Jalapeño
OpenAI's first custom inference chip, improved with model-assisted gains verified on real silicon.
Key Highlights
- Jalapeño is OpenAI’s first custom inference chip, announced with Broadcom for large language model workloads.
- OpenAI reported Jalapeño delivers 1.5–1.9× more AI work per watt and 1.7–3.6× lower end-to-end latency across multiple models.
- On highly interactive workloads, Jalapeño was reported to achieve 2.1–4.1× higher performance.
- OpenAI said Jalapeño reached tapeout in nine months with AI-assisted design and that some AI-generated kernels outperformed human-written versions.
- OpenAI later said model-assisted improvements to Jalapeño were verified on real silicon.
At a glance
- Category
- Hardware
Jalapeño
Overview
Jalapeño is OpenAI’s first custom inference chip, developed with Broadcom and optimized for large language model workloads. It is positioned as infrastructure for ChatGPT, Codex, the API, and future agentic products, with the goal of improving inference performance and energy efficiency as AI usage scales.For AI Product Managers, Jalapeño matters because it connects model performance directly to product economics and user experience. OpenAI’s reported results suggest meaningful gains in work per watt, latency, and interactive performance across major open models, while also showing how AI-assisted chip design can shorten development cycles and improve silicon outcomes.
Key Developments
- 2026-06-25: OpenAI and Broadcom announced a collaboration to release Jalapeño, an inference chip optimized for large language model workloads, aimed at improving inference performance and energy efficiency.
- 2026-08-26: OpenAI reported first results for Jalapeño, saying it delivered 1.5–1.9× more AI work per watt at peak throughput and 1.7–3.6× lower end-to-end latency across GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T. On highly interactive workloads, performance was reported at 2.1–4.1× higher, and on Kimi specifically, Jalapeño achieved about 1.5× higher peak perf/W and about 3.4× lower latency while staying below its 700 W rating, with sustained power at ≤550 W in tests.
- 2026-08-26: OpenAI also said the chip and rack-scale system were co-designed with software and networking to minimize data movement, that Jalapeño reached tapeout in nine months with AI-assisted design, and that AI-generated kernels ran 1.5–1.8× faster than human-written implementations for selected attention and MoE blocks.
- 2026-11-19: OpenAI shared that its models identified up to 14% PPA improvement on AMD flagship GPU IP in under five months and also improved Jalapeño, with gains verified on real silicon. The update reinforced Jalapeño as an example of model-assisted hardware design affecting production chip outcomes.
Relevance to AI PMs
- Model serving costs and margins: Jalapeño’s reported improvements in work per watt can materially change inference unit economics. AI PMs responsible for pricing, gross margin, or scaling strategy should track how hardware efficiency affects feature viability and API economics.
- Latency-sensitive product design: Lower end-to-end latency and stronger interactive workload performance matter for chat, coding, and agent experiences. AI PMs can use these kinds of hardware gains to justify more responsive UX patterns, longer context interactions, or more tool-calling steps within latency budgets.
- Roadmap planning across the stack: Jalapeño shows that product outcomes increasingly depend on co-design across models, kernels, networking, and silicon. AI PMs should plan roadmaps with infrastructure assumptions in mind, especially when launching features that depend on sustained throughput, power efficiency, or fast token response times.
Related
- OpenAI: Creator of Jalapeño and the company positioning it as infrastructure for ChatGPT, Codex, the API, and future agentic products.
- Broadcom: Collaborated with OpenAI on the Jalapeño inference chip.
- ChatGPT: One of the core product contexts for which Jalapeño is framed as enabling infrastructure.
- Codex: Another OpenAI product tied to the broader inference infrastructure strategy around Jalapeño.
- GPT-OSS 120B, DeepSeek R1, Kimi K2.5 1T: Models used in reported Jalapeño performance comparisons for throughput, latency, and efficiency.
- AMD: Referenced in OpenAI’s chip-design update, where model-assisted methods reportedly found up to 14% PPA improvement on AMD flagship GPU IP and also improved Jalapeño.
Frequently asked questions
What is Jalapeño?
OpenAI's first custom inference chip, improved with model-assisted gains verified on real silicon.
What should AI product managers know about Jalapeño?
Jalapeño is OpenAI’s first custom inference chip, announced with Broadcom for large language model workloads. OpenAI reported Jalapeño delivers 1.5–1.9× more AI work per watt and 1.7–3.6× lower end-to-end latency across multiple models. On highly interactive workloads, Jalapeño was reported to achieve 2.1–4.1× higher performance.
Newsletter Mentions (3)
“OpenAI shared that its models identified up to 14% PPA improvement on AMD flagship GPU IP in less than five months, building on designs AMD engineers had spent years optimizing, and improved OpenAI’s first custom inference chip, Jalapeño, with gains verified on real silicon.”
#6 𝕏 OpenAI shared that its models identified up to 14% PPA improvement on AMD flagship GPU IP in less than five months, building on designs AMD engineers had spent years optimizing, and improved OpenAI’s first custom inference chip, Jalapeño, with gains verified on real silicon. OpenAI is also working with Synopsys to bring its model capabilities and Synopsys EDA design and validation tools to more chip designers.
“OpenAI’s custom inference chip Jalapeño delivers 1.5–1.9× more AI work per watt at peak throughput and 1.7–3.6× lower end-to-end latency across GPT‑OSS 120B, DeepSeek R1, and Kimi K2.5 1T (with 2.1–4.1× higher performance on highly interactive workloads), and on Kimi specifically achieved ~1.5× higher peak perf/W and ~3.4× lower latency while running below its 700 W rating (sustained ≤550 W) in tests.”
#2 📝 OpenAI News Jalapeño’s first results show industry-leading speed and efficiency in AI inference - OpenAI’s custom inference chip Jalapeño delivers 1.5–1.9× more AI work per watt at peak throughput and 1.7–3.6× lower end-to-end latency across GPT‑OSS 120B, DeepSeek R1, and Kimi K2.5 1T (with 2.1–4.1× higher performance on highly interactive workloads), and on Kimi specifically achieved ~1.5× higher peak perf/W and ~3.4× lower latency while running below its 700 W rating (sustained ≤550 W) in tests. The chip and rack-scale system were co-designed with software and networking to minimize data movement, were brought to tapeout in nine months with AI-assisted design, and AI-generated kernels ran 1.5–1.8× faster than human-written implementations for selected attention and MoE blocks. Also covered by: @OpenAI News , @OpenAI
“OpenAI and Broadcom announced a collaboration to release a Jalapeño inference chip optimized for large language model workloads, aiming to boost inference performance and energy efficiency.”
The chip is presented as OpenAI's first AI chip and is tied to ChatGPT, Codex, the API, and future agentic products. It is framed as infrastructure for scaling AI access.
Related
AI lab and product company referenced for chip optimization work, infrastructure partnerships, and governance reporting on AI progress.
An AI coding tool mentioned as a supported client for ffmpeg-skill. It is listed with Claude Code and Cursor in the newsletter.
OpenAI’s consumer AI assistant, mentioned here as one of the integrations in HubSpot’s refounding discussion. It remains a baseline product for AI app integrations and workflows.
Semiconductor company whose flagship GPU IP was improved with help from OpenAI's models.
Stay updated on Jalapeño
Get curated AI PM insights delivered daily — covering this and 1,000+ other sources.
Subscribe Free