No, we did not miss the fact that Nvidia did an “acquihire” of AI accelerator and system startup and rival Groq on Christmas Eve. But, because our family was traveling on Christmas Day and The Next Platform was on holiday, we knew we would have to circle back to suss out what Nvidia was shelling out $20 billion to get its hands on.

We did – embarrassed to say – entirely miss that Nvidia did a similar and much smaller acquihire of key personnel and licensing of key intellectual property at network convergence startup Enfabrica back in the middle of September 2025 for a reported $900 million.

Both point to what could end up being a totally new approach to AI inferencing for the GPU accelerator and interconnect maker – so much so that the devices that Nvidia ultimately makes a few generations from now cannot be called GPUs at all.

One could almost make that case with the current crop of datacenter-class GPU accelerators from Nvidia, which look less and less like graphics processing units and more like complex aggregations of vector and tensor engines, caches, and fabric interconnects for doing the relatively low precision mathematics that underpins GenAI and other kinds of machine learning and sometimes HPC.

This deal with Groq is peculiar in a number of ways. The first one is why Groq’s investors sold in the first place. As we pointed out in our analysis of the $10 billion deal between AI model maker OpenAI and AI hardware upstart Cerebras Systems (which was founded around the same time as Groq in 2015 as the AI machine learning was really starting to get traction), the wonder is why Groq would sell now, when low latency, high throughput AI inference is absolutely necessary and Groq is one of the few suppliers who can give Nvidia a run for the money here. Cerebras with its CS-2 waferscale compute engines, Google with its TPUs, and Amazon Web Services with its Trainiums (no one talks about Inferentia anymore because Trainium can do both AI training and inference are the only AI XPUs that have really gotten traction, and Nvidia GPUs dominate both training and inference with AMD getting its share with its datacenter GPUs.

Zoom out and look at this from Groq’s point of view and this is the best time to sell an alternative to Nvidia GPUs, which are expensive even if they are versatile. The acquihire deal has Nvidia licensing the company’s Learning Processing Unit technology and hiring most of Groq’s key engineering people, including co-founder Jonathan Ross and chief operating officer Sunny Madra, for $20 billion. That is a lot of money for a company that had five funding rounds totaling $1.75 billion and a valuation when the last bit of that – $750 million in Series E – came in September 2025 and a valuation of $6.9 billion. Ross had a $1.5 billion commitment from Saudi Arabia to build a massive GroqCloud datacenter in Dammam in his back pocket, but as far as we know that has not happened yet. This will be business that the remaining Groq will chase, since it is basically GroqCloud services, a bunch of intellectual property, and as far as we know, not a plan for a future LPU or GroqWare product line.

Acquisitions are usually both defensive and offensive, and the fully scheduled compiler that Ross spearheaded – and that makes an LPU very different from the initial TPUs that Ross created at Google – is a key asset that Nvidia surely did not want to see fall into enemy hands. Intel needs to buy an AI future, particularly one based on inference, and if it was sniffing around SambaNova, as has been rumored, then it has also been sniffing around Groq as well as Cerebras. But Intel doesn’t have any money, and it has the US government, now an investor, looking over its shoulder. AMD was also a potential suitor for Groq, and if the Groq software stack is truly different, then in theory AMD still has the right to license it as well as any hardware it might think is useful.

Yes, we know. That is truly funny.

A $1.5 billion commitment from Saudia Arabia for a GroqCloud outpost in the desert is not the same thing as an actual contract or better still a check or wire transfer. And on top of that, $1.5 billion is not an Earth-shattering amount of AI iron these days when OpenAI has committed to at somewhere around 30 gigawatts of capacity for AI hardware. Each gigawatt is, depending on who you ask and the circumstances, costs between $35 billion and $50 billion per gigawatt. Call Sam Altman’s capacity planning dream $1.5 trillion for 30 gigawatts. The Groq commitment with the Kingdom is 6.7X smaller than the one Cerebras has just inked with OpenAI, and it is three orders of magnitude smaller than what OpenAI wants to build, give or take.

