OpenAI has revealed that its custom-built Jalapeño chip, developed in partnership with Broadcom, outperforms Nvidia's GB200 and GB300 systems in power efficiency by up to 1.9x. This marks a significant step in the AI industry's shift towards custom silicon solutions.
TL;DR
- OpenAI's Jalapeño chip, designed for inference, outperforms Nvidia's GB200 and GB300 systems in power efficiency by up to 1.9x.
- The chip is expected to begin small-volume deployment by the end of 2026, with a larger ramp in 2027.
- OpenAI continues to use Nvidia hardware for training and may not fully replace it with Jalapeño.
What happened
On August 25, 2026, OpenAI published the first benchmark results for its Jalapeño chip, an inference accelerator built in collaboration with Broadcom. The results were published using InferenceX, a public benchmark maintained by SemiAnalysis, and compared Jalapeño against Nvidia's GB200 and GB300 systems.
OpenAI's chip chief Richard Ho stated that Nvidia's GB300 was the leading publicly benchmarked system available for comparison. He also mentioned that Jalapeño is expected to begin deploying in very small volumes by the end of 2026, with a more significant production ramp arriving in 2027.
The Jalapeño chip is rated at 700W, significantly lower than Nvidia's GB200 (1,200W) and GB300 (1,400W). In sustained testing, Jalapeño ran at 550W or lower, which could lead to substantial cost savings in data centers.
Why it matters
The Jalapeño chip's superior power efficiency could lead to significant cost savings for AI companies, as inference is a major and growing cost center. This is particularly relevant for OpenAI, which operates ChatGPT and other AI services that require extensive inference capabilities.
OpenAI's move to develop its own silicon, along with similar efforts by Google, Microsoft, and Amazon, indicates a broader industry trend towards custom silicon solutions. This could narrow the gap between Nvidia being the only option and becoming the default option, impacting AI infrastructure costs and GPU pricing.
However, OpenAI has not indicated that it plans to abandon Nvidia hardware entirely. The company still trains its frontier models, including the GPT family, on Nvidia hardware and has a $100 billion partnership with Nvidia, including a planned 10 gigawatts of Nvidia systems starting with the Vera Rubin platform.
Key facts
- OpenAI's Jalapeño chip delivers 1.5x to 1.9x more AI work per watt than Nvidia's GB200 and GB300 systems.
- The chip is rated at 700W, significantly lower than Nvidia's GB200 (1,200W) and GB300 (1,400W).
- Jalapeño is expected to begin small-volume deployment by the end of 2026, with a larger ramp in 2027.
- OpenAI tested Jalapeño against three open-weight models: GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T.
- On GPT-OSS 120B, Jalapeño achieved a peak throughput of 85,448 tokens per kilowatt, compared to 44,960 tokens per kilowatt for Nvidia's GB200.
- OpenAI continues to use Nvidia hardware for training and may not fully replace it with Jalapeño.
- Broadcom is a common design partner behind Google's TPUs and OpenAI's Jalapeño, and it has been reported in talks around Anthropic's own custom chip ambitions.
Context
The development of custom silicon for AI inference is part of a broader industry trend. Google, Microsoft, and Amazon have been developing their own chips for years, with Google's Tensor Processing Units, Amazon's Trainium line, and Microsoft's Maia accelerators.
This trend is driven by the need to reduce costs and improve performance for AI inference, which is a major and growing cost center for AI companies. Custom silicon solutions can be tailored to specific AI workloads, potentially offering better performance and efficiency than general-purpose GPUs.
However, Nvidia remains the incumbent in the AI chip market, with a significant share of the market and a strong partnership with OpenAI. The company has a history of introducing new architecture generations roughly every one to two years, which could pose a challenge for custom silicon solutions like Jalapeño.
