OpenAI presented the first benchmarks for its custom inference chip, Jalapeño, on Tuesday, and the comparison it chose was Nvidia (NASDAQ: NVDA). On SemiAnalysis’s public InferenceX benchmark, the chip, built in partnership with Broadcom (NASDAQ: AVGO), delivered 1.5 to 1.9 times more work per watt than an Nvidia. Blackwell system, with end-to-end response times 1.7 to 3.6 times lower. Sam Altman summed it up on X: “we made a chip and it is fast.”
The benchmarks landed at the Hot Chips conference the day before Nvidia reports second-quarter earnings, into a market that treats Nvidia as the only credible supplier of large-scale inference capacity. For an investor, the question is whether custom silicon commissioned by one of the largest buyers of AI accelerators changes the investment case for Nvidia and Broadcom, or whether a first-generation part shipping in low volume in late 2026 is still years from mattering.
we made a chip and it is fast
— Sam Altman (@sama) August 25, 2026
What OpenAI Showed at Hot Chips
Jalapeño is OpenAI’s first custom accelerator, an inference chip, meaning it serves trained models rather than training them, designed specifically for large language model workloads. OpenAI ran SemiAnalysis’s InferenceX benchmark across three open models, GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T, and reported the 1.5-to-1.9x per-watt figure at peak throughput, rising to 2.1 to 4.1 times faster on the highly interactive, low-latency traffic that a product like ChatGPT actually generates.
“The bottom line is that the results show a very, very significant performance advance over state of the art,” said Richard Ho, OpenAI’s head of hardware.
OpenAI’s own benchmark results for Jalapeño on SemiAnalysis’s InferenceX test, across GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T. The perf-per-watt range at previous-best interactivity (8.6x to 104.3x) reflects a narrow throughput-at-matched-latency comparison, not overall speed. Source: OpenAIThese are OpenAI’s own measurements. SemiAnalysis ran the benchmark alongside OpenAI’s engineers in the lab and vouches for the setup, but the chip has not been released for independent testing, and OpenAI selected the models and the Nvidia configuration it was measured against. The benchmark framework is public and reproducible, which anchors the baseline, but the results themselves are the vendor’s. That is the frame for every “beats Nvidia” claim here.
Since announcing Jalapeño, our first custom inference chip, we’ve been testing it and the system around it.
The results show a major advance: more intelligence from every watt and faster responses, delivering both higher throughput and lower latency in one architecture without… pic.twitter.com/vj7VOrA8pP
— OpenAI (@OpenAI) August 25, 2026
Investor Takeaway
The load-bearing numbers are the 1.5-to-1.9x-per-watt and 1.7-to-3.6x-latency figures, which are OpenAI’s own measurements on a public benchmark, so they describe a real result on a reproducible test rather than an independently verified claim about the chip.
Blackwell Is the Wrong Fight, and Jalapeño May Win the Right One
The headline compares Jalapeño to Blackwell, but SemiAnalysis itself calls that comparison incomplete. The fair fight, it argues, is Nvidia’s newer Vera Rubin platform, since both Jalapeño and Rubin use HBM4 memory and OpenAI taped out its chip after Rubin. On that comparison, SemiAnalysis reports Jalapeño still edges Rubin on output tokens per megawatt, even though Rubin uses a speculative-decoding optimization Jalapeño has not adopted yet. On total cost per token, the two come out roughly even.
That last point matters more than the raw performance, and it complicates the bull case. SemiAnalysis notes that part of Jalapeño’s cost advantage comes from “trading Nvidia’s high margins for Broadcom’s lower (though still high) margins.” In other words, some of the edge is not silicon superiority but a margin structure: OpenAI pays Broadcom’s markup instead of Nvidia’s. That is a real saving for OpenAI, but it reframes the story from “OpenAI built a better chip” to “OpenAI found a cheaper path to comparable performance,” which is a different, and for Nvidia less threatening, claim.
Broadcom’s Role and Why AVGO Is the Other Trade
The chip is OpenAI’s design, but the silicon is Broadcom’s, with system integrator Celestica handling board and rack work. This is the same custom-silicon-versus-merchant-GPU thesis running through FinanceFeeds’ coverage of Marvell’s custom AI chip deal with Google: the largest AI buyers commissioning bespoke accelerators to cut their dependence on Nvidia’s expensive GPUs. Broadcom sits on the winning side of that shift, since every hyperscaler that builds custom silicon is a potential Broadcom customer, and the company has guided toward $100 billion in AI revenue by 2027.
The relationship also runs deep on the demand side. OpenAI is both a massive Nvidia customer, the subject of FinanceFeeds’ reporting on Nvidia trimming its OpenAI-linked Ohio financing guarantee, and now a Broadcom design partner, and it is racing toward a public listing, as covered in FinanceFeeds’ look at OpenAI’s planned 2027 IPO. A company building its own chips while still buying Nvidia’s in volume is not defecting; it is diversifying.
What It Does and Doesn’t Mean for Nvidia’s Q2 Print
Jalapeño reportedly has not moved beyond engineering samples, while Rubin systems are already shipping to customers. OpenAI expects to deploy Jalapeño in very small volumes by the end of 2026, with broader scaling into 2027, and it will run those chips alongside the Nvidia systems it already operates. By the time Jalapeño reaches volume, Nvidia’s frontier will have moved again, so the lead measured today may not hold at deployment.
For tonight’s Nvidia earnings, that makes Jalapeño a narrative pressure, not a numbers event. It does not touch a dollar of the revenue Nvidia reports for last quarter, and it does not change what Nvidia ships in 2026. What it does is put a credible, named alternative on the board for inference, the workload growing fastest as AI usage scales, and give investors a concrete reason to ask how durable Nvidia’s inference monopoly really is. The memory-cost pressure FinanceFeeds examined in Nvidia’s AI-server price hike is the near-term margin question; custom silicon from its own largest customers is the long-term one.
Investor Takeaway
Jalapeño is years from volume and still in engineering samples, so it changes nothing in tonight’s Nvidia report and little in 2026, making it a story about the long-term inference market rather than near-term revenue.
