AMD

On 15 October 2025, The Batch issue 323 reported that OpenAI was tightening its ties with AMD, and read the move as a step away from the NVIDIA ecosystem. The stated purpose was hardware suited to OpenAI's training and inference workloads; the newsletter also expected effects on GPU availability and on competition in the AI hardware market.

Six days earlier, an analysis of Bloomberg's reporting on a "web of circular deals" placed AMD inside a different frame: a network in which a small number of companies invest billions in one another, potentially inflating valuations without durable monetization. OpenAI is not projected to be cash-flow positive before the end of the decade. A supply arrangement between two members of that network is hard to read as a plain commercial fact.

Cobus Greyling, writing on 30 October about the DGX Spark workstation, gives AMD its sharpest role, and it is a counterfactual one. "No easy AMD/Intel swaps, but that's the point right?" In that reading AMD is the substitution NVIDIA's stack (open source Nemotron models, developer tooling, fine-tuning methodology, all optimized for NVIDIA hardware) is designed to make impractical.

Both readings hold at once. AMD is the named alternative that gives OpenAI leverage in negotiations, and the alternative whose absence from NVIDIA's software ecosystem measures how deep that ecosystem's lock-in runs. Which of the two the partnership turns out to be depends on whether the circular capital holds.