Quick answer: Jalapeño is OpenAI’s first custom-built AI chip, designed specifically to run ChatGPT and other AI models faster and cheaper than the Nvidia GPUs OpenAI currently depends on. It was announced June 24, 2026, built in partnership with semiconductor giant Broadcom, and completed from design to manufacturing in just nine months, the fastest custom chip development cycle in semiconductor history. Full deployment begins at end of 2026, scaling through 2027 and into 2028.
The simplest way to understand why it matters: every time you type a message into ChatGPT and wait for a response, that response costs OpenAI money, hardware power, electricity, and server time. Jalapeño is designed to make that cost dramatically lower, so OpenAI can serve more users at less expense, potentially charge less over time, and stop being so dependent on Nvidia chips it can’t fully control. It’s a chip built for one specific job: generating AI responses, billions of times a day, as efficiently as possible.
Here’s everything you need to know, without the jargon.
Why OpenAI Built Its Own Chip at All
To understand why Jalapeño exists, you need to understand OpenAI’s core infrastructure problem.
Since ChatGPT launched in 2022, OpenAI has run everything, every conversation, every API call, every Codex suggestion on Nvidia GPUs. These are powerful, flexible processors, and they’re excellent for training AI models. But they’re built to handle a huge range of tasks: 3D graphics, scientific simulations, video games, and yes, AI. That flexibility is also their weakness when it comes to one specific job: answering AI queries at scale.
When someone sends a message to ChatGPT, the GPU doesn’t just run AI inference. It routes that task through computing and memory systems designed to handle all sorts of other jobs too, wasting power and processing headroom on capabilities it doesn’t need in that moment. Multiply that inefficiency across billions of daily requests, and OpenAI ends up paying an enormous, never-ending bill for hardware that’s significantly underused for its specific purpose.
There’s a second problem: control. When your entire business runs on someone else’s hardware, hardware they design, price, and decide how to allocate, you’re permanently at their mercy. If Nvidia changes its pricing, limits supply, or prioritizes other customers, OpenAI feels it immediately. Building their own chip doesn’t just cut costs; it gives OpenAI ownership over a critical piece of their own infrastructure.
What Jalapeño Actually Is
Jalapeño is an ASIC, an Application-Specific Integrated Circuit. Unlike a GPU, which can handle almost any computing task you throw at it, an ASIC is designed to do exactly one thing. Jalapeño’s one thing is LLM inference: the process of generating responses from a large language model, one word (technically one “token”) at a time.
By stripping out everything a general-purpose chip carries and doesn’t need for this job, Jalapeño can pack more useful computing power onto the same amount of silicon, run cooler, and use less electricity per response generated.
Here are the key technical facts confirmed from OpenAI’s announcement:
- Built on a 3-nanometer process — the same cutting-edge chip fabrication used in the latest iPhones and top Nvidia GPUs, manufactured through TSMC
- Developed in nine months from initial design to manufacturing tape-out, a record for custom ASIC development, a process that typically takes 18–24 months
- Co-developed by three companies: OpenAI designed the architecture; Broadcom handled silicon engineering and contributed its Tomahawk Ultra Ethernet networking silicon; Celestica managed board, rack, and system integration
- Already running production workloads in testing, including GPT‑5.3-Codex-Spark, at target frequency and power
- Claims roughly half the inference cost of a GPU according to Broadcom CEO Hock Tan — though this is an internal benchmark, not an independently verified figure
- Part of OpenAI’s 10 gigawatt infrastructure roadmap through 2029 — a massive long-term commitment to custom silicon at scale
The physical design pairs a large compute section with six stacks of high-bandwidth memory. That matters because the biggest bottleneck in AI inference isn’t raw computing power, it’s moving data (the model’s weights and recent conversation context) fast enough to feed the compute. Jalapeño is built around solving that specific data movement problem.
The Nine-Month Record: Why That Timeline Is Remarkable
Traditional ASIC development takes 18 to 24 months from design to tape-out. OpenAI and Broadcom did it in nine. OpenAI President Greg Brockman told CNBC that their own AI models compressed the timeline in ways that genuinely surprised the team, using AI to accelerate the design of AI hardware is exactly the kind of recursive progress that makes this announcement more significant than a typical chip launch.
