OpenAI has unveiled Jalapeño, its first custom AI chip, co-developed with Broadcom for silicon and networking and Celestica for systems integration. Using SemiAnalysis's public InferenceX benchmark on models including GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5, OpenAI's self-reported results show Jalapeño delivering 1.5x to 1.9x more inference throughput than Nvidia's Blackwell-generation systems, with 1.7x to 3.6x lower end-to-end latency and up to 4.1x better performance on interactive workloads.
The chip is inference-only — it runs trained models rather than training them — and took 16 months to develop, with just nine months from design to fabrication in November 2025. VP of Hardware Richard Ho said it "achieves high throughput and low latency simultaneously, a first in the industry." Analyst Dylan Patel called it huge news for custom silicon challenging Nvidia in its very first generation.
The comparison comes with caveats: Jalapeño uses newer HBM4 memory, hasn't adopted speculative decoding optimizations yet, and only engineering samples exist so far. Volume production isn't expected until 2027, with limited deployment later in 2026. The announcement lands the same day Nvidia reports quarterly earnings, with analysts expecting roughly $91-92 billion in revenue.