Broadcom Q3 FY2026: Hock Tan Claims TPU v8i Matches Vera Rubin as Anthropic Commits 5GW in 2027
Broadcom’s Q3 FY2026 earnings, reported September 2, showed custom AI accelerator revenue scaling faster than expected. CEO Hock Tan used the earnings call to make the direct performance case against Nvidia.
“TPU v8i is comparable if not surpasses the Vera Rubin GPU,” Tan told analysts. He also claimed the Jalapeno chip — designed by OpenAI and co-developed with Broadcom and Celestica, manufactured by TSMC — runs OpenAI workloads at half the cost of an equivalent GPU deployment.
Revenue Pipeline
Broadcom has secured supply commitments that set a floor on near-term AI revenue:
| Period | AI Revenue Secured |
|---|---|
| FY2027 | $115 billion |
| FY2028 | $230 billion |
Those figures represent Broadcom’s AI semiconductor revenue guidance, underpinned by securing upstream supply chain capacity — wafers, HBM, and substrates — from Google (TPU), OpenAI, and other hyperscaler partners.
Anthropic’s 5GW Commitment
Anthropic is expected to deploy five gigawatts of TPU v8i capacity in 2027. That represents a step-change from the gigawatt-scale deals announced in early 2026. At inference-workload power densities, five gigawatts of TPU capacity is substantial — the equivalent of several large data center campuses running Broadcom silicon exclusively.
Nvidia Holds Training; ASICs Take Inference
Nvidia’s Q2 FY2027 earnings, reported within the same week, showed data center revenue at $89 billion, reinforcing GPU dominance in training-heavy workloads.
The split developing across both earnings calls: GPU clusters are still the primary tool for large-scale model training; custom ASICs are taking share in inference, where workload shapes are predictable and cost-per-token matters more than raw flexibility. Broadcom’s Q3 results are the clearest evidence yet that this handoff is underway.
Context on Jalapeno
The Jalapeno chip was announced publicly on June 24, 2026, when OpenAI and Broadcom jointly unveiled the design. Tan’s Q3 claim of “half the cost of a GPU” for OpenAI workloads is the first earnings-call validation of those original efficiency claims. It is a vendor assertion, not an independent benchmark result.
Nvidia’s Vera Rubin GPU, announced earlier this year, is architected for long-context inference with 50 PFLOPS FP4 performance. Broadcom’s comparison positions TPU v8i as a direct replacement for that workload class.
What the Revenue Numbers Actually Mean
The $115B/$230B figures are supply commitments. They reflect how much Broadcom’s hyperscaler customers have contractually agreed to absorb — a floor, not a ceiling. If inference demand accelerates (as training compute did in 2024-2025), both figures could revise upward. If any major customer delays deployment, they revise down. The Anthropic 5GW commitment is one line item inside the $115B FY2027 figure.