IA · 25 August 2026 · 4 min read
OpenAI Unveils Jalapeño Benchmarks: Custom Silicon Challenges Nvidia on AI Inference
In brief: OpenAI has published the first official benchmark results for Jalapeño, its proprietary ASIC chip co-developed with Broadcom to accelerate AI model inference. Tested on SemiAnalysis' InferenceX suite, the processor outperforms Nvidia Blackwell architectures in both throughput per kilowatt and token generation speed. The chip is slated for limited deployment in late 2026, with volume scaling planned for 2027.
by Team Mocchi's
At the annual Hot Chips conference, OpenAI provided a detailed look into its hardware ambitions by unveiling the first comparative benchmark results for Jalapeño, its custom Application-Specific Integrated Circuit (ASIC) engineered specifically for AI inference. Developed in close collaboration with Broadcom, the silicon aims to eliminate the traditional trade-off between latency and throughput in large-scale AI operations.
Presented by OpenAI vice president of hardware Richard Ho, the data signals a deliberate shift toward full-stack vertical integration as frontier AI providers race to curb escalating data center energy costs.
Silicon tailored for model inference
Unlike general-purpose accelerators designed for large-scale training, Jalapeño is purpose-built to run production workloads, interactive chat, and autonomous agent loops. As reported by TechCrunch, OpenAI took a holistic design approach, utilizing its own AI models to assist with circuit layout and co-designing models, memory subsystems, and silicon in tandem.
Architecturally, Jalapeño targets the primary friction points of inference: prefill overhead and inter-chip communication delays. By minimizing on-board data movement and optimizing the management of the key-value (KV) cache during sequential token generation, the chip eliminates processing stalls that frequently degrade responsiveness under heavy concurrency.
InferenceX benchmarks: lower latency and higher energy efficiency
To establish standard comparisons, OpenAI evaluated Jalapeño using the InferenceX benchmark suite from SemiAnalysis, measuring performance against Nvidia's flagship GB200 and GB300 superchips. The evaluations spanned multiple open and frontier models, including GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T.
According to details shared by The Verge, Jalapeño delivered 1.5 to 1.9 times more AI work per watt across the tested models compared to Nvidia's systems. On responsiveness metrics, the chip achieved 1.7 to 3.6 times lower end-to-end latency, driving down the Time Between Tokens (TBT) critical for real-time applications.
This balance allows data center operators to pack substantially higher concurrent query capacity into a fixed power envelope without sacrificing the user-facing speed of agentic workflows.
Deployment timeline and co-existence with Nvidia
Despite the promising benchmark figures, Jalapeño's roll-out will be measured. OpenAI announced that early production units will deploy into data centers in small batches toward the end of 2026, with meaningful volume ramp-up scheduled for 2027.
Company leadership emphasized that the ASIC will not immediately displace Nvidia's hardware, but will operate alongside existing partner infrastructure in a hybrid compute strategy. Jalapeño is positioned as a multigenerational platform, with second- and third-generation iterations already in active development.
Mocchi's take
OpenAI's foray into dedicated inference silicon underscores a clear reality: software optimizations alone can no longer offset the energy demands of scaling model usage. For businesses building custom AI workflows and agentic pipelines, vertical hardware innovation is a welcome development. In the medium term, higher compute density per watt will help stabilize API pricing and deliver the low-latency response times essential for high-throughput enterprise automation.