OpenAI had thus far stayed in the AI models space, but it’s now taking steps into moving into the hardware direction.

The company has unveiled Jalapeño — its first custom AI chip — built in partnership with Broadcom. Designed specifically for LLM inference, the chip was conceived from scratch by OpenAI’s engineering teams and taken from design to manufacturing tape-out in just nine months, a timeline that Broadcom describes as potentially the fastest ASIC development cycle ever achieved in high-performance semiconductors. Celestica is also part of the picture, handling board, rack, and system integration to bring the platform to production.

Jalapeño is purpose-built around what OpenAI actually runs: ChatGPT, Codex, its API, and the growing range of agentic products the company is building toward. Rather than adapting a general-purpose accelerator to fit LLM workloads, OpenAI designed the chip around those workloads from the ground up. Engineering samples are already running GPT-5.3-Codex-Spark in the lab at production-target frequency and power. Early results suggest meaningfully better performance per watt than current state-of-the-art hardware, though OpenAI says a detailed technical report will follow in the coming months.

The architecture focuses on reducing data movement and better balancing compute, memory, and networking resources — the kind of tradeoffs that matter enormously at inference scale, where the bottleneck is often not raw compute but how efficiently data flows through the system.

The Infrastructure Play

For a company that has spent years buying Nvidia GPUs at scale, this is a meaningful shift. Inference — the process of serving a model’s response to a user query — represents a large and growing share of the compute bill for any company running at ChatGPT’s volume. A chip optimized specifically for that job, and tuned to OpenAI’s own models, gives the company a lever it previously lacked: the ability to improve inference economics without waiting for Nvidia’s next product cycle.

Broadcom CEO Hock Tan, who delivered a physical sample of the chip to Sam Altman and Greg Brockman, has long made this argument publicly — that companies serious about leading in AI need their own silicon. OpenAI’s Jalapeño is the company acting on that logic. Broadcom’s role isn’t just manufacturing; its Tomahawk networking silicon is woven into the platform, and the company is a critical part of how the chip scales to gigawatt-level deployments planned with Microsoft and other data center partners starting later in 2026.

Jalapeño is explicitly the first step in a multi-generation roadmap. The companies are targeting 10 gigawatts of compute powered by OpenAI-designed accelerators with Microsoft through 2029, which is the kind of commitment that signals this isn’t a one-off experiment.

Where the Rest of the Industry Stands

Google has been doing this for years. Its Tensor Processing Units (TPUs) — purpose-built accelerators that bypass general-purpose GPU architecture — have been in production since 2016. TPUs power much of Google’s AI infrastructure across Search, Translate, and DeepMind workloads. The advantage Google gained from designing silicon that matches its software is precisely what OpenAI is now trying to replicate.

Amazon has Trainium and Inferentia. Meta has MTIA. Microsoft has Maia. OpenAI, until now, was the notable exception among the major AI players — leaning heavily on Nvidia while the rest of the hyperscaler field was building inward.

Anthropic sits in a different position entirely. The company has no proprietary chip program and currently trains Claude across AWS Trainium, Google TPUs, and Nvidia GPUs. It has signed a deal with Google and Broadcom for next-generation TPU capacity, which means it’s still dependent on external silicon. Interestingly, Clive Chan — one of the earliest engineers on OpenAI’s chip team — recently left to join Anthropic, a sign that the company is accumulating hardware talent, but it’s a long road from hiring chip engineers to shipping your own silicon.

What This Means for Nvidia

OpenAI has relied on Nvidia almost exclusively since the beginning, and Jalapeño doesn’t change that overnight. Training workloads, in particular, still run on Nvidia hardware, and OpenAI has acknowledged it’s only exploring whether to extend its custom chip program into training. For now, Jalapeño is an inference play.

The harder question Jalapeño raises for Nvidia isn’t whether it loses OpenAI’s business outright. It’s whether the largest AI operators — Google, Amazon, Microsoft, Meta, and now OpenAI — continue accepting one default architecture for every job. Once a company has its own inference silicon tuned to its own models, the conversation about what to buy from Nvidia becomes more deliberate. That’s a different dynamic than simply waiting in line for the next GPU allocation.

Broadcom, for its part, emerges from this announcement with a stronger position. The company has quietly become the workshop of choice for hyperscalers building custom AI silicon — and landing OpenAI, arguably the highest-profile name in the space, cements that role.

OpenAI’s full-stack ambitions have been visible for a while. Jalapeño is the infrastructure layer becoming real.