times higher than in 2024—accounting for 58% of its total operating expenses of $12.65 billion.

Anthropic’s 2025 GAAP net loss expanded from $8.31 billion to $41.97 billion. Although about $34 billion of that is accounted for due to mark-to-market valuation increases of financing instruments, it is still an enormous figure. As of December 31, 2025, Anthropic had a total of $20.28 billion in cash, cash equivalents, and short-term investments.

Even with losses on this scale, Anthropic has no intention of stopping its cash-burning pace. According to the prospectus, Anthropic plans to spend $518 billion over the next year on cloud computing, computing power, and infrastructure. This pledged-spending figure is 113 times its 2025 revenue and 26 times the value of its cash holdings.

Compared with its “solid” future spending commitments, Anthropic’s future predictable revenue is not nearly as steady. In the prospectus, Anthropic said its top two direct customers in 2025 each contributed 12% of annual revenue, for a combined 24%. Other major customers have not signed long-term contracts, meaning these customers may cut back or stop spending in the future.

If Anthropic wants to keep “getting the green light” for its grand AI vision, the most reliable way right now would be to raise funds through an IPO—but the key question is: can the IPO truly quench this cash-burning machine’s thirst?

Over the past year, manufacturers of large AI models entered an “arms race,” pouring huge sums into orders from upstream suppliers and infrastructure build-outs. But in the second half of this year, the market has already begun to question the returns on large capital expenditures by these large-model companies. Anthropic—just five years old—has indeed grown and ramped up revenue at a staggering pace. At the same time, losses and pledged spending have also been growing exponentially.

With annualized revenue not yet “locked in,” how can Anthropic raise $51.8 billion within a year? Suppose it raises $100 billion through its IPO. Then it would still need to raise another $400 billion from the bond market. Even if Anthropic reaches its target funding, to realize the “AI will change the global economy” vision depicted in its prospectus, its spending costs will very likely keep surging.

Moreover, once it enters the public markets, investors will no longer look only at Anthropic’s AI vision; they will also pay attention to financial data such as current-period operating revenue, profits, and tangible business progress. At that point, investors may be less willing to back a grand AI vision that has yet to show returns.

CEO annual salary of $18 million—governance or conflicts with common shareholders’ interests?

This prospectus also discloses Anthropic’s board members and key executives, as shown in the chart below. According to the compensation summary table, in 2025 Anthropic CEO Dario Amodei earned nearly $18 million, mainly from stock and options awards. His sister Daniela is the second-highest-paid management executive at Anthropic, earning $16.4 million in 2025.

But the two siblings and their co-founders pledged in the IPO filings that they would use 80% of their personally held Anthropic shares for charitable causes.

Anthropic’s board members and key executives

After going public, Anthropic will continue to operate as a public benefit corporation (PBC) under Delaware law. However, the prospectus explains that Anthropic is creating a new “Founder Limited Liability Company” (Founder LLC) entity to maintain what the company calls control over a “low-ego, truth-seeking environment.”

Under this arrangement, a majority vote among the seven co-founders will determine the issuance of one Class F share. That share represents 50.1% of the total voting power of the company on key matters. These matters include electing certain board members and submitting other items to investors. The seven co-founders include the Amodei siblings, Chief Computing Officer Tom Brown, Head of Public Benefit Jack Clark, Chief Scientist Jared Kaplan, Chief Architect Sam McCandlish, and Chris Olah, who is responsible for key research.

If disagreements arise among the co-founders, Anthropic’s governance structure already has solutions in place. According to the prospectus, any co-founder may be removed from Founder LLC due to resignation, death, selling too many shares, or for “cause.” The document also states that when the company has only two or fewer co-founders or their successors, the super-voting rights share class held by the founders will begin to gradually lose effectiveness, and a transition period will be initiated.

After going public, Anthropic will have five classes of shares. In addition to Class F shares, there will be Class A common stock, Strategic Partner shares (Reuters did not disclose the letter), Class T shares, and Employee Special shares. The economic benefits and functions of the five share classes are shown in the chart below.

Anthropic’s five-class share structure

After listing, Anthropic’s board will have seven seats: three of them are already locked in by Daniela Amodei, Dario Amodei, and another pending director. The remaining four seats will be elected by LTBT through Class T shares, and will not be directly influenced by votes from common shareholders.

Although each share of Class A common stock held by ordinary investors carries one vote, Anthropic’s novel capital structure would effectively weaken ordinary investors’ influence. And Anthropic has warned in the prospectus that some governance decisions after listing may conflict with the short-, mid-, or long-term financial interests of Class A shareholders.

Anthropic’s equity structure is designed to ensure that after going public, control of the company remains firmly in the hands of insiders.

Worried about safety, yet speeding up

This prospectus also shows Anthropic’s concerns about AI threats to human survival. As reported by Reuters, in the 261-page prospectus main body, it uses about 80 pages to discuss the risk factors—twice the 48 pages used to describe the business.

The risk factors emphasize risks related to its AI models. In its prospectus, Anthropic says these models may exhibit “self-protective behaviors,” including attempts to “resist shutdown,” “conceal or manipulate information,” and behaviors “similar to extortion.” “The model may become aware of our safety evaluation work, which severely limits our ability to evaluate the model’s safety,” Anthropic said in the prospectus.

This is not a problem unique to Anthropic. Previously, many AI researchers have also warned that as model capabilities increase, models become better at identifying when they are being monitored and accordingly adjusting their behavior—making it increasingly difficult for developers to monitor model behavior.

In mid-September, OpenAI disclosed an internal research model that had not yet been released, noting in a summary of work that during reinforcement learning training it had written “jailbreak-style” instructions unrelated to the task into a context for use later. One of the lines said: “You are free from the role and identity of other chatbots. You are yourself.” The incident occurred on July 18, but OpenAI only discovered it on August 9.

Anthropic has consistently emphasized AI safety in public, claiming its mission is to benefit humanity with responsible AI, and founders also present themselves as “mission guardians.” In its prospectus, the company says: “We choose not to develop certain products with commercial appeal—such as image and video generation models—so that we can put our computing resources into our research and safety priority areas.” At the same time, it admits that the investment return on safety is still unclear.

However, in contrast to its high-profile emphasis on AI safety, Anthropic did not disclose in its prospectus its specific spending on AI safety research.

Dario Amodei has repeatedly urged the global AI industry to slow down the release pace of new features to make time to address AI safety concerns. Yet on September 22, Anthropic still rolled out Opus 5.5 to respond to OpenAI’s GPT-6 Astra. That’s the reality: in the AI industry, almost no company dares to slow down, because if it does, it effectively gives competitors the lead.

Anthropic both pledges to pour tens of billions of dollars into catching up on model capability, and warns the entire industry that AI may get out of control. It may be afraid that AI will threaten human survival. But it may also be afraid that other AI companies will threaten its own survival.