Nvidia entered a new agreement on Wednesday with AI chip startup Groq. The deal focuses on inference technology. It is not exclusive.
The announcement left analysts at D.A. Davidson confused. Analyst Alex Platt said the strategy behind the deal was unclear. He shared his concerns in a note released on Friday.
Platt questioned Nvidia’s real motivation. He wondered if the move was driven by technology or by a need to protect its market position.
The analysts pushed back on claims that Groq has a strong advantage in inference chips. They pointed out that Groq’s current chip only has 230MB of SRAM memory. They called this extremely low.
For comparison, Nvidia’s HGX B300 chip comes with 288GB of HBM3E memory per chip. The gap is massive.
Because of this limitation, the analysts said Groq’s chips can only handle a small portion of inference tasks. They believe the technology cannot support next generation AI models. These upcoming models are expected to need memory measured in terabytes.
Platt said he still struggled to see the logic behind the deal.
As part of the agreement, Nvidia will license Groq’s inference technology. Nvidia will not get exclusive rights. Several key Groq executives will join Nvidia, including founder Jonathan Ross and President Sunny Madra.
Groq will continue operating as an independent company. It will now be led by new CEO Simon Edwards.
Reports suggest the deal could be worth as much as $20 billion in cash. However, Nvidia is not buying Groq and is not taking full control of its technology.
D.A. Davidson analysts said Groq’s chips may perform well in certain prefill tasks. Even so, they noted Nvidia already has its Rubin CPX platform coming soon.
They also questioned whether Nvidia was acting defensively. The analysts said Nvidia’s only real long term concern right now appears to be Google.
Despite all the doubts, the analysts said they trust Nvidia CEO Jensen Huang’s judgment. They said they are willing to give him the benefit of the doubt on the decision.
Comments are closed.