Since 2019, Mellanox Technologies has been Nvidia’s largest acquisition, paying $6.9 billion for the networking company. Now, its biggest purchase technically isn’t a full-scale takeover. Instead, Nvidia’s buying the assets of alternative AI chipmaker Groq. Note that the company itself remains independent. but only on paper. There is also talk of a “non-exclusive license agreement,” although Groq will effectively disappear from the AI chip market. What exactly is going on?
The press release announcing the agreement is possibly the most concise we have ever seen for a deal worth tens of billions. It does contain the most important information: founder Jonathan Ross, president Sunny Madra, and “other members of the Groq team” are now Nvidia employees. Given that Groq’s expertise is limited to a relatively small platform called GroqCloud, little else remains of the AI startup. New designs will be integrated and scaled up at Nvidia by former Groq staff. In other words: don’t expect any future Groq chip designs to be created after today.
Elephant in the room
Nvidia’s huge revenue and market value make it easy to imagine how it could come up with $20 billion for an acquisition. The biggest stumbling block for the AI giant, which controls more than 90 percent of the chip market for the technology, is the legislative branch. For example, it was not allowed to buy Arm for $40 billion because of the potential distortion of competition. Even in 2022, when Nvidia was “only” worth hundreds of billions and not trillions, this was too big a step according to various antitrust authorities.
Groq never got beyond being an AI startup with interesting ideas, so it is not really comparable to Arm. However, the promise surrounding the LPU (Language Processing Unit) was and remains great. It is the brainchild of Jonathan Ross’s team, which was previously behind Google’s TPU (Tensor Processing Unit). Said TPUs are currently considered the only alternatives to GPUs at hyperscale level and are an asset for Google in challenging Nvidia. The LPU changes the game, however, in that it is specifically designed to tackle generative AI workloads. In effect, it’s an ASIC. It just happens to be an ASIC for what’s currently the single most lucrative workload to be running. In terms of inferencing, Groq has always targeted efficiency over anything else, especially when compared to Nvidia’s high-overhead, more versatile GPU offering.
Beyond Groq, there are parties such as Cerebras, intended Intel’s acquisition target SambaNova, and smaller players in various countries. None of these startups are really capable of breaking through in the field of AI training, where OpenAI, Oracle, Microsoft, Amazon, and Google are buying GPUs en masse from Nvidia and, to a much lesser extent, AMD.
As an AI giant, Nvidia could theoretically raise the money to buy every AI chip manufacturer outside of Google. Yet it is not doing so, and neither is Groq. Although it is effectively buying up everything Groq offers for $20 billion, the deal is not technically exclusive and Groq will continue to exist. The question is whether regulators will see through the smokescreen. Bernstein analyst Stacy Reagon argues that the deal’s structure upholds a “fiction of competition”, one that is rather transparent.
Similar constructions
Nvidia is already in the crosshairs of many critics. The key question is whether the AI hype is worth the company’s sky-high valuation. However, as a pickaxe seller in the race for the AI gold mine, its position is relatively solid, because real money does appear in Nvidia’s account from the sale of GPUs. In recent months, this state of affairs has changed, largely because Nvidia has become entangled in a web of circular deals involving OpenAI, Oracle, SoftBank, and others. The fiercest criticism revolves around Nvidia supposedly being a kind of modern Enron, the long-gone American energy company at the center of perhaps the most infamous large-scale accounting scandal in history.
Nvidia’s accounting practices are not that shady in themselves, but the network of AI players that provide each other with tens or hundreds of billions of dollars through press releases is certainly questionable. The practice of not acquiring tech companies but effectively emptying them out is a more familiar tactic in that regard. Meta also did not “buy” Scale AI in June 2025, but it did acquire a 49 percent stake and hired its founder. The same thing happened with Character.AI, whose founders are now back at Google after a $2.5 billion deal to license Character.AI’s models.
The real ‘moat’
We have concluded on several occasions that Nvidia holds a position of power that the rest of the industry will find difficult to shake. Only Google is in a position to challenge it, but that company’s TPUs are not available to third parties. This leaves AMD as the only real alternative, but since every data center GPU that exits the factory finds a buyer, the supply is not particularly competitive. Nvidia’s ‘moat’ also consists of software and deep integrations with all relevant IT players worldwide. However, one aspect has been overlooked until now, and that is the power of the company’s astronomical market value combined with its enormous turnover. These financial resources enable Nvidia to use Groq constructions to destroy any emerging AI alternative long before it has even gained a measurable market share.
This could cause major problems in the long run and is already costing any company with an interest in AI training large sums of money. The chip world is already one of monopolies or dominance, think of TSMC’s reign or that of ASML. The only area outside Google where Nvidia might expect competition is China. Nvidia’s H200 chips will be allowed to be exported from the US starting in February, but the question is whether Beijing even wants that anymore. After all, it hopes to develop its own AI ecosystem with Huawei, SMIC, Alibaba, and DeepSeek. That will be a challenge that requires a lot of stamina. In the West, Nvidia’s purchasing power is so great that only Google, AMD, and regulators offer an alternative. That, and possibly idiosyncratic private start-ups that can refuse 20 billion or more.