When news broke late on Christmas Eve that Nvidia had all-but acquired Groq, two things were immediately clear. First, the announcement had been timed to drop after the markets closed early that day in order to give markets time to understand what was going on before reacting. And second, that no-one without direct knowledge of the deal understood even remotely what on Earth was going on.

Over the last two weeks, there have been many, many armchair analysts posting on social media about supposed reasons the market leader, and world’s most valuable company, would want to all-but acquire a 10-year-old chip startup with an entirely different hardware and software architecture to its own.

Some of these hot takes are just plain wrong. The most popular misspelled Groq as “Grok” (so close!), used ChatGPT-generated “analysis” of the difference between SRAM and DRAM, or implied the deal may have been driven largely by an acqui-hire of Groq’s team, especially Groq CEO (previously Google TPU architect) Jonathan Ross.

If only the picture were that clear. Here is a summary of what we do know and some hot takes of our own, acknowledging that frankly, there are still more questions than answers at this point.

What we do know

The wording of the announcement from both companies said that Nvidia had taken a non-exclusive license for Groq’s technology and hired key members of Groq’s team. In practice, it has become clear that Groq has effectively been gutted; what remains is a small team in place to continue running GroqCloud’s infrastructure, headed by Groq’s CFO.

We assume Nvidia has licensed all of Groq’s technology (hardware IP, chip and system design, its compiler stack, which embodied much of its secret sauce, and its higher-level software stack, too). We also assume that Groq’s hardware infrastructure and its developer community have been retained by GroqCloud.

We saw this same type of deal recently when Nvidia all-but acquired networking chip startup Enfabrica, and it seems likely that the motive is to avoid regulatory scrutiny or antitrust objections. We assume that the husk of GroqCloud will not be attempting to license any part of this technology to anyone else and that the “non-exclusive” wording of the license is to avoid antitrust claims. Nvidia got badly burned when the Arm deal fell through and would certainly wish to avoid the same scenario, though the conditions here are very different.

With Enfabrica, the rumored figure attached to the deal was $900 million; a big amount for an early-stage startup, but not unreasonable in today’s climate. Groq is a bigger, more established startup, but the rumored all-but acquisition price is an impressive $20 billion—more than 20× what Nvidia paid for Enfabrica.

The difference is an order of magnitude. The prevailing narrative on Enfabrica suggested the all-but acquisition was at least partly an acqui-hire. Enfabrica, led by Rochan Sankar and Shrijeet Mukherjee, certainly had a stellar team of very smart people. Groq CEO Jonathan Ross famously always prided himself on hiring only the best of the best, often thinking outside the box (and outside the industry) to do so. But while Groq was certainly also a team of very smart people, the $20 billion figure simply does not make sense as a deal driven primarily by an acqui-hire of a few hundred people.

For what other reasons could Groq be attractive to Nvidia?

It is not a pivot

Groq was founded in 2016 by a group of ex-Google TPU engineers, including Ross. The company built and commercialized its first-generation chip, which has been available since 2019. Longtime readers will know that six years is basically centuries in AI years. In the centuries since launching its first-gen chip, Groq pivoted to automotive (“not a pivot”) in 2020, and again to infrastructure provider (also “not a pivot”) in 2024, latterly taking on significant venture capital to build out its own infrastructure, effectively becoming a hyperscaler in its own right. This last pivot prompted many to write off Groq as a chip company, especially when it emerged that Groq had to deploy in the order of ten racks of its chips to inference a single Llama-70B model.

Nevertheless, Groq has been doing significant deals with GCC companies lately to deploy its infrastructure in the region. It has also done sovereign AI deals in other regions, including Norway.

Hardware options

Will Nvidia manufacture Groq chips, either first- or second-gen? Here are the clues we have.

In an email to employees, Nvidia CEO Jensen Huang reportedly referred directly to Groq’s chips.

“We plan to integrate Groq’s low-latency processors into the Nvidia AI factory architecture, extending the platform to serve an even broader range of AI inference and real-time workloads,” he reportedly wrote.

However, Huang confirmed in a media Q&A at CES yesterday that Groq’s technology would not become a part of Nvidia’s main data center roadmap.

“[Groq is] very, very different, and I’m not expecting anything there to replace what we do with Vera Rubin and our next generation,” Huang said. “There’s no reasonable [or] good way to do something better than Vera Rubin that we know of, and this doesn’t change that. However, we might be able to add their technology in a way that allows us to do something incremental that the world hasn’t been able to do yet.”

Note his use of the word “might”—it sounds like Nvidia may still be planning what it will do with Groq. It seems clear that Groq will not join the main Nvidia data center roadmap, but that the new technology will address some use case that Nvidia is not already addressing with Vera Rubin; specifically, real-time use cases.

