Wink Pings

Anthropic's IPO Debut Wealth Creation: How Much Is Its AI Safety Brand Worth Against a $900B Compute Bill

Anthropic, the pure AI lab behind the Claude large language model, became the first to go public. On its debut day, its share price surged from $83 to $111, creating thousands of millionaires. However, its future $900 billion in compute spending, shareholder profit demands, and its core AI safety positioning create persistent tension. Whether its unique governance structure can uphold safety standards will depend on answers from upcoming financial reports.

Synthetic image of stock exchange and Anthropic CEO Dario Amodei
Image credit: The Verge / Photos: Getty Images

Last Thursday, Anthropic, the AI lab that developed the Claude large language model, listed on the Nasdaq, becoming the first pure AI lab to trade on public markets.

With an IPO price of $83 per share, the share hit a high of $111 during its first day of trading, catapulting its market cap into the ranks of the world's most valuable companies. A host of law firms serving it soon posted congratulations on their official websites, while thousands of new millionaires would be created internally, along with hundreds of centimillionaires and even more billionaires.

The hype around the stock is exciting, but the bill is a reality. According to its S-1 filing, Anthropic will need to pay around $900 billion in total compute costs over the next few years. It must continue investing in cutting-edge model research, implementing AI safety measures, while generating enough revenue to cover compute expenses and deliver sufficient returns to public market shareholders. The combination of these three things creates a delicate tension.

These discussions come from an exclusive interview with Hayden Field, a senior AI reporter at The Verge, on the Decoder podcast, and is the final episode of the media series "AI in Reality."

Why Is OpenAI Willing to Be the Second?

Many wonder why OpenAI, another leading AI lab, let Anthropic seize the first IPO opportunity.

Hayden Field revealed that Sam Altman had set being first to go public ahead of Anthropic as a core goal over the past year. But this year, OpenAI faced a string of negative news, with public discontent toward the AI industry mostly centered on it. The team finally adjusted its pace and stepped back to observe market reactions.

Anthropic's IPO performance served as a market test for the entire AI industry. After historical levels of subscription and share growth, OpenAI can simply apply the same valuation logic when pitching its own IPO story, saving much of the cost of educating the market.

Making Money from Tokens Is Not Easy

Setting aside the valuation story, return to the basic business logic. Users purchase a token—the smallest billing unit for large language model processing—and Anthropic's production cost per token must be lower than its selling price to form a positive cycle. For now, Anthropic is still far from this goal.

Investor Steve Rattner estimated that Anthropic's revenue needs to grow 17 times over the next 10 years to reach the average return level of comparable tech companies.

It currently has roughly three revenue pathways.

The first is adjusting pricing to pass on compute costs to downstream users. Previously, it raised personal subscription prices, and faced a class-action lawsuit from users due to service cuts, finally issuing a public apology and rolling back some price adjustments. Cost pressures have not disappeared, and it may still pass on costs through package adjustments, tiered pricing, etc.

The second is building user migration barriers, or "moats" as the industry calls it. It aims to deeply bind users' daily workflows with its model capabilities and tool ecosystem, making users feel that migrating to other platforms would cost more than the price difference.

The third is pursuing large enterprise and government customers. These customers have strong payment capacity, low price sensitivity, and prioritize compliance, data security, and stability. Anthropic's long-standing "safe and trustworthy" label is essentially a calling card for these clients.

The actual market trends often deviate from AI labs' marketing rhetoric.

Anthropic executives have publicly stated that enterprises will always choose the most cutting-edge models to stay competitive, as using cheaper models would put them at a disadvantage against rivals. But Hayden Field spoke to many small and medium-sized enterprises that are doing the opposite.

These companies build internal evaluation systems to split business tasks by difficulty. Usually, only 10% of high-demand tasks call Anthropic's expensive cutting-edge models, 40% of tasks use open-source models, and the remaining 50% use mid-range models with better cost-effectiveness. The entire set of evaluation data is internal cost-sensitive information and not disclosed to the public.

