OpenAI's Jalapeño Chip: 9 Months to Beat Nvidia Efficiency

OpenAI's first custom inference chip posted 1.5-1.9x Nvidia efficiency in SemiAnalysis-witnessed tests, after just 9 months from design to tapeout. Here is what the numbers really say.

🌶️ A Scorecard That Should Not Exist

On August 26, 2026, OpenAI published the first real benchmarks for Jalapeño, its first custom inference chip, and the numbers read like a typo. In SemiAnalysis's InferenceX tests — witnessed on-site by the analysts — Jalapeño beat Nvidia's GB200/GB300 systems on every model it ran: GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T.

The headline figures: 1.5–1.9x more work per watt, 1.7–3.6x lower end-to-end latency, and 2.1–4.1x better performance on interactive workloads. Running DeepSeek R1, a single user got up to 700 tokens/s on Jalapeño versus about 169 tokens/s on GB300. On GPT-OSS 120B, the gap was 1,459 versus 535 tokens/s. On the trillion-parameter Kimi K2.5, a full request took 1.56 seconds versus 5.31 seconds — and at a fixed quality of 100 tokens/s per user, Jalapeño served more than 9x the concurrent users of GB300.

SemiAnalysis founder Dylan Patel, not known for charity toward chip startups, put it bluntly: "First-generation chips are usually uncompetitive, but OpenAI is beating Nvidia Blackwell — even Rubin." Sam Altman's version was shorter: "We built a chip that's ridiculously fast."

⚡ 9 Months, 700 Watts, One Idea: Stop Moving Data

The most shocking number isn't in the benchmarks — it's the calendar. Jalapeño went from initial design to tape-out in about 9 months, against an industry norm of 18–24 months for a high-end ASIC (Google's first TPU and Amazon's first Trainium each took roughly two years). OpenAI announced the Broadcom partnership in October 2025; within a year, it had third-party-witnessed scores.

OpenAI designed the architecture; Broadcom handled implementation and networking; Celestica built the boards and racks. The chip is rated at 700W (sustained measured under 550W) with HBM4 memory delivering roughly 15.4TB/s of single-package bandwidth, plus a separate I/O die for rack-scale communication.

The design philosophy fits in one sentence: compute is nearly free, moving data is expensive. Jalapeño slices compute cores and memory into matched pairs, so each core keeps its working data local instead of fetching it across the chip. And unlike Nvidia's Rubin generation — which splits prefill onto a dedicated CPX chip — Jalapeño does prefill and decode on the same accelerator, because real traffic mixes keep shifting and a fixed split leaves half the silicon idle.

Even the chip's creation was AI-accelerated: in parts of GPT-OSS's attention and MoE code, AI-generated kernels ran 1.5–1.8x faster than the human experts' versions, and OpenAI used Codex and GPT-Astra to port DeepSeek R1 and Kimi K2.5 in roughly two months — including implementing DeepSeek's MLA attention from scratch.

💰 Why This Is Really About API Prices

SemiAnalysis did the unglamorous math — power, cooling, networking, amortized over each chip. Result: about 1.125kW per Jalapeño, versus 1.87kW for GB200 and 3.3kW for Vera Rubin. Running GPT-OSS, that translates to roughly 53 million tokens per megawatt per second, against about 10 million for a GB200 NVL72 rack. Total cost of ownership came out to $1.56 per chip-hour — statistically identical to an H100's $1.55, while Vera Rubin sits at $3.61.

Every watt saved is margin on the next price cut. That matters because OpenAI also gave competitors' models identical treatment in its tests — a quiet signal that anyone's model could run near-optimally on this chip. A chip that cheaply serves any model is a shot at the pricing power Nvidia currently holds.

⚠️ Read the Fine Print Before the Obituaries

The caveats are real. The benchmark data was provided by OpenAI; SemiAnalysis witnessed the runs but did not independently execute the full suite. Every published result used a relatively easy 8K-in/1K-out single-turn profile — and there are no numbers yet for long-context or multi-turn agent workloads, the fastest-growing and hardest inference category. The tested silicon is the A0 engineering version; the B0 revision is entering TSMC's N3P process. Deployment starts at "very small scale" in late 2026, with volume only in 2027 — and a second-generation chip is already in deep development.

Meanwhile, Nvidia is not exactly fading: its Q2 FY2027 revenue hit $96.2 billion, up 106% year-over-year. But the direction is set. For twenty years the industry adapted models to chips; Jalapeño is the clearest proof yet that the order has flipped — the model makers are now building chips for their own models.