Broadcom’s AI chip revenue triples to $16.7B — custom silicon now undercuts Nvidia

The custom chip business that Hock Tan spent years convincing hyperscalers to trust just tripled in a single year. Broadcom’s AI semiconductor revenue reached $16.7 billion in its fiscal third quarter, a 221 percent increase from a year earlier, and the company is now guiding toward $58 billion in AI chip sales for the full fiscal year. These are not Nvidia-style general-purpose GPUs; they are bespoke silicon built to specification for the specific training and inference workloads of Google, OpenAI, Anthropic, and Meta.

The case for custom silicon is straightforward when you are running compute at hyperscaler scale. A general-purpose GPU handles every AI workload adequately; a chip built around one company’s model architecture handles that workload more efficiently, at lower energy cost per computation. Tan said on the earnings call that the custom Jalapeno chip developed with OpenAI performs ahead of Nvidia‘s Grace Blackwell for OpenAI’s specific workloads — at roughly half the cost per unit of compute. That is a CEO claim, not an independent benchmark, but OpenAI’s commitment to deploy 1.3 gigawatts of Jalapeno next year gives the number structural weight.

Anthropic has signed on with even larger ambitions. Broadcom expects Anthropic to become its largest XPU customer by 2027, with one gigawatt of Ironwood compute already deployed this year and five further gigawatts of Google’s TPU v8i chips lined up for 2027. When Tan outlines a roadmap reaching $230 billion in AI chip revenue by fiscal 2028, the market is increasingly willing to believe him — because three of the world’s five most valuable companies have signed the contracts that would make it real.

The skepticism sits in the guidance. Broadcom projected fourth-quarter revenue of $34.8 billion, a 93 percent increase from a year ago and impressive by any standard except the one Wall Street had set: analysts expected $35.03 billion. The miss is modest in absolute terms but meaningful as a signal. Custom silicon is supply-constrained: Broadcom cannot ramp faster than its semiconductor manufacturing partners, and the gap between the demand Tan is describing and the capacity he can actually deliver is the risk embedded in every line of his roadmap.

There is a structural tension the quarterly figures do not resolve. The better Broadcom gets at building custom silicon for hyperscalers, the more it trains those hyperscalers to think of chip design as a core competency rather than a procurement decision. Google already designs its own Tensor Processing Units; Amazon has Trainium and Inferentia. The question is whether the customers writing Broadcom’s largest checks today eventually bring that capability in-house — which would make Broadcom’s best clients the eventual architects of its displacement.

Total revenue for the quarter reached $29.6 billion, up 86 percent from a year ago. The semiconductor division grew more slowly than the AI segment, reflecting the two-speed reality inside the company: the AI chip business that Tan’s investor narrative has compressed around, and the infrastructure software acquired with VMware in 2023 that quietly supports most of the margin. The infrastructure unit is not an afterthought — it is what allows Broadcom to fund the custom silicon capacity that its hyperscaler clients keep outpacing.

Broadcom’s fourth-quarter results are due in December. If the Anthropic five-gigawatt TPU deployment begins on schedule in 2027, the $230 billion revenue roadmap stops being a projection — and becomes the architecture of the AI decade’s most consequential infrastructure bet.

Tags: nvidia, AI infrastructure, AI chips, Broadcom, custom silicon, Hock Tan