OpenAI's coverage in our recent briefs has been less about a single deal anchor (the way Anthropic's $21bn TPU order has anchored coverage there) and more about the breadth of compute supply chain partnerships.
The multi-vendor compute footprint. OpenAI appears across briefs as customer or partner of:
- Broadcom — for custom silicon (per May 13 brief: "Broadcom has custom AI chip deals with Alphabet, Meta, Anthropic, and OpenAI")
- AMD MI450 / Helios — for training capacity (per May 17 brief: "AMD MI450/Helios with Meta + OpenAI as customers for 6 GW of training capacity")
- Oracle / Crusoe Energy — for compute hosting at Crusoe's Abilene methane-powered DC (per May 13 brief)
- Nvidia — by implication, given OpenAI's underlying frontier model training and inference demand
The diversification signal: OpenAI is hedging across compute vendors more visibly than most frontier model developers. This is consistent with both the substitution-arc thesis (Nvidia is no longer the only supplier) and with prudent supply-chain risk management at the scale OpenAI operates.
The compute-buyer demand-side signal. Per the May 13 brief, OpenAI is named as one of four hyperscaler / AI lab customers (alongside Alphabet, Meta, Anthropic) partnering with Broadcom for XPU custom silicon. The grouping matters: Broadcom's $20bn 2025 AI revenue and projected $100bn+ by end-2026 implies that demand from this four-customer set is the dominant driver of Broadcom's growth curve. OpenAI is one of four credible sources of that demand.
The 6 GW AMD training capacity. This is a substantial commitment. The May 17 brief frames it: "AMD MI450/Helios with Meta + OpenAI as customers for 6 GW of training capacity." Meta + OpenAI together on a single AMD product line at 6 GW is a material counter-balance to Nvidia's training share. If MI450 pricing on launch undercuts B200 (one of the monitor items on the AMD page), this OpenAI-AMD relationship is the most visible test of the substitution thesis.
The DeepSeek V4 competitive context. OpenAI's GPT-5.4 is being commoditised at the per-token pricing level by DeepSeek V4 ($1.74/M tokens, 27% of V3 FLOPs, runs on Huawei Ascend). This compresses the per-token inference economics across all frontier-model developers including OpenAI.
The composite picture: a high-information compute customer with diversified supplier exposure that makes OpenAI one of the cleanest leading indicators for the broader substitution arc. Coverage is breadth-oriented; single-deal-quantification is less central than for Anthropic.