Half has started the final phase of development of its proprietary AI chips and aims to start with production in September 2026. According to Reuters, citing an internal memo, the goal is to cut spending on GPUs at a time of unprecedented component shortages.
The project is progressing at a rapid pace: at least one of the chips has just passed the testing phase six weeks. Meta takes care of the design together with Broadcombut entrusts the physical production to TSMC. RAM memory from Samsung, storage from Sandisk and fiber optic components from Sumitomo Electric complete the supply chain, demonstrating how complex the supply chain behind a single chip has become.
The program MTIA (Meta Training and Inference Accelerator) was illustrated in detail in March 2026, when Meta presented four new generations of chips intended to enter service staggered between this year and next. The architecture is based on modular chiplets, designed to allow upgrades “in months rather than years” as AI workload needs change.
Meta’s strategy to reduce dependence on Nvidia
The new chips will mainly be used for training ranking and recommendation algorithms, for larger AI workloads and for the inference of models used in the group’s apps. However, it is not a goodbye to third-party GPUs: Meta will continue to spend significant amounts with suppliers such as Nvidia and AMD, while lightening their relative weight on the accounts. The production of proprietary silicon, moreover, has been going on since 2023.
However, investments remain huge. Meta has indicated a capital expenditure in April 2026 of between 125 and 145 billion dollars for the current year, a good part of which is intended for AI infrastructures. The plan involves the deployment of 7 gigawatts of computing capacity by the end of the year, with the aim of doubling a 14 gigawatts in 2027.
Added to this are other agreements signed in recent months: an agreement with Arm for the computing power dedicated to recommendation systems, a multi-billion dollar contract with AMD for the Instinct GPUs and another agreement of the same weight with Amazon for the use of the Seattle group’s proprietary CPUs in AI workloads.
Meta is not alone in trying to stem the flow of capital to Nvidia. Last month OpenAI presented its own inference processor developed together with Broadcom, while Anthropic is considering developing its own chips with Samsung. Amazon and Google, for their part, have been designing dedicated silicon for the training and inference of their models for some time, in a market also crowded with a host of specialized startups.
