Sam Altman and Ilya Sutskever Invested in Cerebras But OpenAI Snubbed Its AI Chips

• AI chip startup Cerebras counts Sam Altman, Ilya Sutskever and Greg Brockman as investors.
• In 2017, OpenAI’s founders considered buying Cerebras.
• However, today, OpenAI has pivoted to other options.

Nearly a year after Cerebras launched its third generation 5nm Wafer Scale Engine (WSE-3), the Nvidia-beating chip is attracting serious attention from major AI developers.

However, although DeepSeek, Mistral, Perplexity and others have all embraced WSE-3 for inference workloads, OpenAI hasn’t.

The anomaly is all the more compelling because three of OpenAI’s founders are among the startup’s early investors and once considered buying Cerebras through a deal with Tesla.

The Cerebras Acquisition OpenAI and Tesla Never Made

The revelation that OpenAI mulled buying Cerebras emerged as part of the ongoing legal battle between OpenAI and Elon Musk. Emails released as evidence show that OpenAI’s interest in Cerebras can be traced back to 2017.

In one message to Elon Musk and Greg Brockman, Ilya Sutskever argued that to achieve Artificial General Intelligence, OpenAI would need to “lock down an overwhelming hardware advantage.”

Concluding that Cerebras’ AI accelerator was “far ahead” of Google’s TPUs, Sutskever said, “having exclusive access to them would put us far ahead of the competition.”

Over the next few months, the emails repeatedly refer to a potential Cerebras merger.

In one message, Sutskever wrote that “in the event we decide to buy Cerebras, my strong sense is that it’ll be done through Tesla.” The emails reveal that during 2017, OpenAI’s founders seriously considered forging a relationship with Tesla.

Andrej Karpathy, who was also Tesla’s Senior Director of AI at the time, said that the “most promising option” to secure OpenAI’s funding was to “attach to Tesla as its cash cow.”

In the end, Musk’s departure from OpenAI foreclosed the possibility of joining forces with Tesla, and the startup eventually reached a funding agreement with Microsoft instead.

For unknown reasons, under Microsoft’s sponsorship, OpenAI’s founders never pursued the Cerebras acquisition.

Cerebras Emerges as Nvidia Rival

With Microsoft’s backing, OpenAI’s ability to buy expensive Nvidia chips increased significantly. However, it has never quite attained the overwhelming hardware advantage Sutskever envisaged in 2017.

Post-ChatGPT, OpenAI’s hardware question has re-emerged.

While Nvidia GPUs remain the industry standard for AI training, Cerebras increasingly has the advantage in inference. The difference can be observed by running the same AI models on GPUs versus WSE-3s.

According to company claims, Cerebras chips can run Llama 3.1 16 times faster  and DeepSeek R1 57 times faster than the fastest GPU-based solutions.

This significant speed advantage hasn’t gone unnoticed by AI firms.

AI Labs Embrace Cerebras

In February, Mistral shifted inference for its Le Chat platform to Cerebras. Meanwhile, Perplexity opted to host its latest Sonar model on Cerebras’ inference infrastructure.

So far, however, the biggest AI infrastructure providers don’t offer the new chips. The reasons for this are partly practical.

“It’s hard for them to switch,” remarked Ahmad Shadid, founder of Cerebras-powered AI platform O.XYZ. “To switch to new infrastructure, new types of chips, you need a new type of engineers that can understand how to deploy the models,” he explained.

Meanwhile, as a young startup, he pointed out that Cerebras still has some scaling to do before it can supply the kind of volumes Big Tech customers deal with.

As for OpenAI, Shadid suggested that an “ego battle” between the two startup founders prevented collaboration. Moreover, since merger talks with Cerebras stalled, OpenAI has advanced other hardware projects.

OpenAI’s Hardware Ambitions

To meet the needs of ChatGPT’s 400 million active users , OpenAI needs immense computational resources.

The startup’s infrastructure is currently GPU-based, but OpenAI and its leadership haven’t given up on their chip ambitions.

In early 2024, reports surfaced that OpenAI CEO Sam Altman was raising money for a new AI chip venture. However, that project doesn’t seem to have gone anywhere.

Then, in September, news spread that OpenAI had teamed up with Broadcom to develop a new inference chip to be manufactured by Taiwan Semiconductor Manufacturing Company (TSMC).

Finally, although OpenAI hasn’t embraced Cerebras’ WSE-3s, it has turned to other emerging AI accelerators.

The New Kid on the Block

In 2019, OpenAI put in a pre-order for $51 million worth of chips from Rain AI, another startup backed by Altman.

Between Rain and Cerebras, which both count the OpenAI CEO as an early investor, Altman appears to now favor Rain. He reportedly pitched the startup to OpenAI investors, in 2024.

Comparing the two, Rain’s technology is more radical, proposing an analog alternative to the current dominant digital semiconductors.

Commenting on the company’s “neuromorphic approach” in 2022, Altman said Rain’s chips “could vastly reduce the costs of creating powerful AI models and will hopefully one day help to enable true artificial general intelligence.”

Has OpenAI lost interest in Cerebras as newer, more innovative solutions emerge? Or is it really just a matter of ego, as Shadid suggested? The answer isn’t clear.

What is certain is that the market for inference chips is increasingly competitive.

Alongside a raft of startups, major chipmakers, including Apple, AMD and Samsung, are all developing their own accelerators. So, too, are the hyperscale cloud giants that handle a significant portion of AI processing today.

Against that backdrop, OpenAI has plenty of options.