Anthropic Compute Constraint

A fintech founder just left Y Combinator to go solve a compute problem. That tells you where the real bottleneck in AI has moved.

Tom Blomfield, co-founder of Monzo and GoCardless and a YC general partner since 2023, is taking leave to join Anthropic's compute team under Chief Compute Officer Tom Brown. He follows a striking 2026 run for Anthropic: Andrej Karpathy in May, Nobel laureate John Jumper from DeepMind, and now an operator whose whole career has been about scaling infrastructure under load.

Notice what these hires have in common. The marquee names aren't going into model design. They're going into compute.

For the past two years the enterprise question was "which model do we standardize on." That question is quietly becoming secondary. When frontier labs are pouring their best talent into securing and orchestrating compute, they're telling you that capacity, not capability, is the binding constraint on what AI can actually deliver at scale.

That reframes the buying decision. The vendor with the best benchmark score isn't automatically the one that can serve your workload reliably, at a predictable price, a year from now. Compute access, contract terms, and capacity guarantees deserve the same scrutiny you give model quality.

If your AI roadmap assumes models are the scarce resource, you may be optimizing the wrong variable.

When you evaluate an AI vendor, how much weight do you put on compute capacity versus model performance?