Dave Jimenez
OpenAI's first custom silicon, a chip called Jalapeño, went public on Wednesday through a joint announcement with Broadcom (AVGO). Developed in just nine months and designed entirely by OpenAI, the chip targets inference workloads and, according to early testing data from the companies, outperforms current state of the art processors, a direct challenge to Nvidia's (NVDA) dominant position.
At a Glance
- Jalapeño is OpenAI's first custom inference chip, co-developed with Broadcom.
- Development took nine months from design to early testing.
- Early results show performance above current leading chips, per OpenAI and Broadcom.
- The chip is the first in a planned multi-generation compute platform, with rollout beginning later this year.
- Broadcom shares rose more than 1% on the news.
Nine Months From Design to Announcement
The speed of development is the detail worth examining first. Nine months is an unusually compressed timeline for a custom application specific integrated circuit built to handle the computational demands of large language model inference. OpenAI says Jalapeño is purpose built for running its own models as well as models developed across the broader industry, suggesting the company intends the chip to function in a general inference context rather than as a narrowly proprietary accelerator.
OpenAI president Greg Brockman framed the chip as part of what he called a long term full stack infrastructure strategy. The goal, in his words, is to make compute more abundant, driving AI that is faster, more reliable, and more affordable for both consumers and enterprises. Designing more of the stack internally, he argued, allows the company to deliver more intelligence per unit of efficiency.
That framing is notable. OpenAI is positioning Jalapeño not as a hedge against Nvidia supply constraints, though that is plainly part of the calculus, but as a structural component of its infrastructure roadmap. The multi-generation platform language implies sustained investment in custom silicon rather than a one-off experiment.
The Supply Constraint Problem Jalapeño Is Meant to Solve
OpenAI is among the largest individual buyers of Nvidia GPUs globally. That creates an inherent tension: the same chips the company relies on are competed for by every other major AI lab, cloud provider, and enterprise AI team. Custom silicon offers a path to compute capacity that bypasses that queue.
Owning the inference layer of the stack also enables tighter optimization. A chip designed specifically around OpenAI's model architectures can, in principle, extract better throughput per watt and per dollar than a general purpose GPU tuned for a wider workload range. That efficiency gap compounds at the scale OpenAI operates.
How OpenAI's Move Fits the Broader Custom Silicon Trend
OpenAI is not the first hyperscaler or major AI platform to pursue this path. The table below shows the current landscape of custom AI chip programs among the largest technology companies.
| Company | Custom AI Chip Program | External Sales |
|---|---|---|
| Google (GOOG) | Tensor Processing Units (TPUs) | Yes, via Google Cloud |
| Amazon (AMZN) | Trainium and Inferentia | Yes, via AWS |
| Microsoft (MSFT) | Maia AI accelerator | Internal and Azure |
| Meta | MTIA inference chip | No (internal), cloud entry floated |
| OpenAI | Jalapeño (Broadcom) | TBD, rollout later 2025 |
Amazon and Google have moved furthest toward commercializing their custom silicon, renting capacity to third party customers through their cloud platforms. Meta has floated the idea of becoming a cloud computing provider, a move that would put it in direct competition with Nvidia. OpenAI has not yet indicated whether Jalapeño capacity will be offered externally.
On the competitive side, AMD (AMD) continues pressing its case in the AI data center market, targeting workloads where Nvidia's pricing and availability create openings. Qualcomm and Cerebras are also positioning their respective architectures for inference and edge AI use cases, adding further pressure below Nvidia's flagship products.
What the Broadcom Partnership Signals
Broadcom's role here is that of a manufacturing and design partner rather than a co-brand. OpenAI handled the architecture; Broadcom provided the fabrication and engineering infrastructure to realize it. This mirrors the model Google used when developing its TPUs with Broadcom, and it reflects Broadcom's strategy of building long term chip supply relationships with the largest AI spenders. The more than 1% move in AVGO shares on the announcement is modest but confirms the market reads the deal as incrementally positive for Broadcom's custom ASIC revenue line.
Frequently Asked Questions
What does Jalapeño actually do?
Jalapeño is an inference chip, meaning it is designed to run trained AI models rather than train them from scratch. OpenAI says it is optimized for its own models and for industry models more broadly, suggesting a general inference positioning rather than a single model application.
Does this mean OpenAI will stop buying Nvidia chips?
No announcement to that effect has been made. Custom silicon typically supplements rather than replaces Nvidia GPUs in the near term, particularly for training workloads where Nvidia retains a large technical lead. Jalapeño appears focused on the inference side of the compute stack.
When will Jalapeño be available?
OpenAI says the chip is the first in a multi-generation compute platform set to begin rolling out later in 2025 and continuing in subsequent years. Specific availability dates or capacity figures have not been disclosed.
How does this affect Nvidia?
Every major hyperscaler and AI platform developing custom silicon represents potential demand that bypasses Nvidia's supply chain. OpenAI is one of Nvidia's largest customers, so its move into proprietary inference hardware is a material long term signal, even if near term GPU purchases remain substantial.
What the Jalapeño Launch Tells Us About the AI Chip Market in 2025
The nine month development cycle and the multi-generation roadmap language together suggest OpenAI is treating silicon as a core competency rather than a procurement problem. With Google, Amazon, Microsoft, and Meta all running parallel programs, the inference chip market is fragmenting rapidly. Nvidia retains dominance in training and high end inference today, but the cumulative weight of these custom programs, each optimized for its owner's specific workloads, will keep pressure on that position through the rest of the decade.
Carlos Dixon
