Meta is preparing to begin production of its latest custom artificial intelligence (AI) chip in September as the company pushes to build more of its AI infrastructure with hardware it designs itself.
The chip, known internally as Iris, marks the next stage of Meta’s long-running effort to reduce its dependence on Nvidia as it spends as much as $145 billion on AI infrastructure this year. Reuters first reported the plans after reviewing an internal company memo.
The new chip is part of Meta Training and Inference Accelerators (MTIA), a family of AI chips built specifically for the company’s data centers. Meta is using them to power the AI systems behind Facebook, Instagram, and its wider portfolio of AI products. The company is working with Broadcom on the chip’s design while Taiwan Semiconductor Manufacturing Company will manufacture it.
A Shift Away From Buying Every AI Chip
Meta remains one of Nvidia’s biggest customers, but the scale of its AI ambitions has made relying almost entirely on outside suppliers increasingly expensive.
According to the internal memo seen by Reuters, Meta expects to deploy seven gigawatts of computing capacity by the end of this year before doubling that to 14 gigawatts in 2027. Supporting infrastructure on that scale requires enormous numbers of AI chips, networking equipment, memory, and storage.
Meta has also reportedly signed long-term supply agreements with companies including Samsung, Sandisk, and Sumitomo Electric to secure key components for its growing data center network.
Building more of its own chips gives Meta greater control over costs and allows it to tailor hardware for its own AI workloads instead of depending entirely on general purpose GPUs.
Why Meta Is Building Its Own Chips
Meta’s push into custom chips comes as the company prepares to spend as much as $145 billion on AI infrastructure this year.
This investment covers new data centers, networking equipment, AI chips, and the computing capacity needed to train and run its AI models. Meta is also looking for ways to generate returns from that massive buildout.
Earlier this month, it was reported that the company is developing a cloud infrastructure business that would let outside customers buy access to its excess AI computing capacity and AI models.
As such, Meta bringing Iris into production is expected to help Meta reduce the cost of building its AI infrastructure while giving the company greater control over the hardware that powers its growing AI operations.
Iris Is Part Of A Bigger Roadmap
Iris is not a one-off project. It belongs to a four-generation MTIA roadmap that Meta plans to develop in-house, with a new chip generation expected roughly every six months through 2027. The social media giant said earlier this year that the chips follow a modular design approach, allowing engineers to improve them more quickly as AI workloads evolve.
The latest development also represents progress for a program that has faced setbacks over the years. Iris reportedly completed testing in about six weeks without major issues, suggesting Meta’s custom silicon effort is now moving more smoothly than in previous generations.
Nvidia Is Still Part Of Meta’s AI Strategy
This Meta’s move does not mean it is abandoning Nvidia. The company will continue buying large numbers of Nvidia and AMD GPUs because its custom chips are designed to complement existing AI infrastructure rather than replace it completely. In this sense, Iris is intended to work alongside those GPUs as Meta expands its computing capacity.
And as AI infrastructure spending continues to climb across the technology industry, companies are looking for more control over the hardware that powers their models. For Meta, producing Iris is about supporting its growing AI operations while reducing the long-term cost of building them at an unprecedented scale.
