When AI Builds Its Own Chip: Why OpenAI’s Jalapeño Hints at a Cheaper Future
TLDR
OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first custom AI chip — a purpose-built inference processor designed from scratch for running large language models. The headline is the timeline: it went from design to manufacturing tape-out in nine months, possibly the fastest high-performance ASIC cycle ever, partly because OpenAI used its own models to design the chip. Strategically, it’s a move to control its supply chain, cut dependence on Nvidia, and attack the single biggest cost in its business: inference. And that’s the part worth paying attention to — when a company controls the full stack down to the silicon, the cost of running models can fall, and that pressure eventually reaches the rest of us through cheaper API pricing.
In November 2025, I was in Atlanta to take part in a hackathon themed around Artificial Intelligence. During a coffee break, I tossed a question to the people sharing my table: won’t AI get more expensive over time? After all, these systems demand more and more processing power, and demand simply outstrips supply by a wide margin. The answer surprised me — they were emphatic that they expected AI to get much cheaper.
Well, it seems we now have the answer, or at least part of it.
On June 24, 2026, OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first custom-designed AI chip. On the surface it’s an engineering announcement. Underneath, it’s the clearest signal yet of why those people at the coffee table were so confident — and it has everything to do with who controls the production chain.
To understand why this matters, you have to separate two phases in an AI model’s life. Training is the massive, one-time job of building a model. Inference is the constant job of running it every time someone sends a prompt — every ChatGPT message, every API call, every line your coding assistant completes. Jalapeño is built only for inference. OpenAI calls it an “Intelligence Processor,” and it’s an ASIC — an application-specific chip designed from a blank slate around exactly how OpenAI’s models behave, rather than a general-purpose GPU adapted to the task. By stripping it down to inference and architecting it around its own workloads, OpenAI says the chip cuts the wasteful data movement that bottlenecks conventional GPUs and runs much closer to a chip’s theoretical peak performance. Early engineering samples are already running one of OpenAI’s own models, GPT-5.3-Codex-Spark, in the lab, with what OpenAI claims is performance-per-watt substantially better than current state-of-the-art hardware.
But the number that made the whole industry sit up wasn’t a benchmark — it was the calendar. A custom chip like this normally takes eighteen months to two years to design from scratch. OpenAI and Broadcom did it in nine months. OpenAI President Greg Brockman told CNBC the chip was designed end to end in that window and credited the company’s own AI models with the speed, saying the degree to which the models accelerated the work surprised even them. Read that loop carefully: OpenAI used AI to help design the chip that will run its AI. The same flywheel Anthropic describes when Claude writes most of Anthropic’s code now shows up in silicon. AI is compressing the development of the very hardware it depends on.
Here’s where the parallel I want to draw comes into focus — and it’s the heart of why AI may get cheaper rather than more expensive.
For years, the cost structure of running AI has been dictated largely by one vendor. OpenAI has been one of Nvidia’s single largest customers since the boom kicked off in 2022, and that dependence is both expensive and risky: you pay Nvidia’s premium margins, and you inherit its supply constraints. When a single supplier sits at the center of your most important recurring cost, that supplier sets the price floor for everyone downstream. This is the classic problem of not controlling your production chain — your unit economics are written by someone else’s pricing power.
Jalapeño is OpenAI’s move to take that pen back. By designing its own inference chip, OpenAI can target the cost of running its models directly — the single biggest recurring expense in its business — instead of routing all of that spend and supply risk through one vendor. This is vertical integration in its purest form: own the model, own the software, own the data center, and now own the silicon underneath it. And it’s not just OpenAI. It’s the defining strategic move of 2026. Microsoft built its own MAI models, SpaceX runs its own compute through Colossus, Anthropic uses Claude to build Claude, and now OpenAI designs its own chips. The companies that control the whole stack are building moats that pure-play rivals can’t dig — and, more importantly for the rest of us, they’re building the ability to push their own costs down.
