OpenAI wanted to acquire Cerebras

According to new legal documents, OpenAI once considered acquiring Cerebras, a publicly traded artificial intelligence chip maker.

New evidence in Elon Musk's ongoing lawsuit against OpenAI describes how OpenAI considered acquiring Cerebras around 2017 — a year after Cerebras was founded and a few years after OpenAI began operations.

In an email to OpenAI CEO Sam Altman and Musk, who at the time was financially involved with OpenAI and had some influence over its direction, Ilya Sutskever, OpenAI co-founder and former chief scientist, floated the idea of ​​acquiring Cerebras through Musk’s electric car company Tesla.

“If we decide to acquire Cerebras, I strongly believe it will be done through Tesla,” Sutskever wrote in September 2017. “But why do it if we can also do it within OpenAI? Specifically, there is concern that Tesla’s obligation to maximize shareholder returns for its shareholders is inconsistent with OpenAI’s mission. As a result, the overall outcome may not be the best outcome for OpenAI.”

In a July 2017 email from Sutskever to Musk and OpenAI co-founder Greg Brockman (now the company's president), Sutskever mentioned several Cerebras-related agenda items: "negotiate merger terms with Cerebras" and "conduct additional due diligence on Cerebras."

The merger ultimately fell through, though it’s not clear from the exhibit why, and OpenAI ended up shelving its chip ambitions for years.

Cerebras, based in Sunnyvale, California, develops custom hardware to run and train AI models, and claims its chips are faster and more efficient than Nvidia’s flagship products for AI workloads.

Cerebras has reportedly raised $715 million in venture capital and hopes to roughly double its $4 billion valuation with an IPO. But it faces huge challenges. An Abu Dhabi company, G42, accounted for 87% of Cerebras' revenue in the first half of 2024, and U.S. lawmakers have expressed unease about G42's historical ties to China. Cerebras CEO Andrew Feldman also has a shady past, admitting to circumventing accounting controls while a vice president at publicly traded company Riverstone Networks.

If the acquisition goes through, both companies would benefit. Cerebras could avoid the tricky path to an IPO, while OpenAI could have a significant resource in the race to build its own chips.

OpenAI has long sought to reduce its reliance on Nvidia, which has a huge share of the market for AI-optimized chips. While OpenAI was late to the game in in-house chips (companies like Google and Amazon Web Services have long launched chips designed for AI workloads), it is under pressure to reduce the cost of training, fine-tuning, and running its models. Having its own chips could be one way to achieve the cost reductions it needs.

OpenAI had hoped to build a network of chip manufacturing plants and was considering acquisition targets. But the company reportedly abandoned those plans and instead actively built a team of chip designers and engineers and worked with semiconductor companies Broadcom and TSMC to develop an AI chip to run the model. The chip could be available as early as 2026.

Nvidia's competitors are taking action to break Nvidia's monopoly in the artificial intelligence chip market. They have raised hundreds of millions of dollars in funds and launched new products in the hope of sharing the benefits brought by the boom in artificial intelligence technology.

A host of smaller companies including Cerebras, d-Matrix and Groq are hoping to grab a piece of the multibillion-dollar AI chip market from Nvidia, which has so far dominated the first wave of investment with its graphics processing units (GPUs).

They expect that as chatbots and other generative AI applications become more popular, demand for AI “inference” — the computing power needed for models like OpenAI’s ChatGPT and Google’s Gemini to generate responses to queries — will grow exponentially.

Nvidia’s Hopper GPUs are ideally suited to the resource-intensive task of training top AI models and have become one of the world’s hottest commodities.

Cerebras, d-Matrix and Groq are focused on designing cheaper, more specialized chips for running AI models.

On Tuesday, Cerebras announced its new “Cerebras Inference” platform based on the CS-3 chip, which is about the size of a dinner plate. Cerebras claims its solution is 20 times faster at AI inference than Nvidia’s current generation Hopper chip, at a fraction of the price. Cerebras cited tests conducted by benchmarking provider Artificial Analysis.

“The way to beat the 800-pound gorilla is to bring a better product to market,” Cerebras CEO Andrew Feldman told the Financial Times. “In my experience, the better product generally wins, and we’ve already taken significant customers away from [Nvidia].”

Rather than using a separate high-bandwidth memory chip like Nvidia uses, the CS-3 chip offers an alternative architecture that builds the memory directly into the chip wafer.

Feldman said that limitations on memory bandwidth are a fundamental constraint on the speed of inference on AI chips. Combining logic and memory into one large chip can achieve results that are "orders of magnitude faster," he said.

Founded in 2019 by Sid Sheth, d-Matrix has also launched a new round of funding less than a year after raising $110 million in a Series B round led by Singapore state fund Temasek. Sheth said the company plans to raise $200 million or more later this year or early next year. d-Matrix is ​​in the early stages of the fundraising process and said the final amount of funds raised may change.

d-Matrix plans to fully launch its own chip platform, Corsair, by the end of this year. Sheth said the company is pairing its products with open software such as Triton, which competes with Nvidia's Cuda, a widely used software platform that gives developers tools to build AI applications and optimize the performance of its chips.

Nvidia's largest customers support the use of open software such as Triton. "Application developers don't like being locked into one specific tool," Sheth said. "People are realizing that Nvidia has absolute control over Cuda for training."

Groq, another AI inference competitor led by former founding members of Google’s Tensor Processing Unit team, raised $640 million this month from investors led by BlackRock Private Equity Partners at a valuation of $2.8 billion.

Despite the hype surrounding the sector, semiconductor startups face challenges breaking into the market, one venture capitalist warns.

SoftBank bought chipmaker Graphcore last month for just over $600 million, according to people familiar with the matter, less than the roughly $700 million in venture capital the company had raised since its founding in 2016.

Groq and Cerebras were also founded in 2016. “Public investors have been eager to find and back the next Nvidia,” said Peter Hébert, co-founder and managing partner of venture capital firm Lux Capital. “This is not just about chasing the latest trends. The momentum has also benefited several venture-funded chip startups that have been working for nearly a decade.”

Public account ID: icbank