The Take — The market will fund the safety Anthropic can sell

Share
The Take — The market will fund the safety Anthropic can sell

The Verge spent an hour this week asking whether Anthropic can still afford to put safety first now that it is a public company. I think that gets the mechanism backwards. Public markets will not pressure Anthropic to abandon safety in general. They will fund, generously, the part of safety a customer will pay for — and quietly starve the part nobody can put in a contract. The company's own governance guarantees that no one has the standing to rebalance the two.

That is a narrower claim than "the IPO killed the mission," and a worse one for Anthropic, because it does not require anyone to be a villain. It requires only that the two halves of safety have different revenue lines, which they do.

A brake pedal wired to nothing

The detail that matters from Nilay Patel's interview with Hayden Field is not the pledge. It is the machinery. Anthropic's Long-Term Benefit Trust — the four-person body the company markets as the thing standing between humanity and its founders — is made up of Ben Bernanke, Richard Fontaine, Christopher Stubbs and Neil Buddy Shah, per Anthropic's own description. Field noted on the Decoder episode that none of them has much AI experience, and a former employee told her the founders picked them.

What the trust can do is appoint or remove some board members. What it cannot do, per Field's reading of the S-1 and her reporting on the structure, is intervene in day-to-day decisions. It has no power to look at a model and say "do not ship that" — the one power the OpenAI nonprofit board famously used, and the only one that would ever matter. Even removal runs through a 30-day notice to a board subcommittee chaired by Daniela Amodei, a co-founder, who can waive the notice.

Then there is the share. Seven co-founders hold a single share of stock through Founder LLC carrying 50.1 percent of voting power. Field's summary of what governance experts told her is blunt: shareholders can never vote the executives out. So the public is buying the economics of a public benefit corporation without the control — a gap we argued in October is a problem of standing, not of size, in a share you can sell beats a fund you can't sue. The tape makes the split concrete: priced at $83, opened at $111.50, closed at $97.17 on the first day, as our deep dive on the debut laid out.

I do not think that structure is a scandal. I think it is a disclosure, and it says something specific: there is no actor in Anthropic's governance whose job is to protect the safety work that does not sell.

Which half of safety pays for itself

Here is the uncomfortable part. Field's most useful observation is that safety branding is the enterprise pitch — "the only thing they've really got going for them," she says, because large customers demand data sovereignty and contractual restraint that Anthropic cannot monetise any other way. That makes a big slice of safety spending a revenue-defense line. Refusing to train on customer data, publishing model cards, sitting through the audits procurement asks for: buyers pay for that, so the market will protect it. Being public probably makes that half safer, because it is now a number in an earnings deck.

The other half is interpretability research, pre-deployment evaluations, red-teaming, and the decision to hold a model back. No procurement officer writes a purchase order for "the release we did not do." There is no revenue line attached, and in a compute-constrained race every dollar there is a dollar not spent on what Dario Amodei called the company's "deepest business dilemma" in a livestream with JPMorgan Chase's chief executive, per Business Insider.

That race now runs on a quarterly clock. Anthropic has roughly $900 billion of compute committed over the coming years, according to the Verge interview, and most of those obligations are contracted whether or not demand arrives, as our analysis of the commitment stack showed. Steve Rattner's estimate, which Field cites, is that revenue has to rise about 17-fold in a decade to deliver a peer-comparable return. Those numbers are why the ownership structure matters less than it looks: nobody needs to vote to cut an invisible budget. The board just has to approve a plan.

Field's small-business detail is the tell. She says SMBs — law firms especially — are building internal benchmarks to route about 10 percent of tasks to the expensive Anthropic model, 40 percent to open source and 50 percent to something mid-priced. Price discipline is arriving from below, and it lands on the sellable half, which is where the premium lives.

The strongest case against me

The counter-argument is not that the trust works. It is that founder control is what protects the mission. Field's sources split on this, and one told her the founders are not chasing money so much as unchecked control — because they believe they hold the only right plan. Accept that premise and seven people holding 50.1 percent looks like a firewall against the activist investor who wants safety spending cut, not a licence to cut it. Add the reputational stake: a lab whose differentiation is trust cannot survive being caught trimming it, and a public market detects that fast.

There is also a version where public capital helps. Safety work is expensive and pre-revenue, and a $1.3 trillion currency funds it better than a Series G. Anthropic's line that "you can't do safety from second place," as Field quotes the company, is coherent if you believe the race is real.

Why the take still holds

Because "the founders mean it" is a bet on seven people continuing to agree for a decade, and that is the one variable the structure cannot hedge. Succession is the obvious hole: a trust that can only remove board members, was appointed by the people it oversees, and routes its removal power through a founder-chaired committee is a brake pedal wired to nothing. If one founder leaves, dies, sells, or simply changes their mind about a release, no mechanism converts the mission into an obligation.

And the two-halves problem does not need anyone to change their mind. It happens by default, through budgeting. The sellable safety has customers, metrics and a sales team. The unsellable safety has a headcount line in the same quarterly report as the compute bill. When investors ask why gross margin is what it is, the cheapest cut is the one nobody outside can see — and nothing in the S-1 prevents it.

Dustin Moskovitz's objection, which our deep dive noted, was that a public market rewards precisely the racing he funded the company to prevent. I would sharpen it. The market does not reward racing. It rewards legible spending and punishes the rest.

What would change my mind

Four things, none of them expensive. Safety spend disclosed as a recurring line item, with a floor the board cannot cut without saying so. A published statement of what the trust can actually veto over a model release, plus one documented case of it being used. A successor process for trust seats that does not route through the founders. And one contract — redacted is fine — where a customer paid more for pre-deployment safety work rather than for data handling.

If Anthropic's safety premium is real, at least one buyer is paying for the half that carries no price tag. Until then, the honest description of its post-IPO posture is: well funded where it is visible, unowned where it is not. That is not a scandal either. It is how a budget works when the only people who can stop it are the people spending it.

Should a public lab have to disclose what it spends on safety it cannot sell? Write to us via our contact page.

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

The Take — You can't ban your way to an open-weight ecosystem

The Take — You can't ban your way to an open-weight ecosystem

The House report on Chinese AI models gets the diagnosis exactly right and the remedy backwards. It found that American companies buy Chinese open weights because nothing American competes on price, customization and near-frontier performance — then recommended excluding the Chinese models without building the alternative first. That is not a supply policy. It is a price hike with a national-security preamble. The facts are narrow and worth stating plainly.

The Take — Chinese data isn't the risk; the last fine-tune is

The Take — Chinese data isn't the risk; the last fine-tune is

Nous Research took Alibaba's Qwen3.8-27B, put about fifty million tokens of supervision through it, and cut the share of answers carrying Chinese state propaganda from 22 percent to 0.9 percent without losing measurable capability. Most coverage read that as a story about Chinese censorship. I think it is a story about where a model's politics actually live — and it is not where the people writing export rules think it is. As our brief on the release laid out, Arete-27B is supervised…