Meta’s New AI Chips Will Begin Production in September

As the global race for affordable and reliable artificial intelligence infrastructure intensifies, Meta is moving forward with its plans to reduce reliance on third‑party hardware suppliers. According to an internal memo seen by Reuters, the company is set to begin mass production of its latest generation of custom‑built AI chips this coming September, marking a major milestone in its effort to gain greater control over its computing needs and reduce soaring hardware costs.

Development and Manufacturing Details

The memo reveals that at least one variant of the new chip completed its testing phase in just six weeks, a relatively fast timeline for complex semiconductor development. The design is being developed in close collaboration with Broadcom, while manufacturing will be handled by Taiwan Semiconductor Manufacturing Company (TSMC), the world’s leading contract chipmaker.

To complete the full hardware stack, Meta has secured supply agreements with key component makers: Samsung will provide high‑performance memory chips, SanDisk will supply storage solutions, and Sumitomo Electric will deliver fiber‑optic networking equipment essential for connecting large clusters of chips within data centers.

These new processors are part of Meta’s Meta Training and Inference Accelerator (MTIA) program, first unveiled in March this year. The company introduced four distinct chip designs under this initiative, with some already being deployed in limited environments and others scheduled for rollout through 2026 and 2027.

Meta has adopted a modular chiplet architecture for the MTIA lineup, a flexible approach that allows individual components to be upgraded or replaced independently as AI technology evolves. “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 explained when announcing the program. This strategy is designed to future‑proof its hardware, ensuring it can adapt quickly to changing model requirements without needing a complete redesign every time.

Purpose and Cost‑Saving Strategy

The main driver behind this initiative is to ease the financial pressure caused by the global shortage of high‑end graphics processing units (GPUs) and their rapidly rising prices. For years, Meta, like most large AI operators, has relied heavily on chips from Nvidia and AMD, which remain the industry standard for training and running complex AI models.

The new MTIA chips will be used to power specific workloads across Meta’s ecosystem, including training its recommendation and ranking algorithms, running broader artificial intelligence tasks, and handling inference, the process of using trained models to generate responses and content for its apps. While the company aims to reduce its dependence on external suppliers, it has clarified that it will continue to purchase significant volumes of GPUs from Nvidia and AMD alongside its own hardware, as certain advanced workloads still require the maximum performance offered by those chips.

Meta began producing its own AI chips in 2023, and this latest generation is the most capable and scalable to date.

Massive Investment in AI Infrastructure

The development of custom chips is just one part of a much larger and more expensive expansion plan. In April, Meta announced it expects capital expenditures for 2026 to reach between $125 billion and $145 billion, with the vast majority of that budget allocated toward building and equipping its AI infrastructure.

According to the internal memo, the company is securing power supply and data center capacity around the world to support its growing computing demands. It plans to deploy 7 gigawatts of dedicated compute power this year alone, with a target to double that capacity to 14 gigawatts by 2027, enough energy to power millions of homes, highlighting the scale of its ambition.

To fill gaps while its own hardware ramps up, Meta has also secured several large‑scale partnerships: it signed a long‑term agreement with Arm last year to support its recommendation systems, a multibillion‑dollar deal with AMD for its Instinct series GPUs, and another major contract with Amazon Web Services to use the cloud provider’s custom‑designed CPUs for additional AI workloads.

Part of a Broader Industry Shift

Meta is far from alone in this move. The high cost and limited supply of top‑tier AI chips have pushed nearly every major technology firm to explore building its own hardware. OpenAI recently unveiled its first custom inference processor, also developed in partnership with Broadcom. Anthropic is reportedly evaluating a joint chip‑development project with Samsung. Meanwhile, Google and Amazon have been producing their own AI‑optimized chips for years, and a growing number of startups are entering the market to offer alternatives to the dominant suppliers.

This trend marks a fundamental shift in the industry: as artificial intelligence becomes central to business operations, companies are no longer willing to rely entirely on a small handful of suppliers. Instead, they are building their own hardware ecosystems to control costs, ensure supply security, and tailor performance exactly to their specific software needs.

When contacted for comment, Meta declined to provide additional details beyond what was already disclosed.