Nvidia’s subterfuge to acquire Groq
The GPU giant accelerates its push into inference-dedicated chips
This is a tactic popularized last year by Microsoft that Nvidia has now adopted. In late December, the S graphics-card giant announced a “non-exclusive licensing agreement” with Groq, an US startup specializing in chips designed for inference in generative artificial intelligence models. Behind this wording lies what is effectively a disguised acquisition, aimed at securing key technologies and engineering teams while sidestepping a potential veto from competition authorities.
Structured this way, the deal does not require prior regulatory approval. That said, it cannot be ruled out that antitrust regulators will take a closer look. Last year, the US FTC, the European Commission, and the UK’s CMA examined the partnership between Microsoft and Inflection. While the FTC has yet to issue a decision, its counterparts ultimately concluded that the arrangement complied with competition rules. The financial scale and strategic implications appear greater in Nvidia’s case, however — though any such proceedings would in any event take years to materialize.
90% of employees moving to Nvidia
According to various outlets, Nvidia will pay $20 billion to Groq. Most of that sum will go to its investors, delivering a substantial return — at its last funding round in September, the startup was valued at just $6.9 billion. More significantly, around 90% of Groq’s employees, including founder and CEO Jonathan Ross, will join the Santa Clara-based group. While Groq says it will continue to operate “independently,” it is already seeking to sell off its remaining assets, notably its cloud platform, according to The Wall Street Journal.
Founded in 2016 by former Google engineers who helped design the first TPU, the search giant’s AI chip, Groq long remained under the radar. On several occasions, the startup even came close to bankruptcy. The rise of generative AI, however, allowed it to change scale, positioning itself as one of the new challengers aiming to compete with Nvidia, whose GPUs remain the industry benchmark. Since the summer of 2024, Groq has raised $1.4 billion, funding that enabled it in particular to build its own data centers.
Chips dedicated to inference
Unlike Nvidia, Groq specifically targeted the inference market — the phase in which AI models are run to generate text, images, or video. This process requires far less computing power than training, yet it is still largely carried out on the same GPUs. With its specialized chip, which it calls an LPU (Language Processing Unit), the company promised significantly higher performance: execution speeds up to ten times faster, enabling quicker responses, and energy consumption divided by ten.
The deal with Groq reflects a growing realization inside Nvidia: demand for chips exclusively dedicated to inference is rising sharply, particularly as companies seek to curb investment and operating costs. In September, the company unveiled a GPU specifically designed for this task, scheduled for release at the end of 2026. That model, however, was still based on an architecture similar to its existing accelerators. Nvidia now plans to integrate Groq’s architecture into its own product lineup.
Countering Google?
Jensen Huang’s group thus kills two birds with one stone. On the one hand, it eliminates a potential competitor — even if Groq’s first-generation chips did not necessarily deliver on all their promises. On the other, it acquires technological expertise and patents that should allow it to offer some of the most advanced inference-focused solutions, countering those of Cerebras or SambaNova. As well as Google’s upcoming Ironwood TPUs, designed, according to the company, for “the era of inference.”
In fact, the timing of the operation may not be coincidental. At the end of November, The Information reported that the Mountain View group now intends to sell its TPUs — previously reserved for its own data centers and cloud customers. Meta could be among the first buyers, integrating its rival’s chips into its infrastructure starting in 2027. Google executives reportedly believe they could capture 10% of a market currently dominated by Nvidia.