Nvidia & Groq – Peak Preservation

Nvidia moves to preserve its dominance

• Nvidia’s deal with Groq is an acquisition in all but name, as the part of the company that is left has very little chance of being able to continue developing chips, meaning that Groq is now effectively part of Nvidia.
• On December 24^th 2025, Nvidia announced that it had acquired a non-exclusive license to Groq’s technology and that 90% of its employees would move over to work for Nvidia.
• For this non-exclusive license, Nvidia is paying a thumping $20bn, nearly 3x the company’s last valuation when it raised $750m in September 2025.
• No one in their right mind would pay this much for a non-exclusive licence unless they were very confident that the technology would only see the light of day as part of its products, with the rump of the company unable to compete.
• This is where the recruitment of the 90% of the company comes in, as while Groq will still be able to make and sell the current generation of LPUs, there will be no one around to design the next generation, meaning that Groq is likely to be a cloud service business and nothing more.
• This means that whatever the next generation of Groq is, it will be the sole property of Nvidia and very likely integrated into Nvidia’s product suite.
• There is unlikely to be a next generation of Groq outside of Nvidia, which is how this is effectively an acquisition but comes without the regulatory and time constraints that a formal purchase would require.
• Groq was founded in 2016 by the Google team that created the first TPU and originally focused on the automotive market with its very fast processor explicitly designed for inference.
• When ChatGPT happened in 2022, the company pivoted to the data centre and won a $1.5bn deal to supply inference systems to Humain of Saudi Arabia.
• With its focus on inference and a very different architecture using onboard SRAM rather than off-chip HBM, Groq was becoming a viable alternative to Nvidia for inference, which created a threat to Nvidia’s market share.
• This is because the real growth in the market now is for inference rather than training, and I have long been of the opinion that the vast majority of the market will end up being in inference as opposed to training.
• Nvidia has an iron-clad grip on training, but it is much easier to use chips from other companies to run a model trained on Nvidia, which is where competitors like Groq, AMD, Cerebras, and now Qualcomm have been making progress.
• Nvidia has been aware of both this threat and the threat of the control point in AI moving away from silicon (see here) and has been long taking steps to mitigate this risk and keep both market share and margins intact as the market evolves.
• This is what Nvidia Inference Microservices (NIMs) and Dynamo are all about, and the addition to Groq takes this strategy to a new level.
• Nvidia will be able to improve and expand its offering for inference, and with Groq LPUs becoming available with CUDA, this will create a strong incentive for customers to stick with Nvidia for inference as well as training.
• With Nvidia routinely at least one generation ahead of competitors, it can justifiably make the claim that it offers the best value when it comes to tokens per dollar, even at 70%+ gross margins.
• With Groq effectively gone, the mantle of competing with Nvidia in the merchant market will now fall to AMD, Cerebras, Qualcomm and others, who I suspect are all seeing a boost of interest following this deal.
• When it comes to valuation, there is no question which of these players offers the best value for money, which is why I am happy to hold onto my position in Qualcomm, where some big data centre deals could be a catalyst for a rerating on the shares.
• Nvidia looks stronger than ever and is still the best way to make a direct investment in AI, but I still argue that the adjacencies of inference at the edge and nuclear power remain much better value-oriented ways to get exposure to the AI trend.