Skip to content
1 min read Newsletters

Quick Hits: OpenAI's Mammoth AMD Deal

The massive chip deal between the two and its implications

Quick Hits: OpenAI's Mammoth AMD Deal
Photo by İsmail Enes Ayhan / Unsplash

OpenAI has signed a multibillion-dollar deal with AMD to purchase up to 6 gigawatts¹ of chips, including the MI450, to power AI inference workloads starting in 2026. The agreement includes warrants for OpenAI to acquire up to 10% of AMD at a nominal cost, contingent on performance milestones. This is AMD’s most direct challenge to Nvidia’s dominance and signals a shift in AI chip procurement strategy.

¹ Grounding it in the known MI450 IF128 specs, that’s a continent-scale build — multiples larger than today’s largest single-tenant AI deployments. For comparison: all hyperscale data center capacity added worldwide in 2024 was roughly on the order of ~25–30 GW (all workloads). AMD/OpenAI are claiming ~20% of an entire year’s global data center power footprint, just for one partnership, and very quickly. If it happened, which seems a stretch, it would require massive changes in power, cooling, real estate, networking, etc.

Here are four quick points:

  1. Vendor Financing in Equity Form

AMD is effectively subsidizing demand by issuing OpenAI warrants for up to 160M shares. That makes OpenAI a 10% shareholder: part customer, part financier — a risk transfer from cash to stock, as well as making OpenAI the largest and thus controlling AMD shareholder.

  1. CUDA Lock-in Challenge

Embedding OpenAI as a large AMD partner forces ROCm (AMD's stack) adoption at scale. This is less about FLOPS and more about dragging AI software development out of CUDA’s orbit, with large implications as workloads shift to inference from training.

  1. Data Center Overbuild Risk Grows