Why OpenAI’s Cerebras Chip Deal Matters; What Anthropic Wants to Know About Chinese Rivals
OpenAI executives have repeatedly signaled, both publicly and privately, that the only factor limiting the company’s revenue growth is access to computational resources. Chief Financial Officer Sarah Friar wrote an entire blog post on this topic earlier this year, and OpenAI executives highlighted the message again in a presentation to investors last week.
In the meeting, execs emphasized how the company has struck a variety of deals in recent months with cloud and chip providers including Microsoft, Nvidia, Amazon Web Services, Advanced Micro Devices and Broadcom to ensure it’ll have access to enough powerful servers to continue growing, according to a person with knowledge of the presentation. But in one way the most important deal may be OpenAI’s agreement to buy $10 billion worth of chips from a lesser known chip firm—Cerebras Systems—which OpenAI executives spent extra time highlighting last week, the person said.
It turns out the reason OpenAI struck the deal is highly technical but symbolizes OpenAI’s strategy for turning profitable: using different kinds of hardware in a way that reduces its costs.
Cerebras’ chips, for instance, aim to be faster and more efficient in running models, a process known as inference, than traditional AI chips such as Nvidia’s graphics processing units. Cerebras chips keep data on the processor itself through a type of computer memory known as Static Random-Access Memory, minimizing the need to shuttle data back and forth to separate high-bandwidth memory chips. As a result, the chips (in theory) process data faster for specific tasks involving AI apps.
Nvidia GPUs, in comparison, are much larger and take longer to process data, although they also have more flexibility to handle different types of operations well.
If Cerebras’s approach sounds familiar, that’s because another chip designer, Groq, uses a similar SRAM technology. But Nvidia signed a $20 billion deal in December to license tech from Groq and hire away its top talent, making it harder for other firms to work with the chip designer.
This all matters because OpenAI, as it revealed to investors last week, has sharply increased the amount of money it expects to burn between now and 2029 to $218 billion, including $140 billion on inference costs in that span. It needs to reassure investors it has some path to making money. The company is also working on developing its own inference-focused chip with Broadcom, which could also help bring down inference costs compared to using GPUs.
Of course, this is a lot to ask investors to believe in, especially for a company in a field as fast-changing as AI. And OpenAI already missed its margin projections last year, which doesn’t look great for future profitability goals.
But the company is going to need to convince investors that this is a viable strategy, especially since it’s going to have to keep fundraising to afford the unprecedented cash burn it has projected.