For context: Google started building its own AI chips roughly a decade ago and only this year released Ironwood, its seventh generation of TPU, the first one designed specifically for inference rather than training. Custom chip development is normally a long, generational program. OpenAI moving from zero to a production chip in nine months is a legitimately unusual achievement.
Training vs. Inference: The Difference That Makes Jalapeño Necessary
Most people who’ve heard of AI chips think about training, the massive, months-long process of teaching a model on enormous datasets. That’s what most of the famous Nvidia GPU clusters are used for. Training is a huge upfront cost, but it’s a one-time expense per model.
Inference is the cost that never stops. Every single response ChatGPT generates is an inference job. Every API call from a developer. Every Codex suggestion. At OpenAI’s scale, that means hundreds of billions of tokens generated per day, around the clock, forever. Inference is where the money bleeds out continuously, and it’s exactly what Jalapeño is built to address.
This is why the chip matters even though OpenAI isn’t selling it to anyone. If Jalapeño cuts inference costs meaningfully, ChatGPT becomes significantly cheaper to operate and that saving eventually flows to pricing, capacity, and what OpenAI can afford to build next.
What Jalapeño Gives Up
A chip designed for one job is excellent at that job and limited everywhere else. Here’s what Jalapeño can’t do, and why that matters:
It can’t train AI models. Jalapeño handles inference only. OpenAI still needs Nvidia GPUs or equivalent for training new models. The custom chip supplements the GPU fleet; it doesn’t replace it.
It may struggle with evolving AI architectures. Jalapeño is likely optimized for the transformer-based architecture used in today’s GPT models. If OpenAI’s next generation of models changes significantly as they’ve already done with the shift toward “reasoning” models like o3 the chip may need redesigning to remain fully efficient.
It requires its own software ecosystem. Nvidia GPUs run on CUDA, a mature programming ecosystem with years of tools and libraries built around it. Jalapeño requires custom compilers, runtimes, and kernel libraries written specifically for it. That adds software complexity even as it reduces hardware costs.
Nobody outside OpenAI can use it. This is not a product. There’s no rental market, no API, no way to access Jalapeño’s capabilities. It serves OpenAI’s own infrastructure, period similar to how Amazon’s Trainium chips, Google’s TPUs, and Microsoft’s Maia 200 are all internal-only silicon.
Who Actually Wins From This — And It’s Not Who You Think
The most perceptive take on Jalapeño’s announcement comes from the investment angle: the company whose name is on the chip rarely captures most of the value. Broadcom does.
Google, Meta, and OpenAI all build their custom chips using Broadcom as the design and manufacturing partner. That makes Broadcom something like a toll booth for the custom AI chip economy, it doesn’t matter which AI company wins the model race, because they all pay Broadcom to build their silicon. In the first quarter of fiscal 2026, Broadcom reported $8.4 billion in AI chip revenue, up 106% year over year, and its order book carries a $73 billion backlog in committed AI chip orders roughly half the company’s total backlog.
The chip race that looks like competition between AI companies is, at the infrastructure layer, very good news for the companies that build the hardware underneath all of them.
The Manufacturing Bottleneck Nobody Is Talking About
There’s a quiet constraint on all of this that affects Jalapeño and every custom chip announced in 2026: TSMC’s advanced packaging capacity is sold out through the end of this year.
Jalapeño, like every cutting-edge AI chip, depends on TSMC not just for chip fabrication but for the specialized packaging that bonds compute and memory into a single working part. That packaging capacity is the scarcest resource in the entire AI hardware supply chain, and demand across the industry OpenAI, Google, Apple, Microsoft, Meta, Nvidia runs far ahead of what TSMC can produce. OpenAI can’t skip the line just because they designed their own chip. They compete for finite allocation alongside every major technology company on the planet.
This is why Broadcom CEO Hock Tan described late 2026 as “small prototype development,” with full production ramping in 2027 and “full tilt” output in the first half of 2028. The timeline isn’t set by engineering, it’s set by the physical capacity of the world’s most advanced chip factory.
