OpenAI debuts Jalapeño, its first in-house inference chip
OpenAI has revealed Jalapeño, a custom inference processor built with Broadcom, in its boldest move yet to design its own silicon. What it means for India's chip ambitions.
The News
OpenAI on Wednesday revealed its first custom processor, a chip named Jalapeño that it designed in partnership with Broadcom. The company describes it as an inference processor, meaning it is purpose-built to run already-trained models rather than to train new ones from scratch.
The announcement formalises a collaboration the two firms first disclosed in October 2025. According to OpenAI, Jalapeño has been tuned for real-time coding workloads and, in early testing, delivers what the company calls significantly better performance-per-watt than current state-of-the-art alternatives. The chip is still in a testing phase, and OpenAI did not publish a node size, transistor count, production timeline or manufacturing location.
Greg Brockman, OpenAI's president, framed the project as a consequence of the company knowing its own software better than anyone. "We have a deep understanding of the workload," he said, asking "how can we build something that will accelerate what's possible?" OpenAI added that its own models assisted in the chip's development, part of a broader push to optimise its full stack across architecture, kernels, memory, networking and deployment.
Why It Matters
Jalapeño is less a product launch than a statement of intent. By designing inference silicon in-house, OpenAI is following a path Google and Amazon already cut, with the TPU and Trainium families respectively, to cut costs and loosen the grip of a single supplier. OpenAI was careful to note that pre-training is still expected to lean on Nvidia hardware, so this is a targeted move rather than a wholesale divorce.
The logic is economic as much as technical. Inference is the recurring cost of running a model in production, billed every time a user prompts ChatGPT or fires off a Codex agent. Shaving watts off each query compounds across billions of requests. The last time a frontier lab decided general-purpose GPUs were too blunt an instrument, Google built the TPU and ended up running much of its search and ads infrastructure on it. OpenAI is betting the same calculus now applies to generative workloads at scale.
For Broadcom, the deal cements its role as the go-to design partner for hyperscalers that want bespoke accelerators without becoming chip companies themselves. That positioning, rather than any single customer, is what has made the firm one of the defining names of this compute cycle.
Indian Angle
The Jalapeño story lands squarely in the middle of India's own silicon moment. Under the India Semiconductor Mission, the country has approved fabrication and assembly plants from Micron in Sanand and a Tata-PSMC fab in Dholera, while positioning itself as a design and packaging hub. A marquee fabless chip designed by Broadcom is a useful reference point for Indian policymakers arguing that value sits as much in design as in manufacturing.
That matters because India is already deep inside this supply chain through talent rather than fabs. Broadcom runs large engineering centres in Bengaluru and Hyderabad, and a meaningful share of global semiconductor design work is done by Indian engineers. A custom-silicon arms race among hyperscalers is, in practice, a hiring race for exactly the chip-design skills India produces in volume.
There is a cost angle for Indian builders too. Domestic model startups such as Sarvam and Krutrim still depend on imported Nvidia GPUs priced in dollars, a painful equation when the rupee is weak. If inference-optimised chips like Jalapeño push down the cost-per-token of running models, the savings could eventually reach Indian developers renting capacity from global clouds, even if the chips themselves never ship to India. For MeitY, the lesson is that owning inference economics, not just data, is becoming a strategic question.
FAQ
What exactly is an inference chip?
Inference is the stage where a trained model answers a query, such as generating code or replying in a chat. An inference chip is optimised for that repetitive, latency-sensitive work, as opposed to training chips that crunch enormous datasets to build the model in the first place. Jalapeño targets the former.
Does this mean OpenAI is dropping Nvidia?
No. OpenAI said pre-training is still expected to rely on Nvidia hardware. Jalapeño is aimed at inference, so it complements rather than replaces Nvidia GPUs for now. The move trims dependence on a single vendor without abandoning it.
When will Jalapeño ship?
OpenAI has not given a production timeline. The chip is described as still being in testing, and the company has not disclosed a node size, manufacturer beyond Broadcom, or volume schedule.
How does this affect Indian AI startups?
Not directly today, since the chip is for OpenAI's own infrastructure. Over time, cheaper inference could lower the cost of cloud-hosted models that Indian developers rent, easing a dollar-denominated bill that hits hard when the rupee is weak.
Where can I read the original announcement?
The news was reported by TechCrunch, linked in the attribution below, which carries OpenAI's framing and the quote from Greg Brockman.
This story was reported by TechCrunch. Read the full original coverage at TechCrunch.