Meta readies homegrown AI chips for mass production as it races to cut Nvidia costs

Meta is gearing up to launch production of its latest generation of custom AI chips starting in September, according to an internal memo obtained by Reuters. The move reflects the company's strategy to reduce costs as demand for computing hardware continues to outpace supply across the AI industry.

Development and design process#

The tech giant is designing the chips through a partnership with Broadcom while enlisting Taiwan's TSMC to handle manufacturing. Additional components—including RAM from Samsung, storage from SanDisk, and fiber optic infrastructure from Sumitomo Electric—round out the supply chain.

Notably, at least one prototype cleared its validation testing in roughly six weeks, suggesting an accelerated timeline for getting the hardware into production.

The MTIA program evolution#

These chips stem from Meta's Meta Training and Inference Accelerator (MTIA) initiative, which the company introduced publicly in March with a roadmap covering four new generations. The modular architecture allows Meta to adapt designs as AI workloads shift, with each iteration incorporating fresh insights and emerging hardware technologies.

"Each MTIA generation builds on the last, using modular chiplets, incorporating the latest AI workload insights and hardware technologies, and deploying on a shorter cadence," the company said when announcing the program.

Purpose and deployment strategy#

Meta plans to deploy these accelerators primarily for ranking systems, recommendation algorithms, and AI inference across its applications. The chips are designed to complement—not replace—purchases from vendors like Nvidia and AMD, though they should meaningfully reduce spending on external GPUs. The company has been developing in-house silicon since 2023.

Meta's broader compute ambitions#

The production ramp aligns with Meta's massive infrastructure buildout. The company projected capital spending between $125 billion and $145 billion in 2026, with a substantial portion supporting AI operations. It aims to deploy 7 gigawatts of compute this year and plans to double that capacity next year.

Beyond custom silicon, Meta has sealed major deals across the industry: partnerships with ARM, a multi-billion dollar arrangement with AMD for Instinct GPUs, and another major agreement with Amazon for its homegrown AI CPUs. This diversification strategy reflects competitive pressure to secure computing resources in a constrained market.