OpenAI Jalapeño Chip: 50% Cheaper Inference Targets NVIDIA

OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom AI chip designed for inference, claiming a 50% cost reduction compared to NVIDIA GPUs. Manufactured by TSMC on a 3nm process, the ASIC was developed in nine months using OpenAI's own AI models and targets commercial deployment in late 2026, starting with Microsoft data centers. The chip aims to lower inference costs, which represent the majority of AI compute spending, while OpenAI continues to rely on NVIDIA for training.

OpenAI and Broadcom unveiled Jalapeño today https://techcrunch.com/2026/06/24/openai-unveils-its-first-custom-chip-built-by-broadcom/ — the company’s first custom AI chip, a purpose-built inference processor that claims to cut inference costs by roughly 50% compared to current NVIDIA GPUs. Designed in just nine months using OpenAI’s own AI models to accelerate the process, the ASIC will be manufactured by TSMC at its 3nm process node and is already being tested with GPT-5.3-Codex-Spark in OpenAI’s own facilities. Initial commercial deployment targets late 2026, with Microsoft data centers first in line. What Jalapeño Actually Is and Isn’t Jalapeño is an ASIC — Application-Specific Integrated Circuit — designed exclusively for inference. Not training. That distinction matters more than the chip name. Richard Ho, who leads OpenAI’s hardware initiative, put it plainly: “Jalapeño was designed from the ground up for LLM inference using detailed insights from our close collaboration with OpenAI researchers.” The chip optimizes around the specific memory movement, kernel execution, and networking patterns that transformer-based models demand at scale. The nine-month design-to-tape-out timeline is the headline engineering claim. Broadcom and OpenAI say this is the fastest ASIC development cycle ever achieved for high-performance advanced semiconductors — a timeline compressed in part by using OpenAI’s own models to accelerate chip design decisions. According to The Next Web’s analysis https://thenextweb.com/news/openai-jalapeno-chip-broadcom-nvidia , the chip is manufactured on TSMC’s 3nm node with Celestica handling server infrastructure manufacturing. OpenAI’s president Greg Brockman framed Jalapeño as part of a “full-stack infrastructure strategy to make compute more abundant.” Related: NVIDIA Grove: Open-Source Kubernetes API for AI Inference Why AI Inference Costs Are the Real Fight Training gets the attention, but inference is where the money bleeds. Every ChatGPT response, every OpenAI API call, every agentic loop executing a coding task burns GPU hours at NVIDIA’s market rates. Inference now represents the majority of ongoing AI compute spend — and for a company at OpenAI’s scale, shaving 50% off that cost changes the fundamental unit economics of the business. However, the “50%” figure deserves scrutiny. Broadcom CEO Hock Tan told Bloomberg that cost savings amount to “roughly 50% reduction versus conventional AI GPUs.” OpenAI’s own language is more conservative — “substantially better performance per watt than current state-of-the-art alternatives.” A full technical report is promised “within months.” Treat the 50% as an early-lab target, not a production guarantee. Volume production doesn’t ramp until 2027, and the company’s 10-gigawatt compute target with Jalapeño isn’t due until 2029. OpenAI Isn’t Leaving NVIDIA — Training Still Needs GPUs Jalapeño is inference-only. OpenAI’s training runs remain entirely dependent on NVIDIA hardware, and that’s unlikely to change in the near term. This is a diversification play at the inference layer, not NVIDIA’s obituary. NVIDIA’s inference market share currently sits around 90%, and analysts project it could fall to 20-30% by 2028 as custom silicon scales — but that projection assumes flawless, simultaneous deployment across multiple competing chips. OpenAI joins every major hyperscaler in the custom inference chip race. According to Tom’s Hardware’s custom ASIC analysis https://www.tomshardware.com/tech-industry/semiconductors/custom-ai-asics-examined-from-broadcom-to-mtia , Google’s TPU v6e claims 40% total cost of ownership reduction, Amazon’s Trainium3 targets 30-50% savings, and Microsoft’s Maia 200 aims at inference at Azure scale. Broadcom is the common thread — it’s providing ASIC design capability and Ethernet networking IP across multiple chip programs simultaneously. Furthermore, NVIDIA’s moat isn’t just hardware: decades of CUDA tooling and developer familiarity represent switching costs that custom silicon doesn’t automatically erase. Consequently, stock markets barely blinked — NVIDIA dropped 0.26% on the announcement day while Broadcom gained 2%. Related: Neoclouds Will Own 20% of the $267B AI Cloud Market by 2030 Key Takeaways - OpenAI’s Jalapeño is an inference-only ASIC built with Broadcom on TSMC’s 3nm node — it does not replace NVIDIA for training workloads - The claimed 50% inference cost reduction comes from early internal testing via Broadcom’s CEO; a full technical report is pending and volume production doesn’t start until 2027 - First commercial deployment targets late 2026 at Microsoft data centers — don’t expect API price cuts this year - Broadcom is emerging as the custom chip kingmaker, designing ASICs for multiple AI labs simultaneously, making it the picks-and-shovels winner in the silicon arms race - NVIDIA stock barely moved on the announcement — the market agrees this is diversification, not disruption, at least for now