OpenAI Unveils Custom Jalapeño AI Chip, Challenging Nvidia’s Dominance
OpenAI, with Broadcom, announced its custom ASIC chip "Jalapeño" at the Hot Chips conference. Consuming 700W, it outperforms Nvidia’s 1,400W flagship on inference efficiency—delivering up to 1.9x more work per watt and 3.6x lower latency. Deploying by end of 2026, the chip uses HBM4 memory and targets inference workloads. While analysts see a threat to Nvidia’s margins, experts note the comparison is unfair without Nvidia’s upcoming Rubin platform. OpenAI will still use Nvidia for training.
Editorial responsibility
- No named human review is recorded for this page.
- Reports are grouped by semantic similarity and deterministic rules. Language models may assist titles, summaries, translation and cross-source analysis; the page itself is projected from evidence records.
- Source updates only; analysis pending
Analysis pending
Cross-source coverage
Wire timeline
OpenAI's Jalapeno AI chip beats Rubin on tokens per megawatt, taped out in nine months
OpenAI has unveiled Jalapeno, an AI-designed chip that reportedly outperforms the Rubin chip on tokens per megawatt, a key efficiency metric for AI inference. The chip was taped out in just nine months and is capable of running Doom. The announcement was made via a post on X, promoting a detailed breakdown on SemiAnalysis Weekly featuring Myron and Bryan alongside host Jordan Nanos. The episode is expected to cover comparisons between Jalapeno, Rubin, and GB300 on InferenceX, the importance of tokens per megawatt as a metric, Samsung's HBM4 comeback, AI-written RTL and kernels, the potential draining of the CUDA moat, and Anthropic's upcoming ASIC. This development signals a significant advancement in custom AI hardware design and competition in the AI chip market.
OpenAI's Jalapeño Chip Outperforms Nvidia on Power Efficiency, Not Raw Chip Count
At the Hot Chips conference on August 25, 2026, OpenAI presented first measured results for its custom inference chip, Jalapeño. Benchmarks using the InferenceX standard from SemiAnalysis showed Jalapeño delivering 1.5 to 1.9 times more AI work per watt at peak throughput and 1.7 to 3.6 times lower end-to-end latency compared to Nvidia's GB200 and GB300 rack systems. The tests were normalized to thermal design power ratings: 700 watts for Jalapeño versus 1,200 for GB200 and 1,400 for GB300. SemiAnalysis noted the comparison was 'somewhat incomplete and unfair' because Jalapeño uses HBM4 memory while Nvidia's tested chips use HBM3E. Against Nvidia's upcoming Vera Rubin platform (which uses HBM4), the two produce almost the same output tokens per dollar. OpenAI emphasized that power efficiency, not raw chip performance, is the critical metric as data center electricity demand surged 17% in 2025. OpenAI's hardware head Richard Ho estimated deployment would begin at end of 2026 in very small volumes. The chip's nine-month design cycle from initial design to tapeout was highlighted as a significant achievement.
OpenAI Built an Nvidia-Beating Inference Chip in Nine Months with Broadcom's Help
OpenAI announced its first homegrown inference chip, codenamed Jalapeño, which can perform up to 1.9 times more AI work per watt than comparable Nvidia systems. The chip went from initial design to tape-out in just nine months, a significantly faster timeline than typical chip development. Broadcom Inc. handled Jalapeño's physical implementation and supplied Tomahawk networking, while Celestica managed board and system integration, and TSMC manufactured the chip. OpenAI designed the architecture and used its own models to accelerate development. In InferenceX results, Jalapeño delivered 1.5-1.9x more work per watt and 1.7-3.6x lower latency across three large open models compared to Nvidia GB200 and GB300 systems. However, the chip is an inference accelerator, not a training chip, and was not compared with Nvidia's Vera Rubin. OpenAI and Broadcom plan to deploy 10 gigawatts of OpenAI-designed accelerators through 2029. The article notes that while this poses strategic risks to Nvidia's premium GPU market, OpenAI still relies on Nvidia hardware for training.
Show 3 older updatesHide older updates
Jim Cramer Sees No Real Competitors to Nvidia Despite OpenAI's Fast New Chip
OpenAI announced its first custom inference chip, named Jalapeño, developed with Broadcom, claiming it delivers 1.5x to 1.9x more AI work per watt and 1.7x to 3.6x lower latency than Nvidia's GB200 and GB300 chips on benchmarks. OpenAI plans to deploy Jalapeño by end of 2026 and ramp production in 2027, but says it will continue using Nvidia and other partners for training. CEO Sam Altman touted the chip's speed on X. However, CNBC's Jim Cramer dismissed the threat to Nvidia's dominance, stating he reads about superior chips daily but sees no real competitors. The article highlights the ongoing tension between Nvidia's market leadership and emerging custom AI chips.
OpenAI's Jalapeño AI Chip Threatens Nvidia's Margins as Custom Silicon Gains Ground
OpenAI announced its first custom AI chip, Jalapeño, which beat Nvidia's Blackwell systems on key inference-efficiency tests, according to benchmarks. Analysts told CNBC this poses a threat to Nvidia's margins in the fast-growing inference market and reduces OpenAI's reliance on Nvidia for some workloads. The chip, developed with Broadcom, uses newer HBM4 memory and is designed for inference. While Nvidia still dominates AI compute and has ecosystem lock-in via CUDA, custom ASIC chips from OpenAI, Google, AWS, and Meta are gaining ground. SemiAnalysis noted the comparison to Blackwell is somewhat unfair as Jalapeño uses HBM4, making Nvidia's upcoming Rubin platform a more appropriate comparison. OpenAI plans to deploy Jalapeño by end of 2026 and is already working on next-generation versions.
OpenAI’s 700W Jalapeño ASIC outpaces 1,400W Nvidia flagship GPU
Tom's Hardware reports that OpenAI, in collaboration with Broadcom, has developed a custom ASIC chip codenamed 'Jalapeño'. The chip consumes 700 watts of power but claims to outperform Nvidia's flagship GPU, which consumes 1,400 watts. OpenAI asserts that the Jalapeño ASIC delivers up to 1.9 times the throughput per kilowatt and achieves 3.6 times lower latency compared to the Nvidia competitor. This development signals a major push by OpenAI to reduce energy costs and improve inference efficiency for its AI models, potentially challenging Nvidia's dominance in the AI hardware market.