Here are a few of the possible outcomes.

One possibility is that Nvidia will manufacture and deploy Groq’s SRAM-based chips as a standalone product for data center inference. Despite Huang’s reference to Nvidia’s AI factory stack, this still seems a little far-fetched, because it would mean Nvidia admitting that in some cases, GPUs are not the solution. It would mean destroying years of Nvidia messaging that GPUs are all you need, since their flexibility lends them to rapidly evolving workloads. (Nvidia CEO Jensen Huang reportedly repeated this argument once again in an analyst Q&A ahead of CES.)

Nvidia weakened its own argument slightly when it introduced the Rubin CPX—a prefill-specific GPU intended to work alongside other Nvidia big iron for decode—because it means admitting there is an argument for more specific hardware, and/or less flexible GPUs, for certain workloads. Admitting Groq has a performance advantage would be much, much worse.

Groq obviously has a software stack that works, but does it work well enough to deploy Groq Chips as a product? Sceptics say that could be part of the reason companies like Groq and Cerebras pivoted towards serving tokens for a subset of popular foundation models in the first place. So, we would assume that a non-trivial amount of software work would be required to re-launch Groq Chip as an Nvidia chip-level or system-level product.

More realistic might be a Groq chiplet alongside a big GPU chiplet to handle certain parts of the workload, but again, software would be a sticking point. Integrating Groq hardware into the CUDA ecosystem would not be straightforward. Forget CUDA kernels—Groq’s architecture does not use kernels at all.

So, what does Groq’s hardware offer that Nvidia does not have?

The answer is not simply SRAM. Nvidia GPUs use SRAM. Many types of logic chips, including most AI ASICs on the market, use SRAM. Groq does not use DRAM because it cannot. It literally does not have a DRAM interface because its chip was designed circa 2017 when big AI meant 25 million parameters, not hundreds of billions, so DRAM was not needed. Its inability to use HBM, or any DRAM, means Groq has to use racks and racks of its chips to fit a single instance of today’s mid-size models.

The answer may be single-user token latency—raw speed—but Cerebras, another SRAM-based architecture, is faster. Even with these benefits, as discussed above, the challenges of integrating Groq architecture into Nvidia’s data center hardware and software roadmaps would seem to be huge.

One left-field answer may be determinism—an interesting side-effect of Groq’s architecture it tried to push to the automotive industry in 2020. Determinism has significant implications for applications that require functional safety, including robotics (Huang refers to “real-time” applications in his email above). Previous GTC keynotes have highlighted physical AI as a huge potential market for Nvidia. Again, this would mean admitting GPUs are not good at everything, but this scenario is easier to imagine, even if it would require Nvidia to adjust its messaging to admit that edge is a very different use case.

Ultimately, Nvidia is a big company with massive resources and its own team of very smart people. If it wanted to make an entirely SRAM-based chip, it could, and if it wanted to make something that is not a GPU, it could, and it could do so for a lot less than $20 billion. It could also have all-but acquired a company like D-Matrix or even SambaNova for a lot less than $20 billion (reports suggest SambaNova’s term sheet for acquisition by Intel is in the region of $1-$2 billion).

Commercial factors

Rather than buy Groq purely for its technology or architecture, EE Times suspects that there are also significant commercial factors at play.

Groq has significant partnerships with deep-pocketed GCC companies, with big clusters already deployed in the region. It also has sovereign AI deals inside and outside the GCC, which may have looked attractive to Nvidia. That said, one of Groq’s biggest selling points so far has been that it is not Nvidia—that it is a viable and inexpensive second-source for sovereign AI infrastructure. That is likely now defunct and future sovereign buyers will be subject to Nvidia’s negotiation tactics and supply chain constraints anyway.

Nvidia’s biggest fear is that its hyperscale customers will design and build their own chips and systems, and get to a point where these in-house chips and systems are so good they do not need to rely on very expensive Nvidia GPUs to the same extent. Nvidia’s motives for all-but acquiring Groq may include a desire to stop one of these hyperscale customers from buying Groq, whether for its hardware IP or the infrastructure it has already deployed. Something at the scale and maturity of Groq’s hardware, software, and deployed assets could have made a big difference at Meta, Microsoft, or even OpenAI where in-house hardware plans are still pre-tapeout or have had modest success so far.  

Does Nvidia still truly believe that GPUs can do everything? Does this $20-billion deal mean AI is a bubble, or is not a bubble? What does that figure mean for the valuation of other startups, and new entries, in this space? What does this mean for Cerebras’ forthcoming IPO? What will happen to GroqCloud? We still have more questions than answers, and EE Times will be watching closely in 2026.