So, when Ross and Huang got to talking, maybe 2.9X valuation seemed like a pretty good exit price, given that all of the hyperscalers and cloud builders are creating their own AI XPUs as well as using Nvidia and sometimes AMD GPUs and model builders like Anthropic are committing to using Google TPUs and AWS Traniums. It will be problematic to see Groq LPUs into China, where the other action is, and Europe has not quite figured out how to participate more fully in the GenAI Boom in a unique and indigenous way.

Even without all of the defensive reasons Nvidia might want to take out Groq, you can see why Ross and the Groq investors were cool with this deal. And so, now Jonathan Ross, one of the two co-founders of Groq, is now chief software architect at Nvidia and Sunny Madra is vice president of hardware at Nvidia. So that is that.

The acquihire structure is simple: After seeing how the world’s antitrust regulators dragged their feet on the $6.9 billion acquisition of Mellanox Technologies and quashed Huang’s $40 billion dream of acquiring Arm, Nvidia is leaving a shell behind so it doesn’t look like it bought the whole of Groq. There will be a rule change here by the US government, for sure, but we also assume that Huang got the nod from President Trump to make the move as well.

From where we sit, if the Groq team has been extracted and there will be no future LPU development at the remaining Groq, then Nvidia has left itself open for possible antitrust violations as interpreted by the major governments of the world that, like it or not, have a say over these kinds of mergers and acquisitions. If Nvidia did not want to trigger regulators, it would have done a deal that was less than Groq’s current valuation – a lot less – and then the Groq founders and the company’s investors would have laughed their butts off as they closed the door and made a phone call to AMD. There is a lot of playing chicken going on here.

Here is the other thing: There is no rule that Nvidia has to use the technology it has licensed. It happens all the time that companies get acquired and then sat on because they were going to disrupt the status quo. Our favorite example of this was Transitive, whose QuickTransit emulator was able to run mainframe apps on Unix or Unix apps on Linux with little modification. QuickTransit was used in the “Rosetta” emulation environment that Apple created to move from PowerPC to X86 processors in its PCs, and it worked miraculously well. IBM was going to get hurt real bad by QuickTransit, so it bought Transitive in late 2008. After some mumbo-jumbo about emulating other systems on its Power Systems machines, Big Blue just shut it all down and stopped talking about it in 2011.

The Enfabrica acquihire is similar to the Groq acquihire in that it might signal a change in architecture . . . or not. It might simply be defensive maneuvering disguised as offensive mixing of technologies out on the Nvidia roadmap. (Nvidia has not done that before, but the Nvidia today is not the Nvidia we knew five or ten years ago.)

Enfabrica dropped out of stealth mode back in June 2021, and we didn’t really have much of a sense what the company planned to do. By March 2023, we could see it developing, with Enfabrica’s “Millenium” ACF-S silicon converging extended memory and host I/O all down to one chip, getting rid of network interface cards, PCI-Express switches, CXL switches, and top of rack or leaf switches in the rackscale architecture.

The first product to put ACF-S to work was called a SuperNIC, and it was used to make an extended memory server based on CXL to radically boost the scale and performance of KV caches at the heart of AI inference workloads. This memory godbox, called Emfasys, was launched in July 2025, and significantly, the company’s founders told us at the time that adding a rack of Emfasys memory extenders to four racks of GB200 NVL72 rackscale servers would cut the cost per token in half (which means it was doubling the throughput of the GPUs by having this extended memory).

We think that there is a chance that Nvidia wants to build a much better inference machine not based strictly on its current GPU architecture, and that Groq and Enfabrica technology will play a part in it. But there is an equal chance that these two acquihire deals are really about making sure no one else does. And the odds are that it is actually both at the same time.

“¿Por qué no los dos?” as my oldest two bilingual children taught me to say half their lives ago.