OpenAI turns to Broadcom built inference chip

OpenAI will now use its first custom AI chip “Jalapeño” made by Broadcom. It will be used for the heavy computing work of running AI models for people using ChatGPT and other apps.

The company called it a big step in its plan to build the full stack behind its models and products.

“By designing more of the stack ourselves, we can serve more intelligence with greater efficiency and keep pushing advanced AI toward broader access,” Greg Brockman, OpenAI’s president, said in a statement.

Broadcom’s stock (NASDAQ: AVGO) rose about 2% after the news. The stock has grown 10% so far this year.

OpenAI helped start the rush into generative AI back in 2022, and since then it has been one of the largest buyers of Nvidia’s expensive graphics processing units, the main hardware used to build AI models and run big jobs. Since the demand of OpenAI’s products is too high, the company requires more sources for advanced inference chips.

For this reason OpenAI had also joined hands with Amazon Web Services for its Trainium AI chips. It has also looked for alternatives with Nvidia’s rival AMD and Cerebras which IPO’d recently as reported by Cryptopolitan.

After working together for 18 months, OpenAI and Broadcom had been in talks for a custom chip since October. They planned to build and roll out racks of OpenAI-designed chips starting with the goal of making enough to use 10 gigawatts of power.

What the new chip can do

The Broadcom chip is an ASIC. Industry experts say this type is less flexible than Nvidia’s GPU but costs less and can be built for certain AI jobs. OpenAI said it designed the chip in nine months and also built large parts of the computer system that will run it.

The two companies call the chip an “Intelligence Processor” and describe it as the first piece in a wider platform meant to make AI faster and easier to reach.

Broadcom has built a good reputation in the times of AI boom, with its assistance of custom chips for big cloud firms and leading labs.

A physical sample of the chip will reach OpenAI on Wednesday. The companies said they hope to start using Jalapeño chips by the end of 2026 and expand from there.

The move comes after months of tension over Nvidia hardware

Eight sources told Reuters that OpenAI has been unhappy with some of Nvidia’s newest chips and has looked for other options since last year. Seven sources said the company is concerned with how fast Nvidia’s hardware answers users on certain tasks, such as writing software code. OpenAI required chips would eventually handle about 10% of OpenAI’s inference needs.

However Altman said in a X post, “We love working with NVIDIA and they make the best AI chips in the world. We hope to be a gigantic customer for a very long time.”

Nvidia still leads in chips for training large models, but inference has become a new battleground. In September, Nvidia said it planned to invest as much as $100 billion in OpenAI, a deal that would give the chipmaker a stake in the startup.

In February, Nvidia CEO Jensen Huang also dismissed reports of friction as “nonsense” and repeated that Nvidia planned a large investment. Still, the $100 billion deal was scrapped and replaced with a much smaller one.

Later, Jensen Huang confirmed the full $100 billion was “probably not in the cards,” and he later described the original figure as “never a commitment”, just an invitation to invest up to that amount.

What Nvidia actually finalized was a $30 billion equity stake in OpenAI, committed at a $730 billion pre-money valuation. Huang said it was likely Nvidia’s last investment in OpenAI before the company’s planned IPO.

OpenAI’s new cyber model edges out Anthropic

OpenAI has been busy on another front too. On June 22, it fully launched GPT-5.5-Cyber as part of its Daybreak cyber defense program.

On CyberGym, a test built at UC Berkeley using 1,507 known software flaws from 188 open-source projects, the model scored 85.6%. Anthropic’s Mythos 5 scored 83.8% and Claude Opus 4.7 scored 73.1%.

Anthropic’s Mythos 5 and Fable 5 were pulled offline on June 12 after the Trump administration issued an emergency export control order, citing national security. As of June 23, both remain down with no set return date.

The smartest crypto minds already read our newsletter. Want in? Join them.