Insights - Why AMD and not NVIDIA?

Why AMD and not NVIDIA? I get this question a lot. Today I’m at AMD Advancing AI 2026 in San Francisco, good moment to find answers.

Simple: we don’t train models, we run them.

At stressmaniacs and 4chems.com we run open models on our own hardware, in production. No cloud, no CUDA. Why does this work for us?

• MoE models need memory, not FLOPS. 128 GB unified memory in a machine you can just go and buy. Try getting that elsewhere without a datacenter budget.
• Our customers often need their data in the EU, GDPR-compliant. Open models, open runtime, own hardware — that’s a setup I can explain to any auditor.
• “ROCm is not production-ready” — that was 2023. OpenAI, Meta, Microsoft and Oracle are on stage here, running AMD at scale.
• And I have seen this before. Nobody believed Intel could be beaten. Then came Ryzen and EPYC. Now the same playbook in AI: NPUs, iGPUs, Instinct — one open stack from laptop to rack.

Maybe it’s a Pavlovian reflex from 25 years of stresstesting infrastructure: I don’t like single points of failure. A single vendor for all AI compute is one.

For self-hosted inference of open models in Europe? AMD is a pragmatic choice. Not ideology, the numbers just fit.

Tomorrow Lisa Su keynote: ROCm roadmap, MI450, Helios. Let’s see if my bet holds.