🔴 Has Nvidia Already Won?

Last newsletter of the year, to talk about the deal of the year.

First of all, thank you for your loyalty!

2025 will go down as an absolutely incredible year. The last few newsletters brought in a flood of paid subscriptions, which reinforces my sense that this newsletter is finding its place in a French ecosystem that has a lot of questions about the future.

2026 will not be easy, but we are going to navigate it together! More on that later.

If you are not yet a paying subscriber, subscribe!

Since launch, I have chosen to keep the price very low. Some of these notes are worth several times the price of a subscription on their own. A consultant told me recently that a single one of our newsletters had landed him a year-long contract.

Keeping the price affordable also reflects a desire to reach audiences beyond my initial target: students in particular, but also small businesses, and anyone who wants to think these issues through with us.

💸 Please note: the subscription price will go up before the end of the year.
So I wanted to offer you the option, if you wish, of keeping the current rate by subscribing before the increase. I know this is not a commercially intuitive approach.

• If you are already subscribed at the €60 rate, as a founding member, that rate will not change when you renew
• If you are subscribed at the €99 rate, as a co-founding member, and you subscribe now, that rate will not change either

ℹ️ As a paying subscriber, you get access:

• to the site's complete archives,
• to everything I track, through a private channel,
• to priority replies to your emails,
• access to the events we organize, invitations, ...
• to the meetups held on the sidelines of the conferences where I speak.

The “perks” are listed here

Does the Groq-Nvidia deal matter?

As we said before, there is almost certainly an AI infrastructure bubble that will deflate at some point, but for now the race against the clock is not slowing down, it is accelerating. We had already seen that the cross-deals between Nvidia, Google and Amazon are also deals for access to their AI chips, and a guarantee that those chips will be used in 2026 (since Anthropic and OpenAI were lent the money to buy them).

But we should not forget the battle among the leaders to hold on to their positions. Nvidia, it seems, is the best at this game. As we will see, the $20 billion agreement between Nvidia and Groq that has just been struck is not a conventional acquisition.

Presented as a simple non-exclusive licensing contract, it is in reality a reverse acquihire: a targeted extraction of talent, intellectual property and strategic positioning, while maintaining the illusion that the company remains independent.

This kind of operation, with no official buyout, lets Nvidia neutralize a threat without drawing regulators' attention. It follows a series of similar moves by Microsoft (Inflection), Google (Character), Amazon (Adept) and Meta (Scale AI), all aimed at defusing competitors before they become too powerful.

Groq was the only alternative player already deployed in critical sovereign environments, notably in universities working for the Department of Defense. Its disappearance as an independent player marks a turning point in the consolidation of the world's AI infrastructure.

First, you need to understand how the deal was structured:

• Jonathan, the CEO, along with the president and Groq's key technical teams, have joined Nvidia.
• The core technology, including the Language Processing Unit (LPU), was transferred through a non-exclusive license.
• Groq keeps a legal existence, a new CEO and a minimal level of activity, but without its strategic core. This is often called a zombie startup.

The appeal of this method is that it avoids any regulatory alarm tied to a formal acquisition.

I had seen Jonathan, Groq's founder (and the inventor of the TPU at Google), again at the first edition of Raise.ai, and listening to him I came to understand how his company represented a long-term strategic threat to Nvidia.

• Unlike conventional GPUs, Groq's LPU integrates compute and memory in a single coherent architecture, allowing fast, predictable execution.
• This system offered a direct alternative to Nvidia's CUDA model, with gains in speed, stability and power consumption.
• It was particularly well suited to real-time inference workloads, with deployments already in place at critical institutions such as the Argonne, Sandia and Livermore laboratories, all of which work for the defense establishment and need sovereign AI.
• Had adoption kept growing, it would have allowed other players to do without Nvidia's software ecosystem, something the sector's leader cannot afford.

Jensen, the big boss, therefore acted fast, following an industrial playbook that is now well rehearsed and has already been run several times in AI:

• Microsoft (Inflection) brought in the founders and the teams. The company abandoned the consumer market to refocus on an enterprise offering, with a modest return for investors.
• Google (Character AI) signed an agreement to bring back its co-founders (who had not been listened to at Google back in the day, a sweet revenge). The company stopped developing LLMs to concentrate on its platform. The agreement is now under review by antitrust authorities.
• Amazon (Adept) brought in the executives and refocused the company on specialized AI solutions for businesses.
• Meta (Scale AI) followed a similar logic with its CEO, Alexandr Wang, who took over AI, pushing Yann LeCun out the door.

Curiously, Yann LeCun and Alexandre Lebrun seem to be using a similar strategy to structure a fundraising round: despite a recent $70 million raise, Nabla is seeing its founder join AMI, the new startup launched by LeCun, through a cross-deal.

In every case the model is the same: absorb the potential, hollow out the startup, and sidestep regulatory constraints along the way. For now, it seems to be working.

Groq was not alone, but it was ahead

Several other players are pursuing alternative approaches:

• Cerebras (founded by Andrew Feldman): an original large-scale architecture, but geared toward training rather than fast inference. Cerebras appears to have a collaboration agreement with Mistral AI.
• Tenstorrent (led by Jim Keller, formerly of Apple, AMD, Intel and Tesla): a strong technical team, but weak commercial traction and a positioning closer to processors (CPUs) than to AI accelerators.
• Graphcore (co-founded by Nigel Toon and Simon Knowles): in sharp decline after a promising start; has lost market share and strategic investors.
• Etched (founded by Gavin Uberti, Chris Zhong and Ben Solomon): still in development, focused on a narrow use case (accelerating static transformers).

Only Groq had a fully operational product, already adopted in strategic settings, with a genuinely distinct, inference-oriented architecture.

Because the real battle that is brewing is not a hardware battle but a software one:

• The entire AI industry today rests on CUDA, Nvidia's software platform.
• Millions of developers are trained on it, and tools like PyTorch are deeply integrated with it.
• Leaving that ecosystem costs millions of dollars and months of engineering.
• Groq made it possible to get around that lock by delivering results through an API, with no heavy software rewrite.

Even an excellent chip is not enough without compatible software. Groq sidestepped that dependency. Google, with its TPU, is still far from convincing developers that migration is simple.

But there is another reason Nvidia bought Groq, and it is one the company cannot admit publicly. (the rest of this analysis is for paying subscribers)