In short
OpenAI has revealed Jalapeño, its first custom AI chip built with Broadcom to run inference for ChatGPT and future models. The move is part of a broader push to improve efficiency and reduce dependence on Nvidia.
- Jalapeño is OpenAI’s first custom chip and is designed for AI inference, not model training.
- The processor was developed with Broadcom and is expected to be deployed by the end of 2026.
- OpenAI says early testing shows stronger performance per watt than current state-of-the-art chips.
- The move reflects a wider industry trend toward custom silicon as AI compute demands surge.
OpenAI has taken a major step toward controlling more of the hardware that powers its products, unveiling its first custom chip for AI servers in partnership with Broadcom. The processor, named Jalapeño, is built for a single, increasingly important job: running AI inference at scale, the process that turns a user’s prompt into a ChatGPT reply or executes an action inside tools such as Codex.
The new chip marks an important shift for the company behind ChatGPT. Instead of relying entirely on external suppliers for the processors that keep its services running, OpenAI is now moving into the chip-design business. The company says Jalapeño is only the beginning of a broader, multi-generation compute strategy, with deployment expected by the end of 2026.
OpenAI’s announcement places it squarely in the fast-moving race to build custom AI hardware, where companies including Google, Amazon, Meta and Microsoft are all trying to reduce dependence on Nvidia while tailoring chips more closely to the needs of their own models and services.
Why OpenAI is building its own chip
For years, the AI industry has leaned heavily on Nvidia’s graphics processors to train and run large models. But as demand has surged, access to enough chips has become one of the biggest bottlenecks in the industry. For a company like OpenAI, which serves massive volumes of ChatGPT traffic and increasingly sophisticated agentic workflows, that dependence creates both cost pressure and supply risk.
Jalapeño is OpenAI’s answer to part of that problem. Because it is designed as an application-specific integrated circuit, or ASIC, it is optimized for a narrow set of tasks rather than broad-purpose computing. In this case, the target is inference: the phase in which a model applies what it has learned to generate text, answer questions or complete tasks in real time.
That distinction matters. Training large models is computationally demanding, but inference is what keeps products like ChatGPT responsive to millions of users every day. It is also where efficiency gains can translate into meaningful savings, especially when a service scales globally.
Inference versus training
OpenAI’s chip is not intended to replace every kind of accelerator used in its infrastructure. Instead, it is aimed at the workloads that make up a growing share of ongoing AI costs. Training teaches a model. Inference serves it.
That difference is central to the strategic logic behind custom silicon. A chip optimized for inference can be tuned for latency, throughput and power efficiency, making it a better fit for production traffic than a more general-purpose accelerator.
- Training uses large data sets to build and refine a model.
- Inference handles live user requests and model outputs.
- ASICs are built for specific workloads, which can improve performance per watt.
What Jalapeño is designed to do
OpenAI says Jalapeño is intended to support both current and future large language models, signaling that the chip is being planned with the company’s expanding product portfolio in mind. That includes not only chatbot interactions but also more complex tool use and agent workflows, where a system may need to decide, reason, and act across multiple steps.
The company describes the chip as its first move in a broader multi-generation compute platform. In practical terms, that suggests OpenAI is not treating Jalapeño as a one-off experiment. Instead, it appears to be laying the groundwork for a family of chips and infrastructure designs that could evolve alongside its models.
OpenAI characterized Jalapeño as the starting point for a multi-generation compute platform and said early tests indicate notably stronger performance per watt than current leading solutions.
OpenAI has not yet disclosed final benchmark numbers, and that detail matters. Early claims in the chip market often need to be viewed cautiously until real-world deployments confirm them. Even so, the company’s language indicates that it believes Jalapeño could offer a meaningful efficiency advantage over existing options.
The Broadcom partnership and what it signals
The chip was developed with Broadcom, one of the most important names in networking and custom silicon. Broadcom has become a key partner for companies seeking alternatives to off-the-shelf AI accelerators, bringing manufacturing know-how and design experience to some of the largest infrastructure projects in the sector.
That collaboration is significant for two reasons. First, it shows OpenAI is not trying to build a chip operation from scratch alone. Second, it reflects a broader industry shift toward co-designing hardware and software so the underlying silicon is tailored to a company’s exact workload.
In a Reuters interview, Broadcom chief executive Hock Tan said the new chip matches the performance of Nvidia’s Blackwell processors and Google’s Tensor Processing Units. Those are bold claims, and they place Jalapeño in direct comparison with some of the most advanced AI hardware in the market.
At the same time, the competitive landscape remains nuanced. Even as companies build custom chips, Nvidia continues to hold the lead in overall performance, ecosystem maturity and software support. For many AI companies, custom silicon is less about immediately beating Nvidia everywhere and more about achieving better economics for specific, high-volume workloads.
A wider chip race across Big Tech
OpenAI is not entering this arena alone. Over the past several years, major technology companies have increasingly turned inward to design their own AI chips, with different goals ranging from training frontier models to powering search, recommendation systems and cloud infrastructure.
Google has long used its own Tensor Processing Units for internal workloads. Amazon has developed chips for both training and inference. Meta and Microsoft have also pursued custom hardware efforts to manage costs, improve efficiency and control supply.
