On
Meta Is Building Its Own AI Chips to Break Free From Nvidia's Grip

Meta is making a serious bet on hardware independence. The social media giant has kicked off mass production of custom-designed AI chips built entirely in-house, designed to power sprawling data centers while cutting the cord from external hardware suppliers like Nvidia.

The pace is accelerating because AI inference demands are exploding. According to Yee Jiun Song, Meta's VP of Engineering, this is exactly where the company is pouring resources right now. What's interesting here is that Meta has already invested billions building an internal chip design team—but they're not abandoning partners like Nvidia or AMD entirely. Instead, they're strategically shifting toward proprietary silicon that's been optimized specifically for Meta's workloads. The payoff? Significant energy savings and lower operational costs at scale.

Enter the Meta Training and Inference Accelerator (MTIA) program. After testing two generations—MTIA 100 and MTIA 200—Meta has now mapped out four additional chip variants: MTIA 300, MTIA 400, MTIA 450, and MTIA 500.

Meta develops custom AI chips

MTIA 300 has already crossed into mass production. This chip handles R&D tasks and serves as the foundation for future generations. Right now, it's primarily running the ranking algorithms and recommendation systems that feed content to hundreds of millions of Facebook and Instagram users daily.

Next up is MTIA 400, which upgrades support for generative AI models while maintaining backward compatibility for research and optimization work. This one scales to 72 accelerators and, by Meta's own assessment, trades blows with many commercial solutions currently on the market. The company finished testing and has begun deploying these chips across their data centers.

MTIA 450 doubles down on generative AI inference. The key improvement here is doubled HBM memory bandwidth compared to its predecessor. Meta claims this delivers significantly better performance than many existing AI processors. Expect mass production and widespread rollout in early 2027.

Rounding out the current roadmap is MTIA 500. While still focused on inference, this version cranks up HBM bandwidth another 50% over the 450 and adds critical optimizations for low-precision data processing—essential for modern AI models. It should hit production in the second half of 2027.

To pull this off, Meta partnered with Broadcom on design and adopted the open-source RISC-V architecture. Manufacturing is handled by TSMC.

Industry observers say Meta's development velocity is genuinely impressive—faster than the typical semiconductor industry cadence. That's even more striking considering Meta is a social media company, not a traditional hardware manufacturer.

The real benefit? Meta can engineer silicon that fits their specific AI workloads and ship improvements faster than relying entirely on external vendors. But here's the catch: custom chip development doesn't come cheap and demands sophisticated engineering. Meta is still dropping tens of billions on Nvidia and AMD GPUs, plus they've signed deals to lease AI accelerators from Google to cover surging compute demands.

This year alone, Meta is budgeting $115 to $135 billion for capital expenditures. Most of that goes toward expanding AI infrastructure and building new data centers to fuel the company's AI ambitions.

Related Articles