Meta to put its own AI chip into production in September, aiming to double computing capacity

Meta plans to put its own artificial intelligence chip into production in September and is aiming to roughly double the computing capacity across its data centres, Reuters reported on Thursday, citing people familiar with the matter.

The chip belongs to Meta’s in-house silicon line, the Meta Training and Inference Accelerator, or MTIA, which the company has been scaling up as part of a record spending push and a wider effort to reduce its reliance on Nvidia.

Meta declined to comment on the specifics, and both the September timing and the capacity target come from anonymous sources rather than any public disclosure.

If the timeline holds, it would mark another step in a programme that has moved unusually fast this year. Meta unveiled four new MTIA chips in March, the 300, 400, 450, and 500, and said it would ship them on a roughly six-month cadence rather than the annual pace common across the industry.

Those chips are manufactured by TSMC and co-developed with Broadcom, whose partnership with Meta now runs through 2029 and covers several generations of custom silicon. Broadcom has said the newer MTIA parts will be among the first custom AI chips built on a 2-nanometre process.

The strategic logic is straightforward. Meta remains one of Nvidia’s biggest customers, buying vast numbers of GPUs to train its Llama models and to run recommendation systems for more than 3 billion daily users, and every workload it can move onto its own chips is one it does not have to buy at Nvidia’s margins.

For now, MTIA has largely handled inference, the day-to-day job of serving predictions once a model has been trained. The MTIA 300 is already in production for ranking and recommendation work, while the 450 and 500, aimed at generative image and video inference, are slated for mass deployment through 2027.

A training-capable chip would be a harder test. Training frontier models is the workload where Nvidia’s hardware and its CUDA software have proved stickiest, and where in-house alternatives from Google and Amazon have taken years to mature.

The capacity ambition sits inside an enormous spending plan. Meta has guided 2026 capital expenditure to between $125bn and $145bn, with nearly all of the increase going towards data centres, GPUs, and custom silicon, and Mark Zuckerberg has floated eventual targets measured in gigawatts.

That build-out has grown large enough that Meta is now looking to rent out spare compute to outside customers, echoing a model long used by cloud providers. The company has also hedged its bets on suppliers, signing deals for Amazon’s Graviton5 chips and AMD accelerators alongside its standing Nvidia orders.

Custom silicon is central to that hedge because it changes the underlying economics. Designing a chip for exactly the models Meta runs, rather than buying a general-purpose GPU, can cut power draw and unit costs at the scale the company operates, provided the software stack keeps pace.

The catch is that in-house chips rarely displace Nvidia outright. Analysts tend to frame MTIA as a way to absorb growth and trim the GPU bill at the margins, not to replace Nvidia in the near term, and Meta itself keeps expanding its GPU commitments even as it ramps up custom parts.

Reuters’ report does not specify which MTIA generation enters production in September, nor how any doubling of capacity would be split between new chips and additional data-centre floor space. Meta has not published a figure that matches the framing.

What is clearer is the direction of travel. After years of experiments, Meta’s silicon effort has shifted from a side project into a core plank of its infrastructure strategy, and September, if the reporting proves accurate, would be the next milestone to watch. The bigger question is whether the chips can eventually reach the training workloads that still belong almost entirely to Nvidia.