Big market bets on inference chips
Market bets on inference silicon are massive — OpenAI, Meta and others are steering multibillion-dollar commitments toward non‑NVIDIA chips, per industry commentary argued. The social thread lists tens‑of‑billions commitments across AMD, Cerebras, Broadcom and hyperscaler Trainium plays, signaling wide buyer diversification argued.
OpenAI agreed to buy up to 750 megawatts of Cerebras wafer‑scale inference systems in a multiyear pact reported to exceed $10 billion, with capacity scheduled to roll out in tranches through 2028. (openai.com) OpenAI and Broadcom announced a co‑development and deployment plan for 10 gigawatts of custom AI accelerators, with Broadcom set to start building racks in the second half of 2026 and complete rollout by the end of 2029. (openai.com) OpenAI’s multi‑year supply agreement with AMD commits the company to deploy up to 6 gigawatts of AMD Instinct GPUs — beginning with a 1‑gigawatt MI450 deployment in the second half of 2026. (openai.com) AWS’s Project Rainier brought nearly 500,000 Trainium2 chips online as of October 29, 2025, and Anthropic is projected to scale Claude to more than one million Trainium2 chips by the end of 2025, underpinning Trainium’s rapid growth into a multibillion‑dollar product line. (aboutamazon.com) Meta revealed four generations of MTIA inference chips with the MTIA 300 already deployed and MTIA 400–500 planned for 2026–2027, while Meta’s separate AMD purchase deal includes performance‑based warrants — up to 160 million shares tied to a 1–6 gigawatt purchase cadence. (cnbc.com) Those announcements came on a tight timeline — OpenAI’s AMD and Broadcom pacts were disclosed in October 2025, the Cerebras agreement was announced on January 14, 2026, and Meta’s MTIA roadmap was updated in March 2026 — even as Nvidia earlier committed to support OpenAI with roughly 10 gigawatts of systems under a separate $100 billion‑scale framework. (openai.com)