Microsoft CEO Satya Nadella recently expressed a similar view: in the future, every enterprise will have its own model evaluation standards, always choosing the most suitable and lowest-cost models, and will not pay for unnecessary cutting-edge performance.

Founders' Control Is More Concentrated Than Advised

Beyond business pressure, Anthropic's unique corporate governance structure has also drawn attention.

It has established the Long-Term Benefit Trust (LTBT), which claims to have the power to appoint and remove most board members, monitoring founders to make decisions around long-term public interests rather than being driven by short-term capital gains.

According to the S-1 filing and former employees, the actual power structure differs from the public narrative.

Currently, the trust has four members, including former Federal Reserve Chairman Ben Bernanke, foreign policy analyst Richard Fontaine, experimental physicist Christopher Stubbs, and former GiveWell executive Neil Badhey Shah. None have deep experience in the AI industry. All members are selected by the founders, with former employees revealing that the selection criteria favor those who will not raise strong objections.

In terms of authority, the trust cannot intervene in daily operations—even having no power to halt the release of risky models. It can only initiate director appointments and removal processes, requiring a 30-day advance notice to the board committee led by founders.

Of the seven board seats, the trust can nominate three, the founders can nominate three, and the final tie-breaking seat—though nominally nominated by the trust—also requires founder approval to take effect.

In addition, the seven founders hold a special share of Founders LLC, which has 50.1% voting rights. Ordinary public shareholders have no power to replace management or easily intervene in company decisions. The entire setup leaves founders with almost unrestricted control.

Inherent Contradictions in the Safety Line

Many worry that this structure will erode Anthropic's safety foundations.

Motivationally, the founder team's investment in AI safety is not marketing rhetoric. Multiple core members have publicly expressed concerns about potential consciousness in AI, with their commitment to safety approaching a belief. They built this control structure precisely to prevent shareholders from forcing the team to abandon safety principles for short-term profits.

But this logic comes with built-in contradictions. The team's internal consensus is that only by leading the industry can they qualify to discuss safety. To win the race, any safety investment that slows progress may be sacrificed, and all decisions can be justified by the claim that "to ensure long-term safety, we must first win."

Many former employees fear that in the end, Anthropic will become more like the competitors it originally sought to differentiate itself from, with the safety label slowly turning into a pure marketing symbol.

The Nvidia Variable

Anthropic's IPO process and computing supply chain are deeply tied to Nvidia.

Nvidia is not only its core GPU supplier, but also finalized a $100 million private share purchase before its IPO. Combined with its previously promised $100 million in compute investment, their interests are highly aligned.

Recent news that Jensen Huang stepped down as Nvidia CEO for violating company conduct rules will also indirectly affect Anthropic's cost structure. Previously, many AI infrastructure financing collaborations and chip capacity adjustments were personally negotiated by Huang with financial institutions and cloud providers. It remains unclear whether the new management can maintain the same resource mobilization capabilities.

If Nvidia adjusts its channel strategy to divert chips from gray markets back to the U.S. market, compute supply will increase, lowering Anthropic's procurement costs. If infrastructure financing tightens and compute prices stay high, its profit pressure will further rise.

Signals Worth Watching

For ordinary observers, daily share price fluctuations have limited reference value. Hayden Field, a veteran financial reporter, believes that short-term stock prices of AI companies are mostly driven by sentiment and rumors, and have little correlation with fundamentals.

The real things to monitor are the details in each quarterly earnings report and earnings call. For example: has the safety team been downsized? Are there any safety-related R&D projects terminated? Can revenue growth rate keep up with continuously rising compute expenses?

These subtle changes in the corners of financial reports are the core signals for judging how much Anthropic's safety brand is actually worth.

Decoder podcast logo

The full interview is available on the following podcast, with content condensed and audio quality optimized:

发布时间: 2026-11-21 23:00