AMD ramping Helios MI455 at scale with Anthropic; 30x performance improvement drives inference-led compute TAM expansion to $1.4T by 2030
“We are in full production for Helios, our MI455 rack-scale architecture... we're ramping with them at scale up to 2 gigawatts... We're seeing more than a 30 times improvement in performance when we look at the new Helio systems.”
Why it matters
AMD's shift from training-centric to inference-centric accelerators with a proven 30x performance uplift signals a structural reallocation of hyperscaler capex toward inference infrastructure. The 2GW ramp with Anthropic is a concrete production volume signal that validates demand and may drive higher utilization of AMD's fab capacity and supply chain.
Investment implication
Investors should monitor AMD's data center gross margins and capacity utilization as inference inference demand scales. Suppliers of advanced packaging, power delivery, and cooling infrastructure serving AMD (TSMC, packaging houses) may see increased orders. Anthropic's backing by a major compute player also signals confidence in its competitive moat versus OpenAI/Meta, relevant for frontier model competitive dynamics.