SDSignal Desk

Anthropic’s safety promise now answers to one real boss: its compute bill

· Signal Desk newsroom

Analysis of reporting by The Verge. How we work

Image: The Verge

Shareholders can’t fire Anthropic’s founders and the trust above them has thin powers. That leaves the market and a giant compute bill as the only checks on whether safety survives going public.

Key takeaways

  • Anthropic priced its IPO at $83 a share and traded as high as $111 on its first day, making it the first pure AI lab on a public exchange.
  • Per The Verge’s reading of the S-1, a single share held by Founder LLC, co-owned by the seven founders, carries 50.1% of the voting power.
  • The Long-Term Benefit Trust has four members and no say over everyday decisions such as model releases; it can only appoint or remove certain directors.
  • Hayden Field puts Anthropic’s compute commitments at roughly $900 billion over the coming years, and Dario Amodei has called compute the company’s “deepest business dilemma.”
  • The signal that matters is not the share price but what quietly disappears from filings: sidelined teams or dropped safety work.

Insulated from shareholders, not from cash

Anthropic went public with a structure built to keep Wall Street from steering it. Shareholders cannot vote out executives. Control sits with the founders and a body called the Long-Term Benefit Trust. On paper, that is exactly what a safety-first lab would want: freedom to slow down when a model looks dangerous, without a proxy fight. In practice, the Decoder conversation between Nilay Patel and Hayden Field makes a sharper point. Insulation from shareholder votes does not mean insulation from money, and Anthropic needs a staggering amount of it.

Field cites compute bills of roughly $900 billion over the coming years. Amodei himself told JPMorgan Chase’s CEO in a livestream that compute is the company’s deepest business dilemma: buy too much and you overextend, buy too little and you cannot serve customers or train the next model. That is the real governor on Anthropic now. Not a board, not a trust, but the need to keep revenue growing fast enough to pay for chips. Our earlier look at the S-1 found the same pressure in plain numbers, with about $70.7 billion needed for compute in the next 12 months, as we noted when the founders cashed out at the IPO.

A trust with a short leash

Anthropic often presents the trust as the adult supervision above its founders. Field’s reporting suggests the leash runs the other way. The trust currently has four members: former Federal Reserve chair Ben Bernanke, foreign policy analyst Richard Fontaine, experimental physicist Christopher Stubbs and former GiveWell executive Neil Buddy Shah, who chairs it. None brings deep AI experience. A former employee told Field the group was effectively picked not to rock the boat.

Its powers are narrow. It cannot block a model release or weigh in on day-to-day calls, something even OpenAI’s old nonprofit board could attempt. It can appoint or remove certain directors, and only after 30 days’ notice to a small board committee chaired by co-founder Daniela Amodei. Of seven board seats, Field describes three picked by the trust, three by the founders, and a tie-breaker the trust names only with founder sign-off. Add the single super-voting share held by Founder LLC, and the founders’ control looks close to unchecked. We laid out the formal mechanics in who’s really in charge at Anthropic after the IPO; the Decoder discussion is a useful reminder that formal checks and practical checks are not the same thing.

Why this cuts both ways

There is a reasonable defense here. If the founders genuinely believe in pacing development, as Amodei promised on listing day, founder control is what lets them keep that promise when quarterly numbers disappoint. A traditional board answerable to activist investors would likely push harder for speed, not less. The Nvidia saga, where a conventional board forced out Jensen Huang, shows what ordinary governance can do, but nobody should assume ordinary governance would make a frontier lab more careful.

The weakness is that everything now rests on seven people staying aligned for a decade, in an industry with no external regulator of safety decisions. Field’s sources, including former staff, make the obvious objection: if you believe you have the right plan, prove it to someone else, because nobody elected you. University of Colorado law professor Ann Lipton put the choice bluntly in Field’s reporting: you either trust the founders and the trust to balance society’s interests, or you don’t. That is not a governance system. It is a bet on character, made at the exact moment the business pressure to cut corners is rising.

The squeeze on margins

The commercial picture makes that pressure concrete. At Microsoft Ignite, Satya Nadella made clear to Patel and others that Microsoft wants AI vendors fungible underneath Copilot, routing each task to the cheapest capable model. Field describes small and mid-sized firms, including law firms, building private benchmarks to send only a slice of work to expensive Anthropic models and the rest to open-source or mid-tier options. Anthropic’s answer is that companies will always want the frontier. Patel calls that somewhere between naive and idealistic, and the routing behavior Field describes suggests buyers agree with him.

That leaves safety branding doing double duty: it is both a mission and the core enterprise sales pitch. When a brand is also your moat, the temptation is to keep the label while trimming the expensive work behind it. That is the risk worth naming now, before the first earnings call.

Who carries the risk

Public shareholders
They own economics with almost no governance voice; their only lever is selling.
Enterprise customers
They are buying a safety reputation and should ask for evidence that safety teams and evaluations survive cost pressure.
Safety researchers
Their budgets now compete directly with compute in a company that must show revenue growth each quarter.
Regulators
With internal checks this thin, the case for outside oversight of frontier labs gets stronger, not weaker.

What to watch

  1. Anthropic’s first quarterly report as a public company, and whether growth keeps pace with compute commitments.
  2. Any filing language that sidelines a team or discontinues a safety initiative.
  3. Whether the Long-Term Benefit Trust adds members with real AI expertise or uses its removal power at all.
  4. How Nvidia’s leadership change affects chip supply and pricing for Anthropic, given Nvidia’s planned $10 billion share purchase.
  5. Whether OpenAI cites Anthropic’s market reception as it prepares its own listing.

FAQ

Can Anthropic’s shareholders remove its leadership?

Not in any practical sense, according to The Verge’s reading of the S-1. Founder LLC’s single share carries 50.1% of the voting power, and the Long-Term Benefit Trust, not ordinary shareholders, holds the limited power to appoint or remove some directors.

What can the Long-Term Benefit Trust actually do?

It can appoint or remove certain board members after giving 30 days’ notice to a committee chaired by Daniela Amodei. It has no authority over everyday decisions such as whether to release a model.

Why does compute matter so much to the safety question?

Anthropic faces compute bills The Verge puts at roughly $900 billion over the coming years. Paying for that requires fast revenue growth, and that pressure is the likeliest force to squeeze costly safety work.

Read the original

Continue at the source.

The Verge

Companies: OpenAI, Anthropic

More on OpenAI

All OpenAI coverage →