OpenAI has built its own AI chip, and the company says early benchmark testing shows it running circles around Nvidia's current flagship hardware on the exact job that now eats up most of the industry's computing budget: answering people's prompts. The chip, called Jalapeño, was developed with Broadcom Inc. and is designed strictly for inference, the work a model does after it has already been trained, rather than for training new models from scratch.
According to the Los Angeles Times, OpenAI measured Jalapeño against Nvidia's GB300, described as the leading option on the public benchmarking system used in the test. OpenAI’s chip chief, Richard Ho, is set to detail the results Tuesday at the Hot Chips conference at Stanford University. Per the same report, Jalapeño led the GB300 both in response-return speed and in AI work handled per unit of power, while running on 700 watts.
Independent benchmark numbers add texture to that claim. In testing across models including GPT-OSS 120B, DeepSeek and Kimi K2.5, Jalapeño delivered 1.5 to 1.9 times more throughput per kilowatt and 1.7 to 3.6 times lower end-to-end latency than Nvidia’s GB200 and GB300 systems, according to Tom's Hardware. That efficiency came while Jalapeño ran at a 700-watt package thermal design power, compared with Nvidia’s 1,200-watt and 1,400-watt flagship accelerators, the outlet reports.
Where the Speed Gap Widens
OpenAI itself reported an even larger gap on highly interactive workloads — the kind where an AI agent has to respond immediately rather than churn through a long batch job. The company said Jalapeño provided 2.1 to 4.1 times higher performance on those tasks, and between 8.6 and 104.3 times higher throughput per kilowatt when matched against the GB300's fastest decoding speed settings, per Tom's Hardware's account of OpenAI's own figures. The Los Angeles Times notes that Jalapeño delivered its widest advantages on Moonshot AI’s Kimi model, the largest model OpenAI tested the chip against, with the company publicly testing using third-party models from DeepSeek and Moonshot AI as well as one of its own smaller open-source models.
Jalapeño performed well on some of OpenAI's own unreleased advanced models too, though the company was not specific about which ones, per the Los Angeles Times. OpenAI has said Jalapeño's design becomes more valuable as workloads grow larger and more demanding, according to a company blog post cited by the Times.
A Chip Built in Record Time
What stands out to hardware watchers is how fast this happened. Jalapeño is fabricated on TSMC's 3-nanometer process node and went from initial team hiring and architectural concept in mid-2024 to manufacturing tape-out in late 2025 — roughly 16 months, according to a newsletter published by SemiAnalysis. Typical ASIC design cycles often take two to three years, per the same account. The Los Angeles Times reports OpenAI used its own AI models to help develop Jalapeño more quickly, and that OpenAI has a second Jalapeño version already far along in development, with tape-out for that chip expected in the coming months.
Inside each package, a custom compute die sits on a 2.5D interposer alongside six to eight HBM4 high-bandwidth memory stacks, providing 216 GiB of memory and 15.4 TB/s of memory bandwidth, according to TrendForce, which identifies Samsung Electronics as a key HBM4 supplier. Memory bandwidth is widely regarded as the primary hardware bottleneck for large language model inference, which is part of why OpenAI is emphasizing those specs.
Nvidia Isn't Going Anywhere Yet
None of this means OpenAI is walking away from its biggest supplier. Ho said Nvidia remains a good partner for OpenAI, and the Los Angeles Times reports OpenAI continues to need a lot of Nvidia chips even as it builds out its own infrastructure. Jalapeño was not tested against Nvidia's Vera Rubin generation, which began shipping in 2026, and OpenAI has committed to deploying that upcoming platform as well.
Ho told the Times that OpenAI's chips will bring infrastructure costs down and should reduce costs substantially as deployment expands, with the company saying it will decide which AI models run on Jalapeño and that customers can choose options offering either cost savings or better performance. OpenAI plans to start using Jalapeño chips to support its AI models later this year, and the Times reports that expanded use is conditioned on increased production and material cost savings.
The Jalapeño push builds on a partnership OpenAI and Broadcom announced in 2025. That agreement formalized a multi-year plan to co-develop and deploy 10 gigawatts of custom AI accelerators across OpenAI data centers and partner facilities through 2029, according to Data Centre Magazine. Broadcom, which makes custom chips for a variety of clients, has projected its annual AI semiconductor revenue will reach $56 billion in fiscal 2026, while Wall Street analysts at Citi have forecast the figure could climb to $116 billion in fiscal 2027, per Seeking Alpha.
OpenAI Is Not Alone in Building Its Own Silicon
Jalapeño also won't replace providers such as Cerebras Systems Inc. anytime soon, the Los Angeles Times reports, noting that OpenAI has signed up many different compute providers and that Cerebras technology — which supplies some models OpenAI uses — is best suited to smaller AI models. The Times also points to smaller players chasing the same niche: Etched has raised funds at a $21 billion valuation and is starting to ship a low-voltage chip, while MatX, founded by two alumni of Google's silicon business, is working on a semiconductor targeting high throughput and low latency.
The broader shift toward inference-focused silicon reflects where the money is moving. AI inference workloads are projected to account for 60% to 70% of a nearly $400 billion AI accelerator market by 2026, up from roughly 40% in 2023, according to a report from Six Point Ventures via the Value Add VC. Custom hyperscaler accelerators overall — including Microsoft's Maia 200, Meta's MTIA and Google's TPUs — are estimated to reach roughly 1.9 million units deployed in 2026, per Tom's Hardware.
None of that has dented Nvidia's dominance of the AI chip field. Nvidia currently secures roughly 60% of TSMC's advanced packaging capacity, which the foundry is expanding from 65,000 to 75,000 wafers per month in 2025 to 120,000 to 130,000 in 2026, according to a separate Tom's Hardware report on the industry's manufacturing constraints. Nvidia CEO Jensen Huang said in March that the company sees line-of-sight to $1 trillion in AI hardware orders through 2027, citing what he called an “inference inflection point,” per Constellation Research.
OpenAI, for its part, says it wants to be public about its chip development as it builds AI infrastructure across the globe and works toward handling more of its own AI tasks in-house. The company is already designing concepts for a third-generation chip, according to the Los Angeles Times, even as this second Jalapeño version nears tape-out — a sign that OpenAI's silicon ambitions are meant to run alongside its Nvidia deals for years to come, not replace them overnight.
-4.webp?w=160&h=400&fit=crop&crop=edges)
