OpenAI's Custom "Jalapeño" Inference Chip Beats Nvidia Blackwell on Performance per Watt in First Measured Benchmarks
OpenAI has officially unveiled its first custom-designed artificial intelligence inference chip, "Jalapeño." Developed in partnership with ASIC giant Broadcom and manufactured on TSMC's 3nm process, the chip represents OpenAI's first major physical step toward vertical integration to lower serving costs and reduce its reliance on Nvidia.
In initial benchmarks published ahead of Nvidia's earnings in late August 2026, Jalapeño demonstrated significant efficiency advantages over Nvidia's flagship rack systems:
- Throughput: Delivered 1.5× to 1.9× higher throughput per kilowatt than Nvidia's GB200 and GB300 NVL rack systems.
- Latency: Achieved 1.7× to 3.6× lower end-to-end latency under comparable workloads.
- Sustained Power: Remained at or below 550W during active testing.
Memory Architecture and Samsung HBM4 Integration
Jalapeño bypasses standard memory bottlenecks by integrating high-bandwidth memory directly into the package. Each Jalapeño chip combines its compute die with six HBM4 stacks, delivering 216 GiB of memory and 15.4 TB/s of bandwidth.
By comparison, Nvidia's GB300 features 288GB of HBM3e on a 1,400W power rating, giving Jalapeño roughly 50% more memory per watt. Reports from South Korean media indicate that Samsung Electronics is the primary supplier for Jalapeño's HBM4 stacks. If OpenAI scales production under its 10GW deployment roadmap with Broadcom, it could emerge as a massive, direct competitor for next-generation memory supply.
Benchmark Caveats and Deployment Limits
While OpenAI's chip marks a major technical milestone, key caveats remain:
- Inference Only: Jalapeño is strictly an inference chip and cannot be used for AI training, where Nvidia's Blackwell and Rubin platforms maintain an absolute monopoly.
- Model Compatibility: OpenAI's benchmarks were conducted on smaller open-source models, alongside third-party models like DeepSeek R1 and Moonshot AI's Kimi K2.5. This suggests OpenAI is building a versatile platform capable of serving diverse large-model workloads, rather than just its proprietary GPT series.
- Advanced Packaging Race: Because Jalapeño relies on TSMC's 3nm node and advanced CoWoS packaging, OpenAI remains locked in the same bottlenecked foundry queue as Nvidia and AMD.
OpenAI is already accelerating its custom silicon roadmap. According to industry reports, a second-generation Jalapeño tape-out is expected within months, with a third-generation design already underway.