Two billion for a signature, twelve without

Posted on: 31 August 2026

The cleanest way into this story is to put two numbers next to each other and leave them there. In late 2024 a letter headed Project Prometheus offered Anima Anandkumar a place inside a company that Vik Bajaj was about to found with Jeff Bezos, backed by more than two billion dollars already committed in tranches through to a Series B. She turned it down. In June of this year Bezos and Bajaj closed a Series B for Prometheus worth twelve billion dollars, without her. Reuters reviewed the letter and the minutes of the meetings and published the account last week, on the occasion of the launch of the company Anandkumar founded instead, Accelerated Understanding, based in Pasadena and run with her husband Benedikt Jenik.

Most of the coverage has taken the shape of a parable about courage, which is understandable and not very interesting, because the arithmetic is doing something more precise than that. The offer was not refused into a vacuum. It was refused, and then the round it was meant to unlock closed at six times the sum without the person it had been written for, which tells us with unusual clarity what the two billion was and was not paying for. It was not paying for engineering capacity, since the capacity to raise twelve billion evidently survived her absence intact. It was paying for the time a project with no history needs in order to look old, and that is a different commodity with a different shelf life, one that runs out at the precise moment the first round closes.

Anyone who has watched an eminent name arrive on a British board will recognise the shape of the thing. The convention here is well developed and mostly honest about itself, in that a distinguished figure is brought in to lend an institution the credibility it has not yet earned by its own conduct, and everyone involved knows this is the trade even when nobody says so at the meeting. What makes the Prometheus letter worth reading closely is that it wrote the trade down. It offered Anandkumar three things, namely the public face of the new company, a seat on the board and ownership of the scientific vision, and those three things are worth weighing separately because their sum is smaller than it looks. A public face is an allocation of attention which can be withdrawn by press release. A seat on the board is one vote out of a number of seats that the letter, so far as we know, did not specify, and in a boardroom the number is the whole argument. Ownership of a scientific vision is a phrase with no existence in any legal system on earth, since it cannot be registered, cannot be carried out of the building and cannot be argued in front of a judge. It is a compliment that has been given the shape of a clause.

What the list does not contain is capital allocation, hiring authority, a veto, and above all the schedule by which those committed tranches release against milestones. In a staged financing the person who writes the milestones writes the company, as anyone will know who has watched a business change direction between one release and the next without a single vote being taken on the matter, and the effect is entirely legal and largely invisible from the outside. I have read enough documents of that kind to recognise the construction, which is that the generous part sits on the first page while the part that governs anything arrives in the schedules, where generosity turns into arithmetic. None of this requires bad faith on anyone's part. It requires only that the people drafting the letter act for the people providing the money, which is always the case, and the drafting will then optimise for what transfers well into a press release and against what transfers into a shareholders' agreement. The structural limit on the whole exercise is that the person capable of reading the asymmetry is exactly the person whose signature was wanted, which is why these offers can only be pushed so far before they defeat themselves.

There is a detail in the chronology that tends to be read the wrong way round. Anandkumar and Jenik had already founded their company before that dinner in the restaurant outside Los Angeles. The letter was therefore proposing to buy a founding that had already happened, which is when such offers almost always arrive and also when they are worth least, because the seller has already paid the cost the buyer is offering to cover. There is a second detail that passes as biography and is really governance. The two are married and founded together, he builds training infrastructure while she builds models, and the letter was addressed to her alone, as public face, as director, as owner of the vision. In a company of two people that is not a courtesy extended to the better known of the pair. It settles which of the two speaks to the money and which of the two executes, and partnerships that come apart tend to come apart along a line somebody drew from the outside with the best of intentions.

The cost of the refusal deserves stating too, otherwise this stays a fable. Two billion committed is certainty, and certainty in this trade is expensive to decline. In its place there are arrangements with compute providers who are not named, which means the constraint has not gone anywhere, it has changed counterparty and become less visible, and in an industry where cluster availability is the binding input a creditor without a name is not obviously the safer one to owe. Anandkumar declined to discuss funding. She said she has agreements that have given her access to the clusters and she did not say with whom. Nvidia did not respond when Reuters asked whether it stood behind the arrangement. Jensen Huang, by her own account, was the person who encouraged her to try it alone.

I have written before about the difference between buying insight and buying legitimation, in the context of strategy consultancy, where the client is usually paying for political cover and being sold analytical superiority. The reason this case is worth returning to is that the legitimation carries a public price here, quoted twice, eighteen months apart, by the same buyer. That does not happen often, and it gives anyone holding a similar letter a test that costs nothing to apply. Count how many of the items offered would survive a serious disagreement with the person providing the money. Titles do not survive it. Vetoes do. Everything in between depends on the number of seats, which is the item that in documents of this kind appears last or does not appear at all.

The claim that Accelerated Understanding has processed five trillion data points in a single prompt is the company's own, from internal tests, and nobody outside has seen it. That number is doing a great deal of work in the coverage and it should be kept separate from everything above, which stands or falls on documents Reuters has read.