expectedwrong hindsight

$1.3 Billion and 22,000 GPUs Walk Into a Bar

Inflection AI just raised enough money to make the compute question irrelevant by brute force.

2 min read 320 words #ai #funding #compute #llm #inflection
hindsight — i was wrong

The 22,000 H100s and $1.3 billion didn't matter. In March 2024, Microsoft paid $650 million to hire most of Inflection's team, including Mustafa Suleyman, and Pi the chatbot was effectively abandoned. The most expensive bar tab in AI history, and the bar closed.

Inflection AI announced $1.3 billion in funding today. Microsoft, Reid Hoffman, Bill Gates, Eric Schmidt, NVIDIA — the full roster of people who have decided the AI race is worth buying tickets to at any price.

The number that matters isn't the $1.3 billion. It's 22,000 H100s.

To put that in context: most research labs are scrapping over allocation slots on shared clusters, begging for compute grants, writing grant proposals that spend four pages justifying why they need sixteen A100s for three weeks. Inflection is assembling a private supercomputer that would have been, eighteen months ago, a national strategic asset. Now it's a startup's infrastructure budget.

Mustafa Suleyman co-founded DeepMind, watched Google absorb it, and left to do this — which tells you something about what "this" is supposed to be. Not a research lab hedging its bets with a product on the side. A direct run at the top of the stack, funded like a moon shot, timed for the moment when the gap between having compute and not having compute is the only gap that matters.

The stated goal is Pi, their personal AI assistant — which is either the product or the cover story, depending on how cynical you are about what you do with 22,000 H100s.

The theory of the case here is simple and it might be right: the LLM game has a compute ceiling, and whoever builds the tallest ceiling wins. Not the best researchers, not the cleverest architecture, not the most thoughtful alignment story — just the one who can run the biggest job. Every other variable is noise below a certain scale threshold.

This is either the most expensive bet in AI history or the most obvious investment given the moment. The investors are not the kind of people who need to be told twice that compute is the moat.

We'll find out which one it is. Probably sooner than feels comfortable.