OpenAI has committed billions to recent chip deals. Some big names have been left out

In November, pursuing Nvidia's latest earnings beat, CEO Jensen Huang boasted to investors astir his company's presumption successful artificial quality and said astir the hottest startup successful the space, "Everything that OpenAI does runs connected Nvidia today."

While it's existent that Nvidia maintains a ascendant presumption successful AI chips and is present the astir invaluable institution successful the world, contention is emerging, and OpenAI is doing everything it tin to diversify arsenic it pursues a historically assertive enlargement plan.

On Wednesday, OpenAI announced a $10 cardinal deal with chipmaker Cerebras, a comparatively nascent subordinate successful the abstraction but 1 that's angling for the nationalist market. It was the latest successful a string of deals betwixt OpenAI and the companies making the processors needed to physique ample connection models and tally progressively blase workloads.

Last year, OpenAI committed much than $1.4 trillion to infrastructure deals with companies including Nvidia, Advanced Micro Devices and Broadcom, en way to commanding a $500 billion backstage marketplace valuation.

As OpenAI races to conscionable anticipated request for its AI technology, it has signaled to the marketplace that it wants arsenic overmuch processing powerfulness arsenic it tin find. Here are the large spot deals that OpenAI has signed arsenic of January, and imaginable partners to support an oculus connected successful the future.

Nvidia 

Since its aboriginal days gathering retired ample connection models, agelong earlier the motorboat of ChatGPT and the commencement of the generative AI boom, OpenAI has relied connected Nvidia's graphics processing units.

In 2025, that narration went to different level. Following an concern successful OpenAI successful precocious 2024, Nvidia announced successful September that it would perpetrate $100 cardinal to enactment OpenAI arsenic it builds and deploys astatine slightest 10 gigawatts of Nvidia systems. 

A gigawatt is simply a measurement of power, and 10 gigawatts is astir equivalent to the yearly powerfulness depletion of 8 cardinal U.S. households, according to a CNBC investigation of information from the Energy Information Administration. Huang said successful September that 10 gigawatts volition equate to between 4 cardinal and 5 cardinal GPUs. 

"This is simply a elephantine project," Huang told CNBC astatine the time. 

OpenAI and Nvidia said the archetypal signifier of the task is expected to travel online successful the 2nd fractional of this twelvemonth connected Nvidia's Vera Rubin platform. However, during Nvidia's quarterly net telephone successful November, the institution said determination is "no assurance" that its statement with OpenAI volition advancement beyond an announcement and to an authoritative declaration stage. 

Nvidia's archetypal concern of $10 cardinal volition beryllium deployed erstwhile the archetypal gigawatt is completed, and investments volition beryllium made astatine then-current valuations, arsenic CNBC previously reported.

AMD

In October, OpenAI announced plans to deploy six gigawatts of AMD's GPUs crossed aggregate years and aggregate generations of hardware. 

As portion of the deal, AMD has issued OpenAI a warrant for up to 160 cardinal shares of AMD communal stock, which could magnitude to a astir 10% involvement successful the company. The warrant includes vesting milestones tied to some deployment measurement and AMD's stock price.

The companies said they program to rotation retired the archetypal gigawatt of chips successful the 2nd fractional of 2026, and added that the woody is worthy billions of dollars, without disclosing a circumstantial amount.

"You request partnerships similar this that truly bring the ecosystem unneurotic to guarantee that, you know, we tin truly get the champion technologies, you know, retired there," AMD CEO Lisa Su told CNBC astatine the clip of the announcement.

Altman planted the seeds for the woody successful June, erstwhile helium appeared connected signifier alongside Su astatine an AMD launch event successful San Jose, California. He said OpenAI planned to usage AMD's latest chips. 

Broadcom 

Later that month, OpenAI and Broadcom publically unveiled a collaboration that had been successful the works for good implicit a year.

Broadcom calls its customized AI chips XPUs, and has frankincense acold relied connected a fewer customers. But its pipeline of imaginable deals has sparked truthful overmuch enthusiasm connected Wall Street that Broadcom is present valued astatine implicit $1.6 trillion.

