AMD has been a recurring reference point in Signwl's analysis across two distinct threads.
The MI450 / Helios training threat. Per multiple briefs across the week, AMD's MI450 Helios rack (shipping H2 2026) is positioned as a credible competitive threat to Nvidia B200 / GB200 on training workloads. The Meta and OpenAI customer commitments to AMD for 6 GW of training capacity were referenced as the substantive demand signal. Materially undercutting B200 on-demand pricing on launch is identified in our weekend briefs as one of the few scenarios that would invalidate the current B200 demand-pull squeeze (Ohio at $6.05/hr, +136% 30-day) — making MI450 pricing the single most important AMD data point for the GPU complex.
The GPU-to-CPU ratio shift. AMD itself has guided that the GPU-to-CPU ratio for inference workloads is moving from 8:1 toward 1:1 — i.e. a single CPU-per-GPU configuration replaced by configurations that look much more CPU-heavy. If accurate, this structurally increases the addressable market for AMD's EPYC server CPUs (where it has been gaining share against Intel for several years) and decreases per-workload inference GPU demand. Our analysis reads this as a bullish CPU thesis paired with a more nuanced GPU thesis: AMD wins on the CPU side regardless of which GPU vendor captures inference share.
The MI25 catalog exit. A separate signal: the MI25 (AMD's data-centre GPU launched in 2017) shows is_active=false on AWS and has collapsed in price across remaining listings. This is consistent with broader AMD legacy GPU catalog rationalisation, and points to AMD's data-centre GPU strategy concentrating on MI300X / MI450 rather than maintaining a broad fleet.
The composite picture: AMD is structurally winning the CPU narrative downstream of AI inference, building a credible training-rack alternative, and rationalising the legacy GPU catalog to focus on its highest-conviction next-generation products.