OpenAI Built Its First Chip, 'Jalapeño,' in Nine Months — AI Helped Design the AI Chip

Chips Usually Take Years — They Did It in Nine Months

Here's the deal: OpenAI and Broadcom unveiled OpenAI's first custom AI chip, "Jalapeño," on June 24. It's purpose-built from scratch to do just one thing — LLM inference, the process of running a trained model in response to user commands. OpenAI calls it its first "Intelligence Processor."

The most stunning part is speed. From concept to manufacturing tape-out in just nine months. High-performance advanced chips usually take 2–3 years from design to near-production. Doing it in nine months is "believed to be the fastest ASIC development cycle ever achieved in high-performance advanced semiconductors." And the secret is the fun part — OpenAI's own models were used to accelerate parts of the chip's design and optimization. AI helped build the AI chip. Sci-fi, but it actually happened.

The chip is inference-only. Not training — it's specialized to run already-built models fast and cheap. OpenAI designed it from a deep understanding of its models' fundamentals, kernels, serving systems, and product needs, with Broadcom and Celestica industrializing the platform. Early testing shows performance per watt "substantially" better than current state-of-the-art, with staged deployment starting at the end of 2026.

Here's what we'll unpack: why OpenAI built its own chip, what nine months means, and what changes for Nvidia, developers, and the whole AI industry. Three players: OpenAI, which designed it; Broadcom, which implemented the silicon; and Nvidia, which the chip targets.

The Players — OpenAI, Broadcom, and Nvidia

First, OpenAI. As ChatGPT demand exploded, inference cost became the company's biggest financial variable. Every model run spins GPUs, and most of those GPUs are Nvidia's. So OpenAI was locked into a structure where rising revenue meant rising payments to Nvidia. An in-house chip is an attempt to escape that dependence and control inference cost with its own hands — and the first concrete output of the Broadcom partnership struck in October 2025.

Next, Broadcom — a top name in custom-chip (ASIC) design and implementation. Google's TPU was built with Broadcom's help, too. OpenAI decides "what to build," Broadcom owns "how to realize it in silicon." This deal cements Broadcom's position as "the custom-chip partner for AI big tech."

Third, Nvidia. It doesn't appear directly, but it's the company most affected. Jalapeño's goal is cost reduction versus Nvidia GPUs. The AI inference market is far larger than training and keeps growing — and Nvidia GPUs have all but monopolized it. If a top customer like OpenAI replaces part of that with its own chip, Nvidia's biggest revenue source could wobble.

One line: OpenAI built its own chip to escape inference cost and Nvidia dependence, and Broadcom stamped it into silicon in nine months. That's the spine.

What's New — By the Numbers

Item	Detail
Chip name	Jalapeño
Purpose	LLM-inference-only ASIC
Dev cycle	Concept→tape-out in 9 months (record)
Design assist	OpenAI's own models accelerated design/optimization
Performance	Perf/watt "substantially" above current best
Deployment	Gigawatt-scale data centers from end of 2026
Partners	Broadcom (silicon), Celestica (industrialization)

Two things stand out. First, "inference-only" is the crux. Training and inference demand different compute. Training prizes flexibility, favoring general-purpose GPUs; inference repeats the same operations at scale, where a specialized chip maximizes efficiency. Because OpenAI knows exactly what operations its models perform, it tuned the chip to them and lifted perf/watt.

Second, the nine-month speed is the real news. Chip design is an enormous human effort of placing and verifying circuits — and OpenAI's models accelerated that EDA process. The implication is big: AI helps design chips, chips come faster, those chips run stronger AI, and that AI designs the next chip — a feedback loop made real. The very speed of semiconductor development could change.

Who Wins

OpenAI wins biggest. Controlling inference cost with its own hands lets it escape the structure where cost rose in lockstep with revenue. Margins improve and it's less whipsawed by Nvidia supply and pricing. Plus "we make our own chips" is a huge plus for both technical self-sufficiency and the valuation narrative — a strong card on the road to an IPO.

Broadcom is a clear winner. As AI big tech marches toward in-house chips, its standing as the partner backing that design and implementation grows. It bags OpenAI as a marquee reference and can pull in other AI companies' custom-chip demand.

The strained party is, of course, Nvidia. But don't oversimplify. Jalapeño is "inference-only," and Nvidia still dominates the training market and general-purpose GPU demand. Also, very few companies can afford their own chips. Still, the symbolism of "your biggest customer might leave" is heavy, and ceding part of the inference market is real pressure.

