The OpenAI and Broadcom (AVGO) custom AI chip called Jalapeño marks OpenAI's first foray into proprietary silicon, designed specifically for inference workloads across both its own models and third party AI systems. Built in just nine months, the chip reportedly outperforms current state-of-the-art processors in early testing, a direct challenge to Nvidia (NVDA).
At a Glance
- Chip name: Jalapeño, designed by OpenAI, manufactured in partnership with Broadcom
- Development timeline: nine months from design to completion
- Primary use case: AI inference, compatible with OpenAI models and third party systems
- Rollout: begins later in 2025 and extends into subsequent years as a multi-generation platform
- Market reaction: Broadcom shares climbed more than 1% on the announcement
What Jalapeño Is and Why OpenAI Built It
Inference is the operational heartbeat of any deployed AI system. It is the compute process that runs a trained model to generate outputs, whether that means answering a prompt, classifying an image, or generating code. Unlike training, which is a one-time capital expenditure, inference is continuous and scales directly with user demand. Building a chip optimized for exactly that workload, rather than relying on general-purpose processors, is how you compress cost and latency at scale.
OpenAI president Greg Brockman framed the chip as part of a deliberate full-stack infrastructure strategy. The goal, in his words, is to make compute more abundant so that AI becomes faster, more reliable, and more affordable for both individuals and businesses. Controlling more of the underlying stack also means OpenAI can tune its systems precisely to its own model architectures rather than working around the constraints of hardware it does not design.
The Nvidia Problem OpenAI Is Trying to Solve
OpenAI is one of the largest single buyers of Nvidia's high-powered processors. That relationship is not going away overnight, but it creates a structural vulnerability. Nvidia's chips are in extraordinary demand across the entire AI industry, which means OpenAI competes with every other AI company, hyperscaler, and enterprise deploying large models just to secure supply. Delivery timelines and allocation constraints are real operational risks at this scale.
Developing proprietary silicon through Broadcom gives OpenAI a parallel supply channel. Even if Jalapeño does not replace Nvidia hardware across the board, it reduces dependency and gives OpenAI negotiating leverage it currently lacks. The nine-month development window is notable here: it signals that OpenAI has built internal chip design capability fast enough to iterate meaningfully, rather than treating this as a one-off prototype.
Broadcom's role is significant as well. The company has become the go-to partner for hyperscalers designing application-specific integrated circuits, with Google's Tensor Processing Units among the most prominent examples. Broadcom's stock rising more than 1% on the announcement reflects the market's read that this relationship will deepen and generate meaningful revenue.
How the Competition Maps Out
OpenAI is not pioneering custom silicon so much as catching up to a pattern already well established among the largest technology companies. Amazon, Google, and Microsoft each have custom AI processors either in production or in active development. Amazon's Trainium and Inferentia chips are already rented to third-party customers through AWS. Google's TPUs, available through Google Cloud, have been running inference at scale for years. Meta designs and deploys its own chips for AI and other workloads internally and has floated the idea of offering cloud computing services, which would put it in direct competition with Nvidia.
On the competitive chip side, AMD is pushing hard into AI data center hardware. Qualcomm and Cerebras are each pursuing different angles on the inference market. None of these challengers has displaced Nvidia's dominance, but collectively they are narrowing the moat. Jalapeño adds one more credible alternative to that list, and the fact that OpenAI claims it beats current state-of-the-art chips in early testing will pressure Nvidia to accelerate its own roadmap disclosures.
| Company | Custom AI Chip | Primary Use | Third-Party Access |
|---|---|---|---|
| OpenAI / Broadcom | Jalapeño | Inference (OpenAI and industry models) | Not announced |
| TPU (Tensor Processing Unit) | Training and inference | Yes, via Google Cloud | |
| Amazon | Trainium / Inferentia | Training and inference | Yes, via AWS |
| Microsoft | Azure Maia | Training and inference | Internal / Azure |
| Meta | MTIA | Inference and ranking workloads | No (internal) |
| AMD | Instinct MI series | Training and inference | Yes, via cloud partners |
Who This Chip Is Actually For
The positioning is broader than OpenAI's own product stack. Brockman's statement explicitly mentions compatibility with AI models across the industry, not just OpenAI's own. That phrasing suggests OpenAI may eventually offer Jalapeño-based compute to external customers, similar to the model Amazon and Google have adopted. If that happens, OpenAI's ambitions shift from AI lab to infrastructure provider, a significant expansion of its business surface.
For now, the immediate beneficiary is OpenAI's own inference stack. Running ChatGPT and the API at scale is expensive, and any efficiency gain at the chip level flows directly into gross margin. Given the cost pressures that come with serving hundreds of millions of users, even modest per-query savings compound into material numbers at OpenAI's volume.
Broadcom benefits from a high-profile design win that validates its position as the preferred chip partner for companies that want custom silicon without building their own fabrication capabilities. Nvidia faces a longer-term signal that its most important customers are actively investing in alternatives, even if those alternatives are complementary rather than purely substitutional today.
Frequently Asked Questions
What does Jalapeño do differently from Nvidia's chips?
Jalapeño is purpose-built for inference, the process of running trained AI models to produce outputs, rather than for the broader range of compute tasks Nvidia's GPUs handle. OpenAI says it outperforms current state-of-the-art chips in early testing, though independent benchmarks have not yet been published.
Will Jalapeño replace Nvidia hardware at OpenAI?
OpenAI has not indicated it plans to phase out Nvidia processors. The chip is described as part of a multi-generation platform strategy, suggesting it will run alongside existing Nvidia hardware rather than replacing it outright in the near term.
Could OpenAI offer Jalapeño-based compute to other companies?
OpenAI has not formally announced external cloud access to the chip, but the stated goal of designing it for use with AI models across the industry leaves that door open. Amazon and Google have already moved in exactly that direction with their own custom chips.
What is Broadcom's role in the chip?
Broadcom is the manufacturing and design partner. OpenAI designed the chip, and Broadcom provided the engineering and production infrastructure to bring it to silicon. Broadcom has played a similar role for Google's TPUs and other hyperscaler custom chips.
Where the Custom Silicon Race Goes From Here
Jalapeño is described as the first chip in a multi-generation computing platform, with broader rollout beginning later in 2025. That roadmap language matters: it means OpenAI is committing to iterating on this hardware over time, not treating it as a one-off experiment. For Nvidia, the more concerning data point is not any single chip but the compounding effect of every major AI company building its own silicon. Supply diversification across the industry structurally limits Nvidia's pricing power over time, even if no single alternative matches its current performance ceiling.