What This Means for ChatGPT Users
The most direct impact on everyday users is potential price and speed improvement over the medium term. If Jalapeño delivers meaningful inference cost savings at scale, OpenAI has more room to lower API prices for developers and potentially reduce subscription costs for consumers. The chip also means OpenAI can serve more simultaneous users without proportionally growing its hardware budget, which should eventually translate to faster responses and fewer capacity constraints during peak usage.
None of this happens immediately. Jalapeño deploys in limited prototype form at end of 2026 and scales over the following two years. But the direction of travel for ChatGPT users is better performance at lower cost and Jalapeño is a core piece of how OpenAI gets there.
Frequently Asked Questions
What is the OpenAI Jalapeño chip in simple terms? It’s a custom processor OpenAI built specifically to run AI models like ChatGPT. Instead of using general-purpose Nvidia GPUs for everything, Jalapeño is designed from the ground up to do one thing generate AI responses as efficiently as possible. Think of it like the difference between a Swiss Army knife and a chef’s knife: both can cut, but the chef’s knife does that one job far better.
When will the Jalapeño chip be available? Jalapeño is not a consumer or developer product, it’s internal OpenAI infrastructure. Engineering samples are already running test workloads. Initial prototype deployment begins at the end of 2026, with production ramping through 2027 and full-scale deployment in early 2028.
Does Jalapeño mean OpenAI is done with Nvidia? No. Jalapeño handles inference only. OpenAI still needs GPUs for training new models, and even on the inference side, Jalapeño will supplement the existing GPU fleet rather than replace it outright. The custom chip is the beginning of a long-term shift, not an overnight switch.
How fast was Jalapeño developed? Nine months from initial design to manufacturing tape-out. The previous best for ASIC development at this complexity was roughly 18–24 months. OpenAI says their own AI models helped accelerate the design process, an unusual case of AI being used to build better AI hardware.
Is Jalapeño better than Nvidia’s best chips? For the specific task of running large language model inference, Jalapeño claims roughly half the cost per inference compared to current GPU options. That’s an internally measured figure, not an independent benchmark, so treat it as a direction rather than a precise number. For training, flexibility, and the full range of tasks GPUs handle, Nvidia’s chips remain far ahead.
What does Jalapeño mean for Nvidia stock? The broader question of Nvidia’s dominance is worth watching carefully. OpenAI, Google, Meta, and Amazon all have custom chip programs now, which reduces their long-term dependence on Nvidia. In the near term, however, demand for Nvidia GPUs remains at record levels and training needs aren’t going anywhere. Custom inference chips are a long-term pressure on Nvidia’s market position, not an immediate disruption.
Will Jalapeño make ChatGPT cheaper? Potentially, over time. Cheaper inference at scale gives OpenAI more flexibility on pricing. But Jalapeño doesn’t reach full production until 2027–2028, so any pricing impact for consumers is a medium-term story, not something to expect in the next few months.
Who built Jalapeño? Three companies worked on it together: OpenAI led the architectural design, drawing on its direct experience running billions of daily inference requests. Broadcom handled the silicon engineering and manufacturing partnership, and contributed its Tomahawk Ultra Ethernet networking silicon for large-scale deployment. Celestica contributed board, rack, and system integration expertise to industrialize the design for data center use.
The Bottom Line
Jalapeño is the most significant infrastructure move OpenAI has made since ChatGPT launched. It’s not a product you’ll buy or a service you’ll access directly, but it’s the thing that makes everything OpenAI offers cheaper to run, faster to scale, and less dependent on hardware decisions made by companies OpenAI doesn’t control.
For everyday ChatGPT users, the eventual benefit is better, faster responses at lower cost. For the AI industry, it signals that the era of every AI company depending entirely on Nvidia is ending, replaced by a world where the largest labs build their own silicon tuned precisely to how their models actually behave. And for Broadcom, which sits beneath all of it as the company that designs and manufactures these custom chips for Google, Meta, and now OpenAI alike, it’s another confirmation that the most durable position in the AI chip economy isn’t owning the model, it’s owning the road every model runs on.
