OpenAI's Jalapeño ASIC Redraws the Hardware Map
Yesterday at Hot Chips 2026, OpenAI and Broadcom dropped benchmarks for Jalapeño — OpenAI's first custom inference ASIC. The numbers aren't just competitive. They're a paradigm shift.
A single Jalapeño chip runs at 700W and delivers up to 1.9x the throughput per kilowatt of Nvidia's flagship GB300 (Vera Rubin), while cutting inference latency by 3.6x. That's a 1,400W Nvidia part being outperformed by a chip at half the power envelope. The Register reports that 128 Jalapeño chips in a rack hit 1.7 exaFLOPS with 27 TB of HBM.
This isn't a paper launch. The chip is co-designed with Broadcom — the same team behind Google's TPU and Apple's silicon interconnects — and the benchmarks were presented at a major industry conference. SemiAnalysis called it flatly: "Better than Nvidia Blackwell."
The implications are hard to overstate. For the last three years, the entire AI industry has been hostage to Nvidia's supply and pricing. Every hyperscaler has been designing custom silicon (Google TPU, AWS Trainium, Microsoft Maia), but nobody has published benchmarks that decisively beat Nvidia's best at the system level. OpenAI just did.
Why this matters right now:
- Inference is where the money is. Training gets the headlines, but inference is 80%+ of long-term AI compute costs. A 2x efficiency gain per watt rewrites the economics of serving models at scale.
- The monopoly narrative fractures. If OpenAI can run GPT-5-class models on its own silicon at half Nvidia's cost, the margin structure of the entire AI supply chain shifts.
- Broadcom becomes the shadow king. TPU, now Jalapeño, and likely more to come — Broadcom is quietly the ASIC powerhouse behind the AI boom.
Nvidia's Vera Rubin (GB300) hasn't even shipped in volume yet. The response — whether faster Rubin, a new architecture pivot, or aggressive pricing — will define the next 18 months of AI infrastructure.
This is a signal worth watching closely. The gap is no longer "can custom silicon compete?" — it's "how fast can the rest of the industry catch up?"
flowchart LR
A[OpenAI Jalapeño
700W ASIC] --> B[1.9x perf/watt vs Nvidia
3.6x lower latency]
C[Nvidia GB300
1,400W flagship] --> B
D[Broadcom
co-designer] --> A
E[Cost to serve
inference drops] --> F[Market structure
re-evaluation]