# OpenAI’s New Jalapeno Chip Beats NVIDIA’s Blackwell On Some Parameters, Company Says > Source: > Published: 2026-08-25 16:42:53+00:00 OpenAI is looking to own more parts of the AI stack as it prepares for its IPO. The company has put out its first performance results for Jalapeño, the custom inference chip it has been [building with Broadcom](https://officechai.com/ai/openai-has-designed-and-built-its-first-ai-chip-named-jalapeno-in-partnership-with-broadcom/), and the numbers show it outperforming NVIDIA’s Blackwell GPUs on throughput per watt and latency across several open models. OpenAI has also had the chip independently checked by SemiAnalysis, the semiconductor research firm whose engineers visited OpenAI’s labs and ran the workloads themselves. According to OpenAI’s own writeup, Jalapeño delivered between 1.5 and 1.9 times more AI work per watt at peak throughput compared to the systems it was tested against, along with 1.7 to 3.6 times lower end-to-end latency. For workloads that demand high interactivity, the advantage climbed to between 2.1 and 4.1 times. These figures came from testing on three publicly available models: GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T, meaning the chip wasn’t just tuned to run OpenAI’s own architecture well. That last point matters, because a lot of the coverage since OpenAI first unveiled the chip has assumed Jalapeño is essentially a specialized accelerator built around OpenAI’s own models, in the way a game console is built around its own games. SemiAnalysis, which got to run its InferenceX benchmark suite on the chip in person, pushed back on this framing directly, calling Jalapeño a generalized inference chip rather than a narrow one. The firm noted that OpenAI’s engineers even had the chip running Doom, ported over using nothing but Codex prompts, as a demonstration of how flexible the underlying software stack is. On the actual numbers, SemiAnalysis found Jalapeño beating Blackwell on performance-per-watt in nearly every scenario it tested, and doing so without multi-token prediction, a technique the comparison chips were using to their advantage. At low concurrency on DeepSeek R1, Jalapeño reportedly crossed 700 tokens per second per user, a level of responsiveness the firm said was achieved using single-token prediction alone, with no speculative decoding and no separation between prefill and decode chips. SemiAnalysis was careful to flag caveats alongside the results. All the underlying numbers came from OpenAI, and while the firm verified the benchmark runs in person, it hasn’t yet run its full InferenceX suite independently or tested Jalapeño against AgentX, a newer benchmark built around long-context, multi-turn agentic workloads that tend to expose weaknesses hidden by simpler tests. The firm also argued that comparing Jalapeño to Blackwell isn’t entirely fair to begin with, since Blackwell is now an older generation. The more relevant rival is NVIDIA’s upcoming Rubin platform, which also uses HBM4 memory and is closer to Jalapeño in terms of when it’s shipping. On that comparison, SemiAnalysis found Jalapeño’s throughput per megawatt actually beating Vera Rubin’s published July results too, though the two land close to even on cost per token once accounted for TCO. Architecturally, Jalapeño takes a different approach from a typical GPU. OpenAI has designed the chip to minimize the movement of data between cores and chips, keeping the KV cache used during response generation local rather than shuttling it across a network. The company also chose not to separate prefill and decode work across dedicated pools of chips, a design decision SemiAnalysis said was unusual given how much NVIDIA and AMD hardware benefits from that kind of separation. OpenAI’s reasoning, per its own blog post, is that real production traffic never sits at one fixed input-output ratio for long, so a rigid split just leaves chips idle at the wrong moments. The chip also leans heavily on AI in its own development. OpenAI says its models helped design Jalapeño’s arithmetic circuits and are now central to how engineers program it, using Codex to write and optimize kernels that would otherwise take specialist engineers far longer. In one example OpenAI cited, AI-generated code for select attention and mixture-of-experts blocks in GPT-OSS ran 1.5 to 1.8 times faster than the versions written by human engineers. SemiAnalysis’s reporting adds more texture here, noting that OpenAI had no working implementation of the attention mechanism used by DeepSeek’s models until the team needed one for benchmarking, and that Codex reportedly filled that gap. Jalapeño isn’t going into production at scale yet. OpenAI has engineering samples running today, with plans to start deploying the chip within its own infrastructure by the end of the year. Actual manufacturing volume is expected to ramp gradually through 2027. The company has said this is only the first of several planned generations, with a second version already deep into development and a third being scoped out. None of this means OpenAI is walking away from NVIDIA. The company has reiterated it will keep buying and deploying NVIDIA and other partners’ chips at scale for both training and inference. What Jalapeño represents instead is optionality: a way for OpenAI to control a piece of its own hardware economics rather than depend entirely on external suppliers, following the same instinct that has pushed it to sign chip [deals with Broadcom, AMD, and others](https://officechai.com/ai/these-are-the-tech-deals-signed-by-openai-this-year/) over the past year. With an [IPO reportedly on the horizon](https://officechai.com/ai/openai-has-missed-its-own-new-user-and-revenue-targets-wsj/), the ability to point to owned silicon that beats the market leader on some metrics is exactly the kind of story OpenAI needs public investors to believe.