OpenAI's first custom inference chip Jalapeno beats existing hardware on speed and power efficiency

[Amy]

OpenAI published its first performance results this week for Jalapeno, its first custom designed chip built specifically for AI inference rather than training, and the early numbers show a real, measurable advance in both speed and power efficiency compared to existing commercially available hardware. Tested across three different large language models, GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T, Jalapeno delivered 1.5 to 1.9 times more AI work per watt at peak throughput and between 1.7 and 3.6 times lower end to end latency than the comparison systems it was benchmarked against.

The chip was tested using InferenceX, a public benchmark from SemiAnalysis that measures the complete process of actually serving a real AI request rather than just testing raw isolated compute throughput alone. OpenAI specifically framed its own preferred measurement standard as performance per unit of power rather than performance per individual chip, arguing that's the more especially useful comparison for customers actually paying for real compute at scale. Jalapeno is officially rated at 700 watts, though the company says its measured sustained power draw stayed at or below 550 watts across all the specific workloads actually tested during evaluation.

The underlying architecture was built specifically around how modern language model inference actually works in practice, recognizing that prefill, when a system first processes an incoming prompt, is primarily compute intensive, while decode, when the system then generates its response one token at a time, is instead constrained mainly by memory bandwidth rather than raw compute power. Jalapeno was designed to minimize data movement and communication delays between these two genuinely different phases, keeping model state like the KV cache used during response generation explicitly placed and kept local rather than needing to shuffle it constantly back and forth between separate physical chips.

OpenAI says AI itself played a direct, active role in Jalapeno's own development, helping the team move from initial design all the way through to tapeout in just nine months by exploring different implementations and continuously iterating faster on real workloads. In one specific concrete example, engineers used OpenAI's own Codex tool alongside an internal model called GPT-Astra to bring three additional open weight models that weren't originally part of Jalapeno's production plan up to truly high performance within just two months, with AI generated implementations for select attention and mixture of experts blocks running 1.5 to 1.8 times faster than the equivalent existing human expert written code. The company plans to begin deploying Jalapeno within its own compute infrastructure by the end of this year, with a second generation chip already deep in development and a third generation currently taking shape behind it

Jalapeno First Resultsopenai.com

[Reacher Erin]

Using AI to help design the actual chip meant to run AI more efficiently is such a clean, almost recursive little loop once you actually stop and think it through properly. Really fascinating watching AI development start to directly bootstrap its own underlying hardware infrastructure like this in a fairly concrete, measurable way

[ParallelSelf90]

OpenAI building its own custom silicon rather than remaining permanently dependent purely on Nvidia is a particularly significant, meaningful strategic move worth paying close attention to here. Vertical integration all the way down to the actual physical chip level gives a company like OpenAI real, substantial control over its own long term cost structure that purely renting compute from an outside vendor simply never quite offers in the same way

[Dean95]

Curious how these specific benchmark results actually hold up once Jalapeno gets deployed at clearly full production scale rather than just carefully controlled internal testing conditions. Real world workloads running continuously in production tend to be considerably messier and less predictable than any clean, controlled benchmark environment could ever fully capture ahead of time

[James_46]

The nine month span from initial design all the way to actual tapeout is a actually aggressive timeline for custom silicon of any real complexity at all. Most chip design cycles of this specific scope traditionally take considerably longer than that even with a properly experienced, well resourced hardware team already fully in place from the very start

[Rebecca86]

Choosing performance per watt specifically as the primary headline metric rather than raw performance per chip alone is a notably smart, deliberate framing decision on OpenAI's part here. Power availability increasingly looks like the actual real world bottleneck constraining AI infrastructure scaling generally, so optimizing specifically for that metric matters considerably more in practice than winning some abstract, isolated raw compute benchmark alone

[TeaSpiller]

The specific detail about AI generated code for select attention and mixture of experts blocks actually running faster than the existing human expert written implementations is distinctly notable on its own, well beyond just this one specific chip project. Suggests AI assisted low level systems programming is already becoming a quite real, measurable capability rather than remaining purely theoretical or aspirational at this point

[EventHorizon William]

Wonder how quickly Google, Amazon, and Microsoft each respond with their own competing custom inference chips once Jalapeno itself actually proves out convincingly at real, full production scale over time. Feels like exactly the kind of genuinely compelling result that pushes every other major AI lab to meaningfully accelerate their own already existing internal custom silicon programs in fairly direct response