OpenAI unveils jalapeño, its first in‑house AI ASIC chip

At a glance:

• OpenAI announced jalapeño, an ASIC‑based "intelligence processor" co‑developed with Broadcom.
• The chip aims to give OpenAI a full‑stack AI platform and enable gigawatt‑scale data centers by 2026.
• Development was completed in nine months, which OpenAI calls the fastest ASIC cycle ever.

OpenAI reveals its first custom AI processor

OpenAI’s blog post on Wednesday introduced jalapeño, an application‑specific integrated circuit (ASIC) designed specifically for running large language models. Unlike the general‑purpose GPUs that dominate the AI market and are supplied mainly by Nvidia, jalapeño is a purpose‑built “intelligence processor” that promises higher efficiency for inference and training workloads. The announcement marks the company’s most concrete step toward a fully independent, full‑stack AI platform—controlling everything from model architecture to the silicon that powers it.

The post emphasized that jalapeño is not a GPU but an ASIC, a class of chip that performs a narrower set of tasks with far greater speed and power efficiency. OpenAI president and co‑founder Greg Brockman said the move will let the company “serve more intelligence with greater efficiency and keep pushing advanced AI toward broader access.” By owning its own silicon supply, OpenAI expects to lower costs and reduce latency during peak demand periods.

Partnership with Broadcom and the gigawatt roadmap

The jalapeño chip was co‑developed with U.S. chipmaker Broadcom. In October, the two firms announced a plan to build enough custom AI racks to deliver ten gigawatts of power—enough electricity for roughly 7.5 million homes. Broadcom CEO Hock Tan described jalapeño as “just the beginning of a multi‑generation roadmap,” adding that the chips will be deployed in gigawatt‑scale data centers together with Microsoft and other partners starting in 2026.

The collaboration also signals a shift in OpenAI’s supply chain. While the company currently purchases most of its silicon from Nvidia, it also maintains relationships with Amazon, AMD and Cerebras. Jalapeño gives OpenAI a new, proprietary hardware layer that could eventually replace or supplement those external sources.

Record‑fast ASIC development cycle

OpenAI reported that jalapeño was taken from concept to silicon in nine months, a timeline the company claims is “the fastest ASIC development cycle ever achieved in high‑performance advanced semiconductors.” The speed was attributed in part to OpenAI’s own AI tools, which were used to optimise the chip architecture and verify designs. The blog noted, “The same models served to users are helping improve the infrastructure used to run future models.”

This recursive use of AI—where models help design the hardware that will run the next generation of models—highlights a broader trend of “recursive self‑improvement” in the industry. While it promises rapid advances, both OpenAI and Anthropic have warned that unchecked acceleration could lead to an “intelligence explosion,” prompting calls for an international oversight committee.

What the launch means for the AI hardware landscape

Jalapeño’s introduction could reshape the competitive dynamics of AI compute. Nvidia’s dominance has been built on its GPU ecosystem, but a purpose‑built ASIC from a major AI developer may pressure GPU vendors to innovate faster or lower prices. Broadcom, traditionally a networking and storage chipmaker, now positions itself as a key player in AI silicon, potentially expanding its market beyond data‑center networking.

For customers, the promise of cheaper, faster inference could translate into lower subscription fees for services built on OpenAI’s models. For the broader ecosystem, the move underscores a growing desire among AI leaders to internalise the hardware stack, reducing reliance on external suppliers and gaining tighter control over performance, cost and supply‑chain risk.

Looking ahead

OpenAI’s roadmap projects jalapeño‑powered data centers to go live in 2026, with additional generations of the chip slated for future releases. The company’s statement suggests a phased rollout, beginning with pilot deployments before scaling to the gigawatt‑level capacity described in the Broadcom partnership. As the hardware matures, we can expect OpenAI to integrate jalapeño more deeply into its API offerings, potentially offering dedicated endpoints that exploit the ASIC’s efficiency gains.

The broader AI community will be watching how jalapeño performs in real‑world workloads and whether other AI firms follow suit with custom silicon. If OpenAI’s approach proves cost‑effective and scalable, it could accelerate a wave of vertical‑integrated AI stacks across the industry.