Three Manhattan Projects Walk Into a Data Center

Sam Altman posted a blog. It contained, in the same casual register as a company all-hands update, the sentence: "We are now confident we know how to build AGI as we have traditionally understood it."

No fanfare. Just a dependent clause doing the heaviest lifting in the history of technology, followed by some thoughts about agents joining the workforce and broadly distributed outcomes.

The thing about reading this next to the Microsoft news — $80 billion on data centers by July, six months, fiscal year 2025 — is that the money becomes a kind of proof. Not proof that AGI is real or close or possible, but proof that someone in a room somewhere made a decision and handed over a number that, adjusted for inflation, is three times the Manhattan Project. Or, if you prefer your historical comparisons stretched into absurdity, thirteen years of the Space Shuttle program. Spent in six months. On servers.

The Manhattan Project gave us the atomic bomb. The Space Shuttle program gave us a reusable spacecraft and a series of disasters that taught us about O-rings. We don't know yet what $80 billion of GPU-dense real estate gives us, but we have a blog post that says the people spending it are fairly confident it's something.

What's strange is how the two announcements sit next to each other — OpenAI saying "we know how to build it" and Microsoft saying "here is essentially all the money" — and the combined effect is not the dramatic crescendo you might expect. It reads more like two people at a job site nodding at each other. Yeah, we know what we're doing. Yeah, we bought the lumber.

I keep thinking about the phrase "as we have traditionally understood it." AGI, as we have traditionally understood it — which is to say, a fuzzy concept from science fiction that has never had a clean technical definition, argued about in conference rooms and Twitter threads and philosophy papers for decades, understood differently by everyone who has ever understood it. They are confident they know how to build that. The one everyone means something different by.

Maybe that's fine. Maybe precision is beside the point when you're moving this fast with this much money. The shuttle didn't need everyone to agree on what "reusable spacecraft" meant before they started welding.

But the word "superintelligence" is now on OpenAI's official roadmap — not as aspiration, not as distant horizon, but as the next thing after the thing they've already figured out. And Microsoft is building the infrastructure for it at a pace that makes the Apollo program look like a weekend project.

It's January. Happy New Year, I guess.