What was announced

On January 21, the day after the inauguration, the President stood at a White House press conference with Sam Altman, Masayoshi Son and Larry Ellison and announced Stargate. The stated plan is 100 billion dollars deployed immediately and 500 billion dollars through 2029, all of it for AI infrastructure for OpenAI. SoftBank and OpenAI each hold 40 percent of the venture. Oracle and the Abu Dhabi fund MGX are the other partners. Son chairs it and carries financial responsibility, and Altman has operational responsibility.

The first site is in Abilene, Texas, where ten data centres are reported to be under construction. Reporting on the initial capital puts SoftBank and OpenAI at about 19 billion dollars each and Oracle and MGX at about 7 billion each, which is well short of 100 billion and very far short of 500. Microsoft, OpenAI's existing cloud provider, kept a right of first refusal on OpenAI capacity. Elon Musk said publicly within a day that the partners did not have the money, and Altman said publicly that they did.

The number and the counterfactual

We do not know whether 500 billion dollars will be spent, and we suspect the people at the press conference do not know either. The question we can reason about is what the commitment means whether or not the full figure lands. Even a fraction of it is a dedicated compute pool for one company, of a size that no university consortium, national lab programme or non-profit can approach. The buildout is private, the capacity is reserved, and the announcement language was about a single customer's models.

The counterfactual matters here. Before this week, compute at frontier scale was already concentrated in a handful of labs and their cloud partners, and the rest of us worked in the gap between what those labs published and what a few hundred donated or rented GPUs could reproduce. Stargate does not create that gap. It widens it, and it makes the widening official policy by putting the announcement in the White House rather than in a press release.

What research can still do on the outside

It is worth being concrete about what kinds of work need frontier-scale compute and what kinds do not. Training a frontier model needs it, and that door was already shut to academic groups. Nearly everything else does not. Evaluation, interpretability on open weights, data curation, post-training recipes, small-scale scaling studies and reproductions all run on hardware that a well-funded lab or a university cluster already has. Most of the results we have found useful this past year came from groups that never trained a model above a few billion parameters.

The catch is access to the artefacts. Work of that kind depends on the frontier labs releasing weights, checkpoints and enough detail to study, or on open-weight groups keeping pace closely enough that findings on their models transfer. If a 500 billion dollar buildout pulls the frontier further from the best open model, then results on open models become less informative about the systems people actually use. The larger risk to outside research is that the objects worth studying move out of reach, whatever hardware we have.

Rented, donated and public

There are three ways a non-profit gets compute today. Rent it from a cloud provider at market price, accept donated credits from a lab or a cloud provider with whatever strings come attached, or apply for time on a public system. All three are vulnerable in a market where a single customer has reserved the equivalent of several gigawatts. Rental prices follow scarcity. Donated credits follow the donor's priorities. Public systems are funded on timescales that make a five-year 500 billion dollar plan look nimble.

We do not think the answer is for foundations to try to compete on hardware. The answer, if there is one, is for the research that happens outside the labs to be the research that makes the inside legible. Evaluation methods that labs adopt, interpretability tools that work on their models, open benchmarks that they cannot ignore. That work is cheap by comparison and its value goes up, not down, as the frontier gets more expensive and more closed.

What we will watch for

The first thing is whether the money arrives. The initial commitments reported add up to around 52 billion dollars, so the gap to the first 100 billion is real, and the gap to 500 is a matter of belief. The second is whether any of the capacity is offered outside OpenAI, on any terms, to any researcher. Nothing in the announcement suggests it will be.

The third, and the one that matters most to us, is whether the labs building on this scale keep releasing enough for outsiders to do useful work on their models. If they do, a compute gap is survivable. If they do not, the gap becomes a knowledge gap, and no amount of donated credits closes that.

Sources

  1. Wikipedia: Stargate LLC