Anthropic names compute access, not research, as what keeps Claude at the frontier
In the AMD announcement, Anthropic co-founder Tom Brown framed access to compute as central to keeping Claude at the frontier and meeting customer demand. Stated plainly by a lab founder, that is a claim about where the binding constraint now sits.
Three years ago the stated bottleneck was research talent. Two years ago it was data. The current answer, from the people in a position to know, is megawatts — and unlike the previous two, this one has a delivery schedule attached and cannot be accelerated by hiring.
It also explains the financing structure. If compute is the constraint and compute is capital-intensive with multi-year lead times, then equity-for-hardware arrangements are the rational instrument, whatever they do to the independence of the demand signal.
The first Helios gigawatt is scheduled for the first half of 2027, on MI455X parts with EPYC Venice CPUs and Pensando networking. That date is the checkable part of the announcement, and it is far enough out that several model generations will pass before it arrives.
Pulse 2 — AMD and Anthropic partner on 2-gigawatt AI infrastructure deployment and up to $5 billion investment → · TechTimes — AMD bets $5 billion on Anthropic to fix the one thing holding back its AI chips → · Proactive Investors — AMD, Anthropic announce strategic AI infrastructure partnership with up to $5B investment →