AMD/gpu_share_capture/OpenAI MI450 Deal — AMD's Most Important Catalyst

OpenAI MI450 Deal — AMD's Most Important Catalyst

6 GWOpenAI Deal SizeOct 2025. Potential $90-100B+ revenue. First 1 GW binding

The AMD-OpenAI partnership is the single most important catalyst in AMD's investment thesis. The world's leading AI company chose AMD for its next-generation infrastructure, validating AMD as a credible #2 GPU vendor at the highest level. The MI450 is co-engineered with OpenAI input, making this a deep technical partnership, not just a procurement deal. Meta signed an identical structure one month later, confirming the pattern.

1 GW
Binding commitment
H2 2026. Remaining 5 GW milestone-dependent
160M shares
Warrant dilution
~10% at $0.01/share. Final tranche at $600
6 GW
Meta deal
Feb 2026. Identical structure. Through 2031
320M shares
Combined dilution
~20% if fully vested

The critical nuance: OpenAI simultaneously signed a 10 GW deal with Broadcom for custom 'Titan' ASICs. AMD is one supplier in OpenAI's multi-vendor strategy (NVIDIA + AMD + custom ASIC), not the sole NVIDIA replacement. And only the first 1 GW is binding — the remaining 5 GW depends on MI450 execution meeting OpenAI's performance benchmarks.

The binding vs optional question

Only the first 1 GW per deal is binding. If MI450/Helios benchmarks disappoint or Vera Rubin outperforms, OpenAI and Meta can shift remaining budget to NVIDIA or custom ASICs without penalty. The deal is simultaneously the strongest bull evidence and the biggest execution risk.

The key question

Will OpenAI exercise beyond the binding 1 GW commitment? What performance benchmarks must MI450 hit?

Open questions

?Does OpenAI's 10 GW Broadcom ASIC deal serve different workloads (inference vs training) or directly compete with the AMD deal?
?At what deployment milestones do the warrant tranches vest? What is the probability-weighted dilution scenario?
?Can AMD deliver 1 GW of MI450/Helios racks by end of 2026 given TSMC 2nm and HBM4 supply constraints?
?Will the Meta deal follow the same conversion trajectory as OpenAI, or is one more likely to fully convert than the other?