OpenAI has unveiled Jalapeño, its first custom-designed inference processor, built in collaboration with semiconductor giant Broadcom. The chip marks a significant move by the AI company to reduce its dependence on Nvidia GPUs and control its own hardware destiny.

Unveiled on June 24, 2026, Jalapeño is purpose-built for inference — running pre-trained AI models in response to user commands — rather than the heavy training workloads that Nvidia's H100 and B200 GPUs dominate. Early benchmarks show significantly better performance-per-watt than current alternatives, according to the company.

Why OpenAI Needs Its Own Chip

OpenAI operates some of the world's largest AI workloads across ChatGPT, Codex, and its API platform. Each inference request incurs compute costs, and even marginal improvements in efficiency translate to massive savings. By designing custom silicon, OpenAI follows the playbook of Google (TPU) and Amazon (Trainium), both of which built custom chips to control costs and optimize performance for their specific workloads.

Greg Brockman, President of OpenAI, said: "We have a deep understanding of the workload. We've really been looking for specific workloads that are underserved, and asking how we can build something that will be able to accelerate what's possible."

The Broadcom Partnership and Chip Details

OpenAI's partnership with Broadcom was first officially announced in October 2025, though rumors of a custom chip initiative had been circulating for months. The Jalapeño chip is the first fruit of that collaboration. Broadcom brings extensive experience in custom ASIC design, having previously worked with Google on its TPU line.

The chip design itself was partially assisted by OpenAI's own AI models — a circular efficiency where the company used its intelligence to design more intelligent hardware. Jalapeño is currently being tested internally, with no public timeline for broad deployment.

Inference Economics and the Nvidia Dependency Problem

Nvidia currently commands over 80 percent of the AI chip market, with its H100 and B200 GPUs powering the vast majority of both training and inference workloads. However, as AI inference becomes the dominant cost driver — far exceeding training in aggregate — companies are racing to build specialized hardware.

OpenAI notes that pre-training of frontier models will continue to rely on Nvidia GPUs for the foreseeable future. But inference, which runs constantly for billions of user queries daily, represents a far larger addressable cost. Even small reductions in per-query inference cost can dramatically improve OpenAI's bottom line.

Also read: AMD launches $1,499 AI box to challenge Nvidia $4,000 dominance

Full-Stack Strategy: From Silicon to Software

OpenAI described its approach as full-stack ownership: chip architecture, kernels, memory systems, networking, scheduling, deployment systems, and product experience. Because OpenAI operates across every layer, each can be optimized toward the same goal — making AI models faster, more reliable, and more affordable.

This vertical integration contrasts with Nvidia's horizontal model, where the same GPU silicon serves thousands of different customers with different workloads. Custom silicon allows OpenAI to eliminate overheads that general-purpose chips must accommodate.

Also read: Meta raises 2026 AI capex to $145B — what it means for OpenAI

Impact on the AI Semiconductor Landscape

Jalapeño accelerates a broader trend in the AI industry: the shift from buying chips off the shelf to building custom silicon. Google's TPU, Amazon's Trainium and Inferentia, Microsoft's partnership with OpenAI-linked chip designs, and now OpenAI's Jalapeño — all point toward an industry where the biggest AI companies design their own hardware.

For India, this trend underscores the importance of domestic chip design capabilities. With the India Semiconductor Mission allocating $10 billion to boost local fabrication and design, Indian AI companies may eventually follow the same path of custom silicon optimized for local language models and use cases.

What Comes Next

OpenAI has not announced when Jalapeño will enter production or whether it will be available to third parties. Given the scale of OpenAI's infrastructure needs — the company operates data centers across the US and is expanding globally — the chip could first appear as a power-saving upgrade within OpenAI's own server fleet before any potential commercialization.

The company has also hinted at future chip designs targeting training workloads, suggesting that Jalapeño is just the first step in a broader semiconductor strategy.

Sources: TechCrunch, Reuters, OpenAI Official Blog, Broadcom Investor Relations