Broadcom on Monday detailed progress across its custom AI XPU roadmap, saying deployments with five major customers are advancing and that a sixth — OpenAI — is set to begin volume XPU use in 2027. Management framed the programs as deep, multi‑year partnerships tied to large gigawatt‑scale capacity plans and said Broadcom has secured the supply chain to support the ramps through 2028.
Five customers already in production
During the company's fiscal first quarter earnings call on March 4, CEO Hock Tan and Semiconductor president Charlie Kawwas said Broadcom's custom accelerator ramps are "progressing very well" across five core partners. Google remains a steady growth driver, the company said, benefiting from continued demand for its seventh‑generation Ironwood TPU and is expected to require more capacity from later TPU generations.
Broadcom highlighted one partner, Anthropic, as already in production with about 1GW of TPU compute in 2026 and expecting to exceed 3GW in 2027. Management also said Meta's custom accelerator roadmap is active — contradicting some market reports — with Broadcom shipping today and planning to scale Meta's next‑generation XPUs to "multiple GW" in 2027 and beyond.
Customers four and five — not named in the call — were described as seeing strong shipments this year with volumes expected to more than double in 2027, underscoring Broadcom's view of broad-based multi‑customer demand for custom silicon.
OpenAI added as sixth customer, first volume deployments in 2027
Perhaps the most notable disclosure was the company's confirmation of a sixth customer: OpenAI. Broadcom said it expects OpenAI to deploy its first‑generation XPU in volume in 2027 at "over 1GW" of compute capacity. That addition expands Broadcom's stated roster of large custom‑XPU customers and supports the firm's projection of chip‑only AI revenue "significantly in excess of US$100 billion" in 2027.
Custom XPUs for inference and training; multi‑year cadence
Management emphasized that customers are adopting XPUs for both inference and training. Broadcom said many partners begin with inference deployments but increasingly use XPUs across both functions, and some customers are moving toward an engineering cadence that produces multiple chips per year — for example, separate designs specialized for training and for inference.
The company repeatedly framed these engagements as strategic and multi-generational: customers share multi‑year roadmaps with Broadcom, and design choices are being driven by workload specialization.
Supply chain and production assurances
To support the planned gigawatt‑scale deployments, Broadcom said it has "fully secured capacity" for critical components — leading‑edge wafers, HBM and substrates — for 2026 through 2028. Management stressed execution and high‑volume manufacturability as competitive advantages, arguing that the complexity of design plus the ability to ramp hundreds of thousands of parts at acceptable yields will keep Broadcom ahead of nascent customer‑owned chip efforts.
Article edited by Joseph Chen