Precedents — Wins and Misses

We've seen this before. Google's TPU, Amazon's Trainium/Inferentia, Apple's own silicon are the templates. Beyond a certain scale of core compute, big tech concludes "building our own chip is cheaper." Just as Google controlled search/AI costs with TPU, OpenAI walks the same path. The strategy is powerful when sufficient volume is guaranteed.

The key to the win is scale. Custom chips have astronomical development costs, so you must run them enough to recoup. OpenAI has ChatGPT's giant in-house demand, satisfying that condition. The failure risk is just as clear: one stumble in chip development brings yield, power, or software-stack problems that delay production — and in the meantime GPUs advance another generation. "You catch up just to be overtaken" can repeat.

Another lesson is software. Many argue Nvidia's real moat isn't the chip but the software ecosystem — CUDA. Even a fast custom chip won't deliver real-world efficiency if the serving software on top isn't mature. OpenAI co-designing chip and serving software is precisely to dodge that trap.

It helps to separate two different bets bundled in this announcement. One is the obvious one — cheaper inference. The other, quieter bet is strategic optionality: by owning a chip design, OpenAI gains leverage in every future negotiation with Nvidia, with cloud providers, and with capital markets. Even if Jalapeño only ever handles a slice of inference, its existence changes the bargaining table — "we can build our own" is a different position than "we depend entirely on you." Google's TPU arguably delivered as much value as a negotiating chip against Nvidia as it did in raw silicon. OpenAI is buying that same option, and nine months is how fast it bought it.

Rival Counter-Plays

Nvidia's counter is "strengthen inference-specialized products" and "software lock-in." Push inference-optimized GPU lines more aggressively and hold customers with CUDA's convenience. It can also flex pricing and supply to reduce big customers' incentive to build their own — proving "using ours is cheaper than building it."

In-house-chip veterans like Google and Amazon welcome this trend — validation that "custom chips are the right direction." At the same time, OpenAI entering the chip market as a rival makes cloud/infrastructure competition more complex.

Other AI labs (Anthropic, etc.) face a fork: do they have the volume to build their own, or keep using GPUs and third-party chips? OpenAI stamping a chip in nine months suggests the barrier to in-house chips could be lower than thought. If true, more companies gain a rationale to jump in.

So What Changes

If you're a developer — your code doesn't change today. Jalapeño is OpenAI's internal infrastructure, so API users call the OpenAI API the same regardless of the chip. But over the medium term, lower inference cost could feed into API price cuts. GPT-5.6's aggressive pricing (especially Luna at $1/$6) isn't unrelated to this cost-structure shift.

If you're an investor — it signals the AI-chip landscape slowly shifting from "Nvidia monopoly" to "distributed in-house chips." ASIC partners like Broadcom and the inference-specialized semiconductor value chain may draw attention. But weigh, in balance, that Nvidia's moat (software, generality, training market) is still deep.

If you work in AI — the biggest takeaway is that the "AI designs chips" feedback loop is now real. If it accelerates, hardware progress speeds up, which feeds model progress. We're heading into an era where the line between semiconductors and AI keeps blurring.

🥄 Three Things You're Probably Wondering

— Is OpenAI done with Nvidia now? No, it's not that simple. Jalapeño is inference-only; training still needs GPUs. And deployment is staged from year-end, so for a while OpenAI runs Nvidia GPUs and its own chip in parallel. Not full replacement — "moving some inference to its own chip."

— Does nine months mean chips are easy now? Don't misread it. OpenAI had a peerless combo: its own models, massive capital, Broadcom, and Celestica. It doesn't mean a typical company can stamp a chip in nine months. But the "AI accelerates design" methodology is validated, so the ceiling on development speed did rise.

— Is Nvidia stock done? Too early to call. The pressure of ceding some inference is real, but the moats — training market, CUDA ecosystem, generality — are still deep. "A signal the monopoly is wobbling" and "collapse" are entirely different things. Market-structure shifts usually unfold slowly over years.

Sources

• OpenAI and Broadcom unveil LLM-optimized inference chip — OpenAI
• OpenAI and Broadcom reveal Jalapeno, first AI chip in partnership — CNBC
• OpenAI unveils its first custom chip, built by Broadcom — TechCrunch
• OpenAI and Broadcom Unveil LLM-Optimized Intelligence Processor — Broadcom
• Broadcom and OpenAI unveil custom-built Jalapeño inference processor — Tom's Hardware

Numbers and criteria are as of announcement and may change.