OpenAI's Jalapeño chip beats Nvidia superchips on power and latency

Share
OpenAI's Jalapeño chip beats Nvidia superchips on power and latency

Two hardware stories today, both about who sets the price of running AI: OpenAI's first custom inference chip just posted benchmark numbers ahead of Nvidia's best, and one of autonomous trucking's quietest operators just bankrolled its expansion.

OpenAI's custom Jalapeño chip delivered 1.5–1.9x more AI work per watt than Nvidia's GB200 and GB300 superchips, with 1.7–3.6x lower end-to-end latency, according to results the company presented at the Hot Chips conference on Tuesday. The numbers come from InferenceX, an inference benchmark run by research firm SemiAnalysis, tested across GPT-OSS 120B, DeepSeek R1 and Kimi K2.5 1T. OpenAI hardware chief Richard Ho called it a "very, very significant performance advance over state of the art," noting that most systems force a trade-off between throughput and latency — Jalapeño escapes it by keeping each model's working state next to the compute serving it, instead of shuttling data across the machine mid-response.

The caveats matter as much as the headline. These are OpenAI-commissioned results on a benchmark its chip was tuned for, and the comparison ran against Blackwell-generation Nvidia systems — while Jalapeño ships only in small volumes late this year and ramps through 2027, by which point Nvidia will have moved too. Still, the direction is the news: an ASIC co-designed with Broadcom, built purely for inference, beating general-purpose GPUs on their own economics metric is the scenario Nvidia's biggest customers have been waiting for. Ho was careful to say Nvidia remains one of OpenAI's "very good partners" and that Jalapeño won't replace the whole fleet — which tells you OpenAI expects its own silicon to carry most of the inference load, not all of it.


Self-driving truck startup Gatik raised $200 million — its largest round yet — led by Qatar Investment Authority and Koch Disruptive Technologies, two months after signing a multi-year deal to run driverless Frito-Lay deliveries for PepsiCo across Texas, Arizona and Arkansas. Millennium Management, ARK Invest and Intact Private Capital joined the round, bringing Gatik's total funding to roughly $500 million since it emerged from stealth in 2019. The company says it holds $600 million in contracted revenue and operates dozens of fully driverless box trucks on commercial routes for Walmart, Kroger, Loblaw and Tyson Foods. While robotaxi operators still burn cash per ride, middle-mile freight has quietly become the corner of autonomy that actually gets paid.

What to watch: whether anyone outside OpenAI reproduces the Jalapeño numbers independently — and what Nvidia's next generation looks like before the 2027 volume ramp.

If a first-gen startup chip can nearly double the work per watt, how safe is Nvidia's pricing power? Write to us via our contact page.

Source: OpenAI

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

Samsung puts HBM base dies on a logic node to reclaim bandwidth

Samsung puts HBM base dies on a logic node to reclaim bandwidth

The memory stack feeding AI accelerators is getting a quiet overhaul. At Hot Chips 2026, Samsung laid out its plan to build High Bandwidth Memory base dies on a logic process node — a step change that frees up die area it wants to fill with memory controllers, sensors, and eventually compute. Samsung revealed it now fabricates its HBM4 and HBM4E base dies on a 4nm logic node, breaking with the long-running practice of building them on a DRAM process.

Devin factored RSA-260 on spare GPUs — for about $400,000

Devin factored RSA-260 on spare GPUs — for about $400,000

Cognition has finally published how its record factorization happened, and the answer is more interesting than the record: an engineer with no expertise in the math pointed a fleet of coding agents at a spare-compute problem and got a world-first in three weeks. Cognition engineer Eric Lu posted the write-up we asked for: RSA-260, the largest number from RSA Laboratories' 1991 challenge ever cracked, fell in about three weeks of part-time work at an estimated 4,900 GPU-days — roughly $400,000…