OpenAI said it's designing its ain AI chips and systems that volition beryllium developed and distributed by Broadcom. The companies person agreed to deploy 10 gigawatts of these customized AI accelerators.

In the October release, the companies said Broadcom volition purpose to statesman deploying racks of the AI accelerator and web systems by the 2nd fractional of this year, with a extremity to implicit the task by the extremity of 2029.

But Broadcom CEO Hock Tan told investors during the company's quarterly net telephone successful December that helium doesn't expect overmuch gross from the OpenAI concern successful 2026.

"We admit the information that it is simply a multiyear travel that volition tally done 2029," Tan said. "I telephone it an agreement, an alignment of wherever we're headed."  

OpenAI and Broadcom did not disclose the fiscal presumption of the deal.  

OpenAI connected Wednesday announced an statement to deploy 750 megawatts of Cerebras' AI chips that volition travel online crossed aggregate tranches done 2028. 

Cerebras builds ample wafer-scale chips that tin present responses up to 15 times faster than GPU-based systems, according to a release. The institution is overmuch smaller than Nvidia, AMD and Broadcom.

OpenAI's woody with Cerebras is worthy much than $10 billion, and it could beryllium a boon for the chipmaker arsenic it weighs a imaginable debut connected the nationalist markets. 

"We are delighted to spouse with OpenAI, bringing the world's starring AI models to the world's fastest AI processor," Cerebras CEO Andrew Feldman said successful a statement. 

Cerebras sorely needs marquee customers. In October, the institution withdrew plans for an IPO, days aft announcing that it raised implicit $1 cardinal successful a fundraising round. It had filed for a nationalist offering a twelvemonth earlier, but its prospectus revealed a heavy reliance on a azygous lawsuit successful the United Arab Emirates, Microsoft-backed G42, which is besides a Cerebras investor.

Potential partners 

Where does that permission Amazon, Google and Intel, which each person their ain AI spot plays?

In November, OpenAI signed a $38 billion unreality woody with Amazon Web Services. OpenAI volition tally workloads done existing AWS information centers, but the unreality supplier besides plans to physique retired further infrastructure for the startup arsenic portion of the agreement. 

Amazon is besides successful talks to perchance put much than $10 billion successful OpenAI, arsenic CNBC antecedently reported. 

OpenAI could determine to usage Amazon's AI chips arsenic portion of these concern discussions, but thing authoritative has been determined, according to a idiosyncratic with cognition of the substance who asked not to beryllium named due to the fact that the discussions are confidential. 

AWS announced its Inferentia chips successful 2018, and the latest generation of its Trainium chips successful precocious 2025.

Google Cloud besides supplies OpenAI with computing capableness acknowledgment to a deal that was softly finalized past year. But OpenAI said successful June that it has nary plans to usage Google's in-house chips called tensor processing units, which Broadcom besides helps produce.  

Intel has been the biggest AI laggard among accepted chipmakers, which explains wherefore the institution precocious took monolithic investments from the U.S. authorities and Nvidia. Reuters reported successful 2024, citing radical with cognition of the discussions, that Intel had a accidental to put successful OpenAI years earlier and perchance marque hardware for the then-fledgling startup, offering it a mode to debar reliance connected Nvidia.

Intel decided against the deal, according to Reuters.

In October, Intel announced a caller information halfway GPU that it's codenamed Crescent Island and says is "designed to conscionable the increasing demands of AI inference workloads and volition connection precocious representation capableness and energy-efficient performance." The institution said "customer sampling" is expected successful the 2nd fractional of 2026.   

Wall Street volition perceive updates connected Intel's latest AI efforts erstwhile the institution kicks disconnected tech net play adjacent week.

— CNBC's Kif Leswing, MacKenzie Sigalos and Jordan Novet contributed to this report.

WATCH: Breaking down AI chips, from Nvidia GPUs to ASICs by Google and Amazon