OpenAI Jalapeño Turns Custom Inference Silicon Into Nvidia Rack Competition

As of August 29, 2026, OpenAI’s Jalapeño is no longer just a custom-chip ambition sitting somewhere on an infrastructure slide. On August 25, OpenAI published its first measured benchmark results for the inference processor, positioning it as a purpose-built accelerator for serving large language models rather than a general-purpose GPU replacement. The company tested Jalapeño with SemiAnalysis’s InferenceX, a benchmark focused on end-to-end AI request serving, which is the part of the stack users actually feel when a model responds. OpenAI says the chip is its first custom inference processor and part of a multi-generation platform co-developed with Broadcom, with Celestica involved in board, rack, and system integration. (openai.com)

Jalapeño is aimed at high-volume inference: the repetitive, latency-sensitive work of turning prompts into tokens across products such as chatbots, coding agents, APIs, and other interactive AI services. That matters because training grabs the headlines, but inference is where the daily operating cost piles up. OpenAI describes Jalapeño as a blank-slate design for modern LLM inference, tuned around kernels, memory movement, networking, and serving patterns rather than adapted from older accelerator assumptions. Publicly confirmed hardware details remain limited, so this is not a full spec-sheet moment. What OpenAI has disclosed is that the benchmark comparisons used a 700 W package TDP for Jalapeño, against 1,200 W for GB200 and 1,400 W for GB300 in the listed comparisons. (openai.com)

View NVIDIA DGX Spark Personal AI Desktop Supercomputer on partner website

The headline-grabbing part is the Nvidia matchup, but the useful reading is narrower than saying Jalapeño replaces GPUs. In OpenAI’s InferenceX results, Jalapeño showed higher peak mixed tokens per second per kilowatt and lower latency in selected tests against Nvidia GB200 or GB300 configurations, depending on the model. For GPT-OSS 120B, OpenAI lists about 1.9x higher peak mixed TPS per kW and about 1.7x lower end-to-end latency versus GB200. For DeepSeek R1 670B, it lists about 1.7x higher peak mixed TPS per kW and about 3.6x lower latency versus GB300. For Kimi K2.5 1T, the figures are about 1.5x higher peak mixed TPS per kW and about 3.4x lower latency versus GB300. These are OpenAI-published benchmark claims, not independent production fleet data, but they are still notable because they measure complete request serving rather than only raw math throughput. (openai.com)

See ASUS GeForce RTX 5090 price

The more realistic takeaway is that Jalapeño gives OpenAI another lever in its compute portfolio. GPUs remain highly flexible, especially for training, rapid model experimentation, mixed workloads, and the broader CUDA software ecosystem. OpenAI itself says it will continue to widely deploy accelerators from Nvidia and other partners for both training and inference. That makes Jalapeño less of a one-chip takeover and more of a specialized lane: put the custom ASIC where request volume, latency, and power efficiency dominate, while keeping GPUs and other accelerators for workloads that need flexibility or peak capability. In other words, the collision course is about economics at scale, not a clean substitution of one platform for every job. (openai.com)

Buy Lenovo Legion Pro 7i laptop with Core Ultra 9 and NVIDIA RTX 5090 24GB here

Jalapeño also shows where the AI hardware market is heading. Large AI labs are no longer only buying compute; they are shaping more of the stack beneath their models, from chips and memory systems to networking, scheduling, and deployment. OpenAI says Jalapeño is planned for deployment inside its compute infrastructure by the end of 2026, while future generations are already in development. If those deployments match the early benchmark story, the impact could be felt in faster responses, better capacity during demand spikes, and improved serving cost for large open and proprietary models. For Nvidia, the point is not that its rack systems suddenly become irrelevant. The point is that major AI customers now have stronger reasons to mix merchant GPUs with tightly optimized custom silicon when inference becomes large enough to justify the effort. (openai.com)

See SANDISK Optimus GX PRO 8100 PCIe 5 SSD price