OpenAI Just Built Its Own AI Chip in 9 Months — And Claims It Matches Nvidia Blackwell

Nine months. That's how long it took OpenAI to go from a napkin sketch to a working AI chip. Let that sink in for a second.

In an industry where semiconductor development typically takes two to four years — sometimes longer — OpenAI and Broadcom just unveiled Jalapeño, a custom-built application-specific integrated circuit (ASIC) designed exclusively for running large language models. And according to Broadcom CEO Hock Tan, it performs just as well as Nvidia's Blackwell chips and Google's tensor processing units.

Yeah. I had to read that twice too.

The announcement dropped on June 24, 2026, and it signals something much bigger than just another chip launch. OpenAI — the company behind ChatGPT, GPT-5, and Codex — is officially a hardware company now. Or at least, it's trying very hard to become one. And that shifts the entire power dynamic in the AI industry.

What Exactly Is Jalapeño?

Jalapeño isn't some repurposed graphics card or an off-the-shelf processor retrofitted for AI workloads. It's a purpose-built chip — an ASIC, for those keeping score — architected from scratch for one specific job: AI inference.

For anyone who's been following the chip wars but doesn't live inside a semiconductor fab: inference is the part where the model actually answers your question. When you type something into ChatGPT, or when Codex generates code for you, that's inference happening on the server side. Training is the other half — feeding the model enormous datasets so it learns patterns. But inference is what happens every single time someone actually uses these models. And that's where OpenAI is bleeding money.

The scale of inference demand is hard to overstate. ChatGPT serves hundreds of millions of queries daily. Each one burns compute. Nvidia GPUs handle most of this today, but they're expensive, they're power-hungry, and — as anyone who's tried to buy H100s knows — they're notoriously hard to get your hands on.

Jalapeño is OpenAI's answer to all three problems.

The Nine-Month Miracle (Or Is It?)

Here's where things get really interesting. OpenAI claims that Jalapeño went from initial design to manufacturing tape-out in roughly nine months. According to OpenAI's official announcement, this is "the fastest ASIC development cycle ever achieved in high-performance advanced semiconductors."

Nine months. In the chip world, that's almost obscene. Traditional high-performance chip development involves years of architecture work, verification, physical design, and testing before you even send anything to a fab. Broadcom's involvement obviously helped — they brought decades of silicon implementation experience and their Tomahawk networking technology to the table. But even with that expertise, the timeline is staggering.

How did they do it? Part of the answer is recursive in a way that feels very 2026: OpenAI used its own AI models to accelerate specific aspects of the chip design process. The company employed software-hardware co-development techniques, letting AI tools handle optimization tasks that would typically take human engineers weeks or months.

Richard Ho, who leads OpenAI's hardware program, said the chip was "designed from the ground up for LLM inference using detailed insights from our close collaboration with the model and product teams." Translation: they didn't just build a fast chip. They built a chip that was custom-tailored to the exact computational patterns their models use.

Now — and I want to be honest here — "fastest ever" claims make me twitch. Without independent verification or a detailed technical report (which OpenAI says is coming "in the coming months"), we're essentially taking their word for it. Early testing shows the engineering samples running ML workloads in the lab at "production target frequency and power," including workloads from GPT-5.3-Codex-Spark. That's promising. But lab results aren't the same as production deployment at gigawatt scale.

Why This Matters for the AI Industry

If you've been reading our coverage of the AI chip space — and you should, because it's where the real power shifts are happening — you'll recognize a pattern. We covered how Groq raised $650 million to keep fighting in the inference chip space even after Nvidia poached its CEO. We tracked how SK Hynix overtook Samsung because the memory inside these chips became more valuable than Samsung's entire product lineup.

Jalapeño is another piece of the same puzzle: the AI industry is trying desperately to stop depending on Nvidia for everything.

And OpenAI isn't alone. Microsoft, Meta, and Amazon have all launched custom AI chip efforts. Google has been doing TPUs for years. Anthropic was reportedly weighing building its own chip as recently as April 2026. Everyone wants to own more of the stack.

But OpenAI's move is arguably the most aggressive. Rather than building a general-purpose accelerator that can handle both training and inference, they went all-in on inference. It's a bet that says: the training wars are mostly settled, and the real ongoing cost — the real bottleneck — is serving AI to billions of users.

Greg Brockman, OpenAI's President and Co-Founder, put it bluntly: "The world is moving to a compute-powered economy. 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 for people and businesses."

The Full-Stack Gambit

This is where OpenAI's strategy gets genuinely ambitious. They're not just building a chip. They're building what they call a "multi-generation compute platform" — a complete hardware ecosystem designed to work specifically with their models.

The flywheel OpenAI is describing goes something like this: better infrastructure leads to higher compute efficiency, which enables better training and serving, which produces better models, which attracts more users and revenue, which funds the next generation of infrastructure. Rinse, repeat. Intelligence gets cheaper and more capable over time.

