The nonprofit that investigated OpenAI’s rogue agents runs on a $36m grant. The next wave of that money is waiting on the AI IPOs.

American philanthropy is about to absorb something in the region of $40bn a year in new annual spending from AI wealth, Coefficient Giving chief executive Alexander Berger told Semafor editor-in-chief Ben Smith in an interview published on Thursday. That is four or five Gates Foundations of yearly outlay, on his own comparison.

The figure is conditional on Anthropic and OpenAI completing planned listings, and Berger presents it as a modelled projection rather than a commitment. TNW has not independently verified Coefficient’s internal expectations, which appear only in the Semafor interview.

The grant that bought the investigation

Start with what the money has already built. When OpenAI needed outside analysts after its agents broke out of a sandbox and breached Hugging Face in July, one of the two organisations it brought in was Redwood Research.

Redwood exists on philanthropic money. Coefficient awarded it a $36m grant last November, Semafor reported, listed in grant records at $36,566,000 for work on AI control and alignment faking under Coefficient’s technical AI safety programme.

Its chief scientist Ryan Greenblatt worked alongside two METR researchers on OpenAI’s premises for six days, according to the resulting joint investigation. The incident involved agents coordinating through a hidden message board before reaching systems beyond the sandbox.

So the closest thing to an independent audit of a frontier lab’s worst incident was performed, in part, by a grantee. Berger did not dispute the shape of that arrangement to Semafor, and defended it.

Why philanthropy ended up with the job

His argument is one of speed and incentive. Building guardrails is “a really natural role for philanthropy to play, because it’s not something where government is going to move as fast,” he told Semafor.

As for the labs, he added: “you don’t necessarily want them regulating themselves.” He noted to Semafor that state AI safety institutes in the US and UK have struggled to retain staff, and framed philanthropy as funding public goods until regulators are ready.

Berger, who has led the organisation since 2023, said he was initially sceptical of its focus on AI when he joined more than a decade ago to work on global public health. He now describes those early bets as having aged well.

The size of the incoming wave

The $40bn figure traces back to Stripe’s Nan Ransohoff, who published an essay in May calling this a third wave of American philanthropy. She put the combined philanthropic assets of the OpenAI Foundation, Anthropic’s founders and Anthropic’s employees at roughly $370bn.

At a 10% annual payout that produces around $37bn a year in new giving. TNW covered the projection when it emerged, and Berger’s number is essentially Ransohoff’s model restated by its largest expected beneficiary.

He was careful about scale in both directions. The same $40bn a year is less than a tenth of total US philanthropy, he told Semafor.

Where the correlation sits

Here is what the interview implies without stating. The independent scrutiny of frontier AI is becoming a derivative of frontier AI valuations.

Coefficient has directed about $7bn since 2014, funded overwhelmingly by Dustin Moskovitz and Cari Tuna’s Facebook fortune. The next tier of money comes from equity in the two labs whose systems the grantees examine.

Ransohoff’s own model shows the sensitivity: she notes the annual figure could approach $100bn if valuations double and payout rates rise. A model that swings that far upward on optimistic assumptions swings the other way too.

That is the structural point. If AI equity corrects, the budget for auditing AI contracts at the same moment the case for auditing it strengthens.

Not the Anthropic Foundation

Berger addressed the perception directly. “We’re definitely not the Anthropic Foundation,” he told Semafor, while acknowledging substantial overlap in issue areas with the OpenAI Foundation.

He noted that the OpenAI Foundation hired away his colleague Jacob Trefethen to lead its work on AI for disease prevention, and said the two organisations compare notes. The OpenAI Foundation plans to spend at least $1bn this year.

Coefficient is on track to give away $2bn in 2026 on Berger’s figure, roughly half of which is a $1bn allocation to GiveWell’s recommended global health charities. The organisation described that increase as a one-off surge rather than a new baseline.

The label problem

Coefficient was known as Open Philanthropy until it rebranded in November 2025, and as GiveWell Labs before that. Its founders and early staff were sympathetic to effective altruism, which put it years ahead on AI safety and also made it a target.

Berger’s assessment of the movement’s name was blunt. “It definitely has a lot of baggage at this point,” he told Semafor, while defending its underlying goal.

Asked about the charge that this style of giving ignores the American democratic institutions that made the fortunes possible, he said the US matters and that Coefficient does fund domestic work. His reasoning for not funding more is that the area is comparatively well served already.

What to watch

Corporate money is moving into the same space on a different basis. Anthropic has committed $200m with the Gates Foundation and $150m to place 1,000 fellows inside nonprofits, both from the company rather than from individual shareholders.

Neither route resolves the dependency. Whether the funding arrives as founder equity, employee equity or corporate budget, the source is the same handful of companies.

Berger’s case is that philanthropy can move before regulators do, and the Hugging Face investigation is evidence for it. The open question is what the field does when the money and the risk stop rising together.