Nvidia is partnering with MediaTek in a deal reported to be worth around $3.5 billion, presented as a response to big tech companies increasingly developing their own AI chips. It is the company's third defensive move in a fortnight.
We have covered back-to-back OpenAI's in-house chip and Nvidia's acquisition of Hugging Face. This announcement completes the picture, and the picture is coherent.
Nvidia's singular position
Let's recall the problem this company has to solve: its biggest customers are becoming its competitors.
Google has been designing its own accelerators for years, Amazon too, OpenAI has just published the first results from its own, and we have seen that GLM-5.3-Flash would run entirely on Chinese silicon.
None of these players will leave Nvidia overnight: demand exceeds supply, and the software ecosystem built over twenty years remains a considerable advantage. But the trajectory is clear, and a company whose main customers are insourcing sees its pricing power erode.
The logic behind the three moves
Taken together, the recent moves sketch out a readable strategy.
Securing the tooling. By buying the platform where developers discover and deploy models, you retain them through tools rather than hardware.
Occupying the inference segment. The dedicated accelerator stemming from the Groq acquisition addresses the fact that serving a model and training one do not require the same qualities.
Extending coverage. A partnership with a manufacturer positioned in other segments — notably chips for devices and consumer markets — makes it possible to occupy spaces that the big labs' in-house chips do not target.
This last point deserves emphasis, because it ties into a trend we are tracking. Running a model locally is becoming realistic, as shown by Meta's agent on a single card or Qwen 3.8 in 27 billion. And we reported on Tuesday that half of the Fortune 500 would favour in-house hosting. If inference shifts towards devices and enterprise servers, the chip market ceases to be concentrated in a few giant data centers. It becomes a more fragmented market, where the game is played differently.
What this says about the cycle
There is a broader reading, and it is a classic one in industrial history.
A dominant supplier of a critical component always sees its customers seek to reduce their dependence. It happened in automotive, in telecoms, in computing. The usual response is to move: either upstream, or towards software layers, or towards adjacent markets.
Nvidia is doing all three simultaneously, and with considerable resources. That guarantees nothing, and it explains why a company that dominates its market spends so much to defend it: dominance is more fragile than it looks when it rests on customers who have the means to replace you.
What to take away
For an observer, these moves are a good indicator of where power in this industry really lies.
Two years ago, the question was "who has the best model". Today, models are converging and being given away for free, while three companies spend billions to control chips, platforms and tooling chains.
This is what we wrote about AI becoming an industry: competition has shifted towards infrastructure, with barriers to entry closing. For everyone else, the practical conclusion remains the same — build on what these players make available, without depending on any single one of them.