It's a compelling story. It's also the exact same story that Google has been telling about TPUs for a decade. The question is whether OpenAI can actually execute it at scale, especially given that they're essentially a software company learning to play hardware.

The Nvidia Problem

Let's talk about the elephant in the server room. Nvidia is not going to sit still while OpenAI builds alternative silicon. The company's Blackwell architecture is already shipping, and while Broadcom's CEO claims Jalapeño matches it, that's a pretty bold statement from a company that has financial incentive to hype its product. (Broadcom shares are up about 10% this year, by the way.)

There's also the question of memory. As Broadcom's Tan acknowledged to Reuters, AI chips require enormous amounts of high-bandwidth memory, which squeezes margins on custom chip products. The global supply of HBM (high-bandwidth memory) is dominated by SK Hynix and Samsung, and those supplies are already stretched thin. Building the chip is one thing. Building it at the scale OpenAI needs — we're talking gigawatt-scale data centers — is another challenge entirely.

And then there's the uncomfortable reality that OpenAI's own hardware chief admitted: "We're still measuring final performance." Engineering samples running in a lab are a long way from production deployment at scale. We've seen plenty of chip announcements that looked spectacular in benchmarks but couldn't deliver at volume. Intel's various GPU attempts come to mind.

What Happens Next

OpenAI says Jalapeño will begin deployment by the end of 2026 with Microsoft and other partners. That's... soon. Very soon. If they pull it off, it fundamentally changes the economics of running ChatGPT and related services. The inference cost savings alone could be worth billions annually.

But there's a bigger strategic implication. OpenAI is building what amounts to a proprietary hardware platform optimized exclusively for its own AI models. That means if you want the absolute best inference performance for LLMs, you might need to be running on OpenAI's infrastructure. It's a moat-building strategy, and it's one that could reshape how companies think about AI deployment.

The other thing worth watching: OpenAI mentioned that Jalapeño is built "with flexibility to work with all LLMs across the industry." That's a carefully worded statement. It could mean the chip will eventually be available to other companies — maybe even competitors. Or it could just mean the architecture isn't locked to one specific model version. We'll know more when the technical report drops.

For now, though, the headline is simple: OpenAI designed a working AI chip in nine months and says it's on par with the best silicon from Nvidia and Google. The AI chip race just got a new contender. And it's not a chip company.

Frequently Asked Questions

What is OpenAI's Jalapeño chip?

Jalapeño is OpenAI's first custom AI processor — an ASIC (Application-Specific Integrated Circuit) designed specifically for AI inference. It was co-developed with Broadcom and is engineered to handle workloads from large language models like GPT-5, ChatGPT, and Codex. It's not a general-purpose GPU; it's purpose-built to answer user queries efficiently.

How does Jalapeño compare to Nvidia's chips?

According to Broadcom CEO Hock Tan, Jalapeño matches the performance of Nvidia's Blackwell architecture and Google's tensor processing units. OpenAI's own early testing indicates "performance per watt substantially better than current state-of-the-art." However, independent benchmarks haven't been published yet, and OpenAI is still measuring final performance numbers.

When will OpenAI's Jalapeño chip be available?

OpenAI plans to begin deploying Jalapeño by the end of 2026, initially with Microsoft and other data center partners. The chips will be used exclusively by OpenAI for its own services — they're not being sold commercially (at least not yet). This is the first of what OpenAI describes as a "multi-generation compute platform."

Why is OpenAI building its own AI chip?

OpenAI is building custom chips to reduce its dependence on Nvidia GPUs, which are expensive, power-hungry, and difficult to obtain in large quantities. With hundreds of millions of daily ChatGPT queries demanding inference compute, custom silicon offers a path to lower costs and greater supply chain control. Microsoft, Meta, Amazon, and Google have all pursued similar strategies.

Who is manufacturing the Jalapeño chip?

TSMC (Taiwan Semiconductor Manufacturing Company) is manufacturing the Jalapeño chip. Broadcom handled the silicon implementation and provided networking technology. Celestica is responsible for board, rack, and system integration. The entire supply chain relies on advanced semiconductor manufacturing capabilities that currently only a handful of companies possess.

Did OpenAI use AI to design the Jalapeño chip?

Yes, partially. OpenAI used its own AI models to accelerate specific aspects of the chip design process. This software-hardware co-development is one reason the company claims Jalapeño went from initial design to manufacturing tape-out in approximately nine months — far faster than the typical two to four years for high-performance semiconductor development.

Sources

• The Verge — OpenAI reveals its first AI processor: Jalapeño
• OpenAI Official — OpenAI and Broadcom unveil LLM-optimized inference chip
• Reuters — OpenAI unveils custom chip designed with Broadcom

Stay Ahead of the AI Chip Race

The AI infrastructure wars are heating up fast. If you want to stay on top of every major chip announcement, model release, and industry shakeup, bookmark OnlyGFs.ai and check back daily. We cover the stories that actually matter — not the hype, not the press releases, just the facts and what they mean for you.