On 25 August, OpenAI published the first results from its in-house inference accelerator, which it claims is strong on speed and energy efficiency. Nvidia's top global customer is now making its own silicon. This is not a technical detail: it is a strategic move whose early signs we have been tracking for months.
Yesterday we covered Nvidia's inference accelerator entering production. Here is the same move, seen from the customer's side rather than the supplier's.
Why a lab makes its own chips
The motivation is not primarily performance. Three stronger reasons stand out.
Margin. Nvidia earns hefty margins on its accelerators. A player buying tens of billions of dollars' worth eventually works out what it would cost to make its own, even imperfectly, rather than pay top dollar.
Targeted optimisation. A general-purpose chip has to suit everyone. A chip designed for your models and your software can drop unneeded features and focus resources where they matter. This is exactly the logic we described regarding the gap between architectures: specialisation pays off when the use case is stable.
Independence. This is probably the decisive reason. Depending on a single supplier for the most critical resource in your business is an uncomfortable position, especially when that supplier also serves all your competitors and controls delivery allocations.
Two constraints we have documented explain the timing. First, the memory component shortage, which puts everyone in competition for limited output. Second, inference has become the dominant cost: when you serve hundreds of millions of daily users, every point of efficiency per token translates directly into millions of euros. An in-house chip does not need to be better than Nvidia's to be profitable. It needs to be good enough and cheaper.
What this changes in the balance of power
Nvidia finds itself in a singular position: its biggest customers are becoming its competitors. Google has had its own accelerators for years, Amazon too, and OpenAI is joining the movement.
This does not weaken Nvidia in the short term, as demand far outstrips supply. But it changes the trajectory: a supplier whose main customers are developing alternatives sees its pricing power erode gradually.
It is also worth noting the consistency with what we wrote about GLM-5.3-Flash running on Chinese chips. On both sides of the Pacific, the same move: reducing dependence on a single silicon supplier. The motivations differ — geopolitical on one side, economic on the other — and the direction is identical.
The caveats
The results are self-reported. No independent measurements are available, and a company publishing the first figures for its own hardware chooses the comparisons that flatter it.
Designing is not producing. A chip that performs well in the lab must then be manufactured at volume, in fabs whose capacity is already booked years in advance. The bottleneck is not design.
Software matters just as much. Much of Nvidia's strength lies in its software ecosystem, built up over nearly twenty years. An alternative chip is only useful if the toolchain follows.
What to take away
This announcement joins a list that is starting to form a coherent picture: twenty-year electricity leases, orbital computing projects, protected buildings, and now in-house silicon.
What all this tells us is that the AI competition has shifted to infrastructure. Models are converging, prices are falling, and what differentiates a player over the long term is its ability to produce compute more cheaply than others.
This is a heavy-industry competition, with barriers to entry closing shut. It leaves little room for new entrants, and plenty of room for those building on what others have built.