That cost reduction is the bridge to cheaper APIs. Think of the chain of consequences. When you control your own silicon, you can optimize it for exactly your workload and squeeze out efficiency a general-purpose chip can’t reach. When AI helps you design that silicon, you can iterate faster than the industry’s traditional two-year hardware cadence — meaning each generation can be cheaper and faster than the last, on a much tighter clock. When inference gets cheaper at the source, the company can lower the price it charges to serve a token. And because OpenAI doesn’t operate in a vacuum, that pressure cascades: as one player drives down the cost of running models, the entire industry races to match its unit economics or lose customers. The era of inference being dominated by one expensive vendor is starting to crack — and that competition is precisely what flows through to the API pricing that developers, startups, and businesses pay.
Get Alessandro Silveira’s stories in your inbox
Join Medium for free to get updates from this writer.
There’s an additional twist that reinforces this. OpenAI is heading toward a public offering, and it’s currently losing money on the sheer cost of compute. If Jalapeño meaningfully lowers inference costs, that’s a direct path to better unit economics and eventually profitability — exactly the story public-market investors want to see. In other words, OpenAI has a powerful structural incentive to make inference cheap, because cheap inference is what makes the business viable at scale. The interests of the company and the interests of API consumers point, for once, in roughly the same direction.
Now, the honest caveats, because this isn’t a clean victory lap. Jalapeño is inference-only — the heavy lifting of training next-generation models will still lean on Nvidia, AMD, and others. OpenAI even took a $30 billion investment from Nvidia earlier this year as part of a deal to deploy gigawatts of systems, so this is diversification and leverage, not a divorce. It’s a strike at Nvidia’s pricing power, not at Nvidia’s existence. And nothing ships at scale until late 2026, with Microsoft reportedly committing to a large share of the first production run — so the real economic impact is a 2027-and-beyond story, not something that shows up on your invoice next quarter. There’s also a counter-current worth naming: Broadcom’s CEO described AI compute demand as “insatiable” through 2028. If demand keeps outrunning even a faster, cheaper supply, prices could stay sticky for a while regardless of how good the chips get. Cheaper per token doesn’t automatically mean cheaper bills if everyone simply uses far more tokens.
So the people at my coffee table in Atlanta were right, but with an asterisk. The cost of intelligence — what it takes to generate a given amount of AI output — is on a clear downward path, and Jalapeño is one of the most concrete pieces of evidence yet. Vertical integration over the production chain is the mechanism, AI-accelerated chip design is the multiplier, and competitive pressure is the transmission belt that carries those savings to the APIs we build on. Whether your monthly spend actually shrinks depends on how much more you decide to do with intelligence that keeps getting cheaper per unit. If history is any guide, we’ll spend the savings on doing more — which, for anyone whose product runs on tokens, is a pretty good problem to have. The quiet engine under all of this is a flywheel: AI designing the hardware that runs more AI, faster and cheaper, so it can build better AI. Nine months is the first time we’ve watched that crank turn on physical silicon. It won’t be the last.
Sources
- OpenAI and Broadcom unveil LLM-optimized inference chip — OpenAI
- OpenAI and Broadcom Unveil LLM-Optimized Intelligence Processor — Broadcom (press release)
- OpenAI and Broadcom reveal Jalapeño, first AI chip in partnership — CNBC
- Broadcom and OpenAI unveil custom-built Jalapeño inference processor — Tom’s Hardware
- OpenAI unveils its first custom chip, built by Broadcom — TechCrunch
- Jalapeño is the first AI chip from OpenAI and Broadcom — Engadget
- OpenAI REVEALS AI Chip — World of AI (original brief)
Tags
#OpenAI #Jalapeño #Broadcom #AIChips #Inference #ASIC #Nvidia #VerticalIntegration #AICost #APIPricing #LLM #SemiconductorIndustry #AIInfrastructure #ArtificialIntelligence
