OpenAI's first chip Jalapeño takes aim at the cost of inference

The News

OpenAI has unveiled its first piece of custom silicon, a chip named Jalapeño, built in partnership with Broadcom and aimed squarely at the cost of running large language models in production.

Announced on 24 June 2026, Jalapeño is what OpenAI calls its first "Intelligence Processor", an accelerator designed around the company's own view of how inference - the work of generating answers from an already-trained model - should be handled. OpenAI led the design while Broadcom handled the engineering and manufacturing build. The two firms first disclosed their collaboration in October 2025, and OpenAI says it moved from initial design to manufacturing tape-out in roughly nine months.

The companies claim Jalapeño delivers performance per watt "substantially better" than current state-of-the-art alternatives, though final testing is still under way and detailed specifications are due later. Deployment is expected to begin in late 2026. Greg Brockman, OpenAI's president, said the team had "a deep understanding of the workload" and had hunted for tasks that were "underserved" by existing hardware.

Crucially, Jalapeño targets inference rather than training. Pre-training the largest models is likely to stay on Nvidia hardware for now.

Why It Matters

The economics of artificial intelligence have quietly shifted from training to serving. Building a model is a one-off capital expense; answering billions of user queries is a recurring bill that never stops. By designing a chip tuned to its own software, OpenAI is trying to attack that recurring cost and loosen its dependence on Nvidia, whose graphics processors still command the lion's share of the market.

The move echoes an earlier turning point. When Google built its first Tensor Processing Unit last decade, it signalled that the biggest AI buyers would no longer be content renting general-purpose silicon. Amazon followed with its Trainium and Inferentia parts. OpenAI joining that club, alongside its reported 10-gigawatt commitment with Broadcom, is the clearest sign yet that the vertical-integration playbook favoured by the hyperscalers has reached the model labs themselves.

For Nvidia, this is not an immediate threat to revenue, but it is a long-term warning that its largest customers are quietly becoming its rivals.

Indian Angle

For India, the headline is cost. Indian AI startups such as Sarvam and Krutrim, along with thousands of smaller teams, pay for compute in dollars, and inference is the line item that scales with every user they add. If custom silicon pushes down the global price of serving models, the cloud rates Indian developers pay could ease over time, improving the unit economics of building consumer AI for a famously price-sensitive market.

There is a sovereignty thread too. India's semiconductor mission has poured incentives into fabrication plants in Gujarat and Assam, yet the country still has no leading-edge logic fab. Jalapeño is a reminder that the high-value intellectual property sits in chip design, an area where India already supplies deep talent - Broadcom itself runs large design centres in Bengaluru and Hyderabad. Policymakers weighing where to place their bets may note that design, not just assembly, is where the margin lives.

Finally, for Indian enterprises piloting AI, cheaper inference is the variable that decides whether a proof of concept survives contact with the finance team. Lower serving costs widen the set of use cases that can clear an internal return-on-investment hurdle, which matters more here than in markets where compute budgets are looser.

FAQ

When will Jalapeño be available?

OpenAI says deployment begins in late 2026. The chip is still in testing, and full technical specifications have not yet been published. The partnership behind it was first announced in October 2025, so the timeline reflects roughly a year of development before live use.

Does this replace Nvidia for OpenAI?

No. Jalapeño is built for inference, not training. OpenAI is expected to keep using Nvidia hardware to pre-train its largest models, so the new chip complements rather than replaces its existing fleet of graphics processors.

What does it mean for Indian startups?

Indirectly positive. Cheaper inference at the frontier labs tends to filter into lower cloud pricing over time, which helps Indian teams that pay for compute in dollars and run models at scale for cost-conscious users.

Who actually built the chip?

OpenAI led the design while Broadcom handled the engineering and manufacturing. The collaboration was first disclosed in October 2025, and the firms say development took about nine months from design to tape-out.

Where can I read the original announcement?

OpenAI published the news on its official blog, which is linked in the attribution paragraph below.

This story was reported by OpenAI. Read the full original coverage at OpenAI.