**
OpenAI has officially entered the chip business. In partnership with Broadcom, the company unveiled “Jalapeño” — its first custom AI inference processor — marking the beginning of the end of Nvidia’s GPU monopoly in AI infrastructure.**
Announced on June 24, 2026, Jalapeño is a purpose-built ASIC architected from the ground up for LLM inference. With a claimed ~50% reduction in inference cost per token and deployment planned for gigawatt-scale data centers by end of 2026, this is OpenAI’s declaration of hardware independence.
Bloomberg Television reports on OpenAI and Broadcom’s custom Jalapeño inference chip and its potential to cut AI inference costs by ~50%.
The Hardware Marvel: A 3nm ASIC in Record Time
Jalapeño is not a modified GPU. It’s a blank-slate ASIC built on TSMC’s 3nm process — what Tom’s Hardware calls a “massive reticle-sized ASIC” pushing the absolute limits of lithography for maximum compute density.
The design philosophy centers on reducing data movement — balancing compute, memory, and networking to hit real-world utilization “much closer to theoretical peak” than general-purpose GPUs. That’s a direct challenge to Nvidia’s architecture, where memory bottlenecks often leave performance on the table.
Why Jalapeño Matters: The De-Nvidification of AI
Consider the numbers. OpenAI’s operating loss hit $20.92 billion in FY2025 on $13.07B revenue. Its single biggest cost? Compute — over $10.59 billion paid to Microsoft alone. Jalapeño is a survival play dressed as a technology story.
As Broadcom CEO Hock Tan put it: “The chip is as good as Blackwell or Google TPU.” If that holds in production, OpenAI can roughly halve its largest expense center — reshaping the economics of the entire AI industry.
Jalapeño doesn’t replace Nvidia overnight. All vendor deals remain in place — the $30B Nvidia investment, $50B AWS Trainium commitment, and signed AMD Instinct MI450 agreements. Nvidia keeps its training stronghold. But inference is where the volume lives, and that’s exactly the battle OpenAI is picking.
This mirrors the hyperscaler playbook — Google (TPU), Amazon (Trainium), Meta (MTIA). Even Microsoft built its own Majorana 2 quantum chip. The difference: OpenAI is doing it as an independent company, not a trillion-dollar giant, sending a powerful signal about its IPO ambitions.
The 9-Month Debate: Genuine or Framing?
The most contested claim is the 9-month design timeline. On Hacker News, chip veterans expressed deep skepticism — pointing out that “9 months” could mean RTL freeze to tape-out (standard) rather than concept to tape-out (unprecedented).
The “AI-accelerated design” claim also drew fire. Skeptics called it “like saying development was accelerated by Microsoft Office.” But there’s substance here: AI tools excel at generating HDL testbenches for chip verification. OpenAI’s models likely contributed meaningfully, even if the broader framing oversells it.
One detail cuts through the noise: Richard Ho, OpenAI’s Head of Hardware, is a former Google TPU lead and ex-Lightmatter SVP. This isn’t OpenAI’s “first chip design” — it’s Ho’s umpteenth. That expertise explains how a software company moved so fast in silicon.
The Competitive Ripple
Jalapeño sends shockwaves through an $80 billion AI chip market. Nvidia still dominates training, but with GPU margins around 75%, a 50% cheaper inference alternative is margin compression by another name. Vera Rubin arrives in 2027, giving Nvidia a narrow window to respond.
Broadcom emerges as the silicon kingmaker. Already powering Google’s TPU and Meta’s MTIA, it now adds OpenAI to a roster that makes it the most important chip design house in AI — stock up 18% YTD and roughly 7x since end of 2022.
Groq raised $650M on the same day as the announcement, pivoting to a “neocloud” model. Chinese competitors are accelerating too — Alibaba’s Zhenwu M890, Huawei’s Ascend 950DT due in August, and ByteDance negotiating with Qualcomm for custom ASICs. The race for custom inference silicon is now global.
OpenAI President Greg Brockman and Broadcom CEO Hock Tan discuss the Jalapeño chip’s development, performance claims, and strategic vision in an exclusive CNBC interview.
The Recursive Advantage
Here’s the deeply meta angle: OpenAI used its own models to help design Jalapeño. AI helped build the hardware that will run future AI. This recursive loop (AI → chip → faster AI → better chip) creates a compounding advantage that general-purpose GPUs can’t replicate.
It’s the same principle behind AI harness engineering — where AI systems optimize the infrastructure that runs AI. When model and chip are co-designed, you unlock efficiencies generic hardware simply can’t match.
As Greg Brockman put it: “Jalapeño is part of our long-term full-stack infrastructure strategy to make compute more abundant, resulting in AI which is faster, more reliable, more affordable.” That vertical integration — from silicon to model to deployment — looks increasingly like Apple’s strategy, applied to AI.
FAQ: OpenAI’s Jalapeño Chip
{
"@context": "https://schema.org",
"@type": "FAQPage",
"mainEntity": [
{
"@type": "Question",
"name": "What is the OpenAI Jalapeño chip?",
"acceptedAnswer": {
"@type": "Answer",
"text": "Jalapeño is OpenAI's first custom AI inference processor, co-designed with Broadcom. It's a purpose-built ASIC on TSMC 3nm, designed specifically for LLM inference workloads."
}
},
{
"@type": "Question",
"name": "How much cheaper is Jalapeño compared to Nvidia GPUs?",
"acceptedAnswer": {
"@type": "Answer",
"text": "Broadcom CEO Hock Tan claims approximately 50% lower inference cost per token versus Nvidia GPUs, with substantially better performance per watt."
}
},
{
"@type": "Question",
"name": "When will Jalapeño be deployed?",
"acceptedAnswer": {
"@type": "Answer",
"text": "OpenAI plans deployment in gigawatt-scale data centers with Microsoft by end of 2026. Engineering samples are already running in labs."
}
},
{
"@type": "Question",
"name": "Does Jalapeño replace Nvidia for OpenAI?",
"acceptedAnswer": {
"@type": "Answer",
"text": "No. All existing deals with Nvidia ($30B), AWS ($50B), and AMD remain in place. Jalapeño is additive — targeting inference while Nvidia retains training dominance."
}
},
{
"@type": "Question",
"name": "How was the chip designed in 9 months?",
"acceptedAnswer": {
"@type": "Answer",
"text": "The timeline is contested — it may refer to RTL freeze to tape-out rather than concept to tape-out. Key factors include Broadcom's silicon expertise, Richard Ho's TPU background, and AI-assisted verification tooling."
}
}
]
}
Looking Ahead
Jalapeño is Gen 1 of a multi-generational roadmap. With Gen 2+ already in planning, OpenAI signals a long-term commitment to custom silicon. The biggest question isn’t whether the chip works — samples are already running GPT-5.3-Codex-Spark at target specs. It’s whether OpenAI can scale fast enough to meaningfully dent its $20.9B operating loss while pushing frontier models forward. The gigawatt-scale data center infrastructure needed is a challenge that extends far beyond the chip itself.
One thing is certain: the era of AI companies outsourcing all compute to GPU vendors is ending. The de-Nvidification of AI has officially begun.
Photo: OpenAI CEO Sam Altman and Broadcom CEO Hock Tan with the Jalapeño Intelligence Platform wafer. Source: ServeTheHome / OpenAI.
Sources: OpenAI Official Blog, Broadcom Investor Release, TechCrunch, Tom’s Hardware, Hacker News Discussion
Originally published on TekMag
