Meta's Fourth-Gen MTIA AI Chip Enters TSMC Production in September
Meta will begin manufacturing its fourth-generation MTIA AI accelerator at TSMC in September, per an internal memo, as it doubles compute to 14 GW by 2027.
Key Takeaways
Meta will begin volume production of its fourth-generation MTIA AI accelerator at TSMC in September, after the chip cleared validation in about six weeks, according to an internal memo cited by Reuters.
Broadcom collaborated on the chip's physical design and interface architecture, with Samsung supplying high-bandwidth memory, Sandisk providing storage, and Sumitomo Electric supplying fiber-optic gear.
The chip will train and serve Meta's ranking and recommendation systems and handle inference for consumer apps, though Meta will still buy Nvidia H- and B-series GPUs and AMD Instinct accelerators for the most demanding training runs.
Meta plans $125-145 billion in capital expenditures this year, targeting 7 gigawatts of AI compute online in 2026 and 14 gigawatts by 2027, including a CA$13 billion Alberta data centre.
The move intensifies the custom silicon race, alongside OpenAI's Broadcom-built inference chip, Anthropic's reported Samsung talks, Amazon's Trainium, Google's TPUs, and Meta's own $6.5 billion Samsung Foundry 2nm deal for a future MTIA generation.
Kaan Tınmaz
Meta will begin manufacturing the latest generation of its in-house AI accelerator in September, according to an internal memo cited by Reuters, in a push to cut its dependence on Nvidia and AMD GPUs at the same time the company is doubling its compute footprint.
Fourth-Gen MTIA Enters The Fab
The new chip is the fourth generation of the Meta Training and Inference Accelerator, the MTIA family Meta first detailed publicly in 2023. It sailed through validation in about six weeks, the memo says, clearing the way for volume production later this year. Meta collaborated with Broadcom on the physical design and interface architecture, while Taiwan's TSMC will manufacture the chip on its advanced process node. Meta is also sourcing high-bandwidth memory from Samsung, storage from Sandisk and fiber-optic gear from Sumitomo Electric, the report says.
A Modular Answer To GPU Scarcity
Meta described the current MTIA roadmap in March, framing it as a modular chiplet approach that lets the company refresh generations on a shorter cadence as AI workloads evolve. The company will use the fourth-gen MTIA to train and serve its ranking and recommendation systems, along with broader AI workloads and inference for its consumer apps, though it still plans to buy large volumes of Nvidia H- and B-series GPUs and AMD Instinct accelerators for the most demanding training runs.
Chasing 14 Gigawatts By 2027
Meta has warned investors it will spend between $125 billion and $145 billion on capital expenditures this year, most of it on AI infrastructure. It plans to bring 7 gigawatts of AI compute online in 2026 and double that to 14 gigawatts in 2027, deals for which include a CA$13 billion Alberta data centre and a multi-billion-dollar CPU purchase with Amazon Web Services.
Custom Silicon Race Intensifies
Meta is not alone. OpenAI last month unveiled its first inference chip built with Broadcom, Anthropic is reportedly in early talks with Samsung on custom silicon, and Amazon and Google continue to iterate on Trainium and TPU generations. Combined with Meta's earlier $6.5 billion deal to use Samsung Foundry's 2nm process for a future MTIA generation, the September start of Iris production tightens the vise on Nvidia's grip over hyperscaler AI budgets.
Reporting based on coverage from TechCrunch and Reuters.
It is the latest version of the Meta Training and Inference Accelerator, Meta's in-house AI chip family first detailed in 2023, built on a modular chiplet approach that allows faster generational refreshes as AI workloads evolve.
When and where will the new MTIA chip be manufactured?
Production begins in September at TSMC in Taiwan on an advanced process node, following a roughly six-week validation period, per an internal Meta memo cited by Reuters.
Will Meta stop buying Nvidia and AMD GPUs?
No. While the MTIA chip is meant to reduce dependence on external suppliers, Meta still plans to purchase large volumes of Nvidia H- and B-series GPUs and AMD Instinct accelerators for its most demanding training runs.
How much is Meta investing in AI infrastructure?
Meta expects $125 billion to $145 billion in capital expenditures this year, mostly on AI infrastructure, with plans to bring 7 gigawatts of AI compute online in 2026 and double that to 14 gigawatts in 2027.
