According to reports from mid-July 2026, Meta is reportedly planning to start manufacturing a chip designed specifically for its artificial intelligence needs as early as September. Stated goal: cut costs and stop relying entirely on Nvidia and AMD, as components grow scarcer and their prices climb. Meta is not alone: this is now the industry-wide trend.
There is one company that wins almost every time in the AI revolution, regardless of how the battle between labs plays out: the one selling shovels during the gold rush. Nvidia has held that position for years, and it brings in staggering sums. The problem, when you are a customer of that company, is that its margin comes straight out of your pocket. Hence the discreet but massive shift we are observing in 2026.
What Meta is planning
According to information reported in mid-July, Meta is reportedly preparing to launch production of a new chip dedicated to its AI workloads as early as September. The logic is twofold. First, a cost logic: designing your own chip lets you escape the supplier's margin, which becomes profitable once you buy in very large volumes. Second, a supply logic: cutting-edge AI components are becoming scarce and increasingly expensive, and depending on two suppliers for such a critical resource is a strategic vulnerability.
The context gives a sense of the scale of the need. Meta is estimated to have a fleet of several hundred thousand graphics processors, and its next flagship model would reportedly be trained on the equivalent of around one million chips. At that level, every percentage point saved on hardware represents hundreds of millions of dollars.
Meta is late, not early
It is worth clarifying to put this news in context: Meta is not setting a trend, it is catching a train already in motion. Google has been designing its own processors for AI for years. Amazon has developed its in-house range for training and inference, used by several labs. Microsoft launched its own inference-dedicated chip in January 2026, fabbed at 3 nanometres, which claims over 30% more performance per additional dollar compared with competing solutions.
This silicon battle even has direct consequences for alliances between players. Anthropic, for example, is reportedly in preliminary discussions with Microsoft to run part of its computations on these new in-house chips via the publisher's cloud. That would be consistent with its diversification strategy: the company already spreads its needs across Nvidia cards, Amazon's chips and Google's, limiting its dependence on any single supplier.
An Nvidia graphics processor is designed to do many things adequately. An in-house chip can be tailored to do one thing perfectly: exactly the operations your model needs, in the format you use. It is the difference between a Swiss Army knife and a bespoke tool. You lose versatility, but on your specific workload, you gain performance and, above all, cost per operation. At the scale of billions of requests, that gap becomes decisive.
The constraint no one can bypass
There is, however, a limit this movement does not solve. Designing a chip is not the same as manufacturing it. The giants design their processors, but actual fabrication remains concentrated in a handful of foundries, mainly in Taiwan. The physical bottleneck remains, and it stays geographically and politically sensitive.
That is what makes global competition on this front so tense, as we discussed in our article on the compute race. Control of silicon has become an instrument of national power, and export restrictions are its most direct expression. Notably on the other side: the release of Kimi K3 showed that Chinese labs are now building world-class models while working around these hardware constraints.
What to take away
Meta's move illustrates a silent shift in the AI economy. After the race for models and the race for data centers, the race for bespoke silicon is now opening up. The companies that consume the most compute are all seeking to internalise their hardware, for the same reason a carmaker eventually wants to build its own engines: beyond a certain scale, depending on a supplier becomes more expensive than doing it yourself.
For the end user, the effect will likely be invisible but beneficial. More custom-designed chips means more available capacity, less pressure on prices, and ultimately AI costs trending downward. The real potential loser in this story would be the one selling the shovels, watching its best customers learn, one by one, to make them themselves.