Truly Irrational: Nvidia/Groq Deal
What the actual fuck?
Irrational Analysis is heavily invested in the semiconductor industry.
Positions will change over time and are regularly updated.
Opinions are authors own and do not represent past, present, and/or future employers.
All content published on this newsletter is based on public information and independent research conducted since 2011.
This newsletter is not financial advice, and readers should always do their own research before investing in any security.
Feel free to contact me via email at: irrational_analysis@proton.me
First, let’s go over the deal structure.
Nvidia pays Groq $20B for a non-exclusive technology license.
Nvidia hires everyone important from Groq.
GroqCloud continues to operate.
Put on your thinking cap for a moment.
Suppose you are someone else who also wants to license Groq IP.
Where is the technical support coming from?
Is there a roadmap or are you buying IP with no development future?
Acquisitions require regulatory approval. This is a licensing deal, totally different but the outcome might be similar if you have more than two brain cells and think for a moment.
Groq is NOT a TPU clone. It is not a normal ASIC.
The architecture is very different. 144-wide VLIW.
Wrote a detailed post on this last year.
![[V]ery [L]ong [I]ncoherent [W]riteup](https://substackcdn.com/image/fetch/$s_!lVhT!,w_140,h_140,c_fill,f_webp,q_auto:good,fl_progressive:steep,g_auto/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1335e244-45bb-4add-a677-d9ab1ed74702_875x957.png)
Need to slightly update the thumbnail for this post…
Google TPU is 8-wide VLIW for control and normal architecture for the compute. Includes normal memory hierarchy.
Groq is a 144-wide VLIW for EVERYTHING and has ZERO memory. Only SRAM. It’s crazy and I hate it.
Genuinely don’t understand why Nvidia made this deal. Enfabrica made sense. This does not. Must be missing something.
Scratching my head as well! That’s too high to be an Acquihire kind of thing and the company is apparently going to continue to exist in some form.
It will be important for the AI semi ecosystem to figure out just what the heck is happening
Edit: On thinking a bit more, one way this could make sense is that there is some key implementation IP that Groq has which nVidia cannot workaround for an inference optimized alt. arch. solution. nVidia has a team that has been looking at alternative non-GPU like architectures. This would also explain licensing and getting the brains to make it really happen inside nVidia. Clearly, they think that despite others having the same license it's not the same at the actual implementation microarchitecture level as exemplarily demonstrated by Apple. Nobody else managed an out of order high performance ARM CPU that could actually beat x86 till Apple showed the world that it can be done. Some similar dynamics apply here. They have the scale, the customers, the money, the installed base, the supply chain/wafer capacity lock-in and the cumulative experience to make it happen.
Some guesses:
1. Jensen Huang knows capturing value from exploding robotics market comes from cost competitiveness in the lower latency range of the interactivity-throughput curve. Groq currently is not cost competitive (even at the very low latencies demoed in GroqCloud with artificially low pricing via rate limits) but their architecture is oriented towards this part of the curve. Perhaps next-gen is big uplift here, validating real structural advantage vs Nvidia architecture for everyone to see (and vs TPU which has similar degradation at low concurrency), which would be big threat to CUDA in robotics. So Nvidia wants to own this tech and/or prevent competing ecosystem traction while it works out its own CUDA solution.
2. Sacks did a favour for his friend and offered to Jensen that he would convince Trump admin to lift H200 sanctions in return for Nvidia bailing out Chamath’s position in Groq.