This wave of in-house or co-designed processors reflects a simple reality: as AI usage explodes, the economics of renting or buying conventional accelerators can become unsustainable. The companies with the largest AI businesses have strong incentives to optimize the silicon stack around their own workloads.
| Company | Custom chip strategy | Primary use case | Strategic goal |
|---|---|---|---|
| OpenAI | Jalapeño with Broadcom | Inference for ChatGPT and future models | Reduce reliance on Nvidia and improve efficiency |
| Tensor Processing Units | Training and inference | Optimize internal AI services | |
| Amazon | Custom AI chips | Training and inference | Lower cloud costs and expand infrastructure control |
| Meta | Custom silicon efforts | Inference and internal workloads | Support large-scale consumer AI systems |
| Microsoft | In-house AI chips | Cloud and model workloads | Improve supply resilience and cost structure |
Why Nvidia still matters
Custom chips may be spreading, but Nvidia remains the gravitational center of the AI hardware market. Its chips still lead in broad performance, and more importantly, Nvidia has spent years building a software ecosystem that makes its hardware easier to deploy at scale.
That matters because AI infrastructure is not just about raw speed. It also depends on tooling, developer support, compatibility and operational maturity. In many cases, custom chips are best understood as supplements to Nvidia hardware rather than instant replacements.
For OpenAI, Jalapeño likely serves as a pressure valve: a way to redirect some high-volume inference traffic onto a chip designed specifically for its needs. If successful, that could ease supply constraints and improve operating margins over time.
Performance per watt is the key metric
One of the most important numbers in AI infrastructure is not just how fast a chip runs, but how much power it consumes while doing it. Electricity and cooling costs have become central to the economics of AI deployment, especially as data centers expand.
OpenAI’s emphasis on performance per watt suggests that power efficiency may be the chip’s biggest selling point. If Jalapeño can do more work for every unit of energy consumed, it could help lower the cost of serving millions of prompts a day.
- Lower power use can reduce operating costs.
- Better efficiency can ease data center thermal constraints.
- Higher throughput can improve response times for users.
What this means for ChatGPT and future agents
Jalapeño’s initial job is tied to the infrastructure behind ChatGPT, but the chip could have broader implications as OpenAI leans more heavily into agentic AI products. Systems that browse, code, call tools and chain multiple model steps together can generate substantial inference load.
As AI products become more interactive and more autonomous, the infrastructure burden grows quickly. That makes efficient inference hardware more important than ever. A custom chip tuned for OpenAI’s own workload patterns could support more predictable latency and lower costs for these next-generation applications.
There is also a strategic dimension. Owning more of the compute stack gives OpenAI more control over how future products are built and scaled. In a market where compute access can determine the pace of product rollout, that kind of vertical integration is increasingly valuable.
The timeline so far
OpenAI’s chip effort has moved quickly from announcement to reveal. The company first said it would work with Broadcom on custom silicon nine months ago. Now, it has provided a name, a technical category and a rough deployment target.
That pace is notable, though the remaining steps still matter. Final performance validation, manufacturing readiness and large-scale integration into OpenAI’s infrastructure will determine whether Jalapeño becomes a meaningful part of the company’s compute stack or remains an early-stage milestone.
| Milestone | Details |
|---|---|
| Initial Broadcom deal | OpenAI announced plans to co-develop custom chips roughly nine months before the reveal |
| Processor reveal | OpenAI introduced Jalapeño as its first AI processor for inference |
| Performance testing | OpenAI says early testing shows improved performance per watt |
| Expected deployment | End of 2026 |
Industry implications beyond OpenAI
OpenAI’s move will likely intensify pressure across the AI supply chain. If more major model developers conclude that custom inference chips are viable, demand for specialized design services, advanced packaging, memory and networking components could rise further.
It may also sharpen competition between the companies that provide AI infrastructure. Cloud platforms, chip designers and model developers are increasingly overlapping in the same strategic territory, each trying to control more of the stack.
That convergence has several implications:
- Model developers may seek greater independence from scarce external chips.
- Chipmakers will need to prove they can offer both flexibility and performance.
- Cloud providers may face more pressure to adapt to custom hardware ecosystems.
- Customers may eventually benefit from lower costs and faster services if efficiency gains are passed through.
Open questions remain
Even with the announcement, many of the most important details remain undisclosed. OpenAI has not publicly shared the chip’s exact specifications, manufacturing process, memory architecture or deployment footprint. It also has not provided an independent benchmark against leading Nvidia hardware in real production conditions.
That leaves several questions for the months ahead. How many Jalapeño units will OpenAI deploy? Which services will use them first? How much of the company’s inference traffic can realistically be shifted onto custom hardware? And how quickly can the chip move from promising tests to meaningful operational impact?
The answers will determine whether Jalapeño is merely a symbolic first step or the start of a more consequential re-architecture of OpenAI’s infrastructure.
A strategic bet on control, efficiency and scale
For OpenAI, the logic behind Jalapeño is straightforward. If the company can lower its dependence on external suppliers, improve energy efficiency and build a chip stack more closely matched to its own models, it may gain important advantages in cost, reliability and speed.
That does not mean the path will be easy. Building custom silicon is expensive, technically demanding and slow by software standards. But in AI, where infrastructure has become a competitive moat, the investment may be worth it.
OpenAI’s first chip does not end the Nvidia era. But it does show that the companies driving the AI boom are no longer willing to leave their most important workloads entirely in someone else’s hands.
The company’s broader message is clear: to keep scaling ChatGPT and the next wave of AI agents, OpenAI wants more control over the hardware beneath the models.
What to watch next
Over the next year, the key indicators will be whether OpenAI expands Jalapeño from a prototype or early deployment into a meaningful share of its inference infrastructure. Watch for details on manufacturing, server integration, and whether OpenAI follows this first chip with additional designs tailored to other workloads.
If the project succeeds, it could become a template for other AI companies trying to cut costs and improve performance. If it struggles, it will still stand as evidence of how far the industry’s appetite for compute has pushed major players to redesign the hardware layer itself.
