AMD's MI450 ramps into OpenAI's 6GW pact, opening a second front on Nvidia

Summary

Amd is ramping its Instinct MI450 into Openai's 6GW multi-generation GPU pact, the first 1GW deploying in 2H 2026, its most serious push yet against Nvidia's data-center dominance. The flagship MI455X packs 432GB of HBM4 and tens of petaflops of low-precision throughput; AMD's double-wide Helios rack (Q3 2026) bundles 72 MI455X accelerators for ~3 AI exaflops, pitched directly at Nvidia's Rubin racks. Meta has also signed multi-gigawatt MI450 deals. The OpenAI agreement carries warrant/equity terms tying AMD's upside to OpenAI, another strand of the AI financing loop. Execution against Nvidia's software moat and supply chain is the open question.

By the numbers

• 6GW, total AMD Instinct GPUs committed to OpenAI (multi-generation).
• 1GW, first MI450 tranche, deploying 2H 2026.
• 72 / ~3, MI455X accelerators per Helios rack / AI exaflops per rack.
• 432GB, HBM4 memory on the MI455X.
• Q3 2026, Helios rack availability.

Why it matters

A credible second supplier reshapes the economics of the buildout: it pressures Nvidia pricing, gives OpenAI and Meta leverage, and spreads frontier compute across two vendors. But the warrant-linked terms knit AMD into the same circular financing that worries skeptics about the whole AI capex cycle.

What to watch

• Whether the first 1GW of MI450 deploys on the 2H-2026 schedule.
• AMD software/ROCm maturity versus CUDA in real training workloads.
• HBM4 supply and how AMD's gains read in Nvidia's data-center share.