Meta signs $6.5B Samsung foundry deal for 2nm third-gen MTIA chips

Meta has reportedly finalized a $6.5 billion agreement with Samsung Foundry to manufacture its third-generation MTIA (Meta Training and Inference Accelerator) chips on a 2-nanometer process, according to Calcalist. The deal notably shifts a major chunk of Meta's advanced-node production away from TSMC to Samsung, a rare win for Samsung's foundry business as it fights to close the gap with the Taiwanese leader.

The strategic logic is vertical integration and NVIDIA de-risking: custom silicon lets Meta control cost and supply for its inference-heavy workloads rather than paying NVIDIA's premium and competing for scarce GPUs. It directly supports Zuckerberg's stated goal of reaching 5 gigawatts of computing capacity by 2030 and underpins the Meta Compute cloud ambitions and frontier models like Watermelon.

For Samsung, the win compounds a banner quarter (an expected 18-fold profit jump on AI memory) and validates its 2nm process as competitive for a hyperscaler's flagship accelerator. It also intensifies the foundry battle just as demand for leading-edge AI silicon far outstrips supply.

Caveats: MTIA generations have historically targeted Meta's internal recommendation and inference workloads rather than general training, so this is unlikely to fully displace NVIDIA in the near term. What to watch: yield and volume ramp on Samsung 2nm, and whether Meta offers MTIA-backed capacity externally through Meta Compute.