Via browseract.com
Four new generations of in-house AI chips and a Broadcom partnership signal Meta's intent to reduce its dependence on Nvidia's GPUs.
Meta isn’t just buying Nvidia’s chips anymore. It’s building its own.
The company unveiled four new generations of its Meta Training and Inference Accelerator (MTIA) chip family on March 11, 2026: the MTIA 300, 400, 450, and 500. Each is designed for specific AI workloads that Nvidia’s general-purpose GPUs handle today, but at lower cost and with tighter integration into Meta’s sprawling infrastructure.
The silicon strategy takes shape #
The MTIA lineup uses a modular chiplet design that lets Meta iterate roughly every six months, far faster than the traditional GPU development cycle.
The earlier MTIA generations, the 100 and 200 series, have already been deployed in production. Hundreds of thousands of those chips are running inside Meta’s data centers, handling ranking and recommendation workloads that power everything from your Instagram feed to Facebook’s ad-targeting engine. Some have even been tested with Meta’s Llama large language models.
The newer MTIA 450 and 500 variants will push further into generative AI inference, featuring improved high-bandwidth memory (HBM) to handle the data throughput that transformer-based models demand.
An internal memo surfaced in July 2026 revealing that production of the latest MTIA chip, code-named “Iris,” would begin in September 2026 following six weeks of successful testing.
Broadcom enters the picture #
Meta isn’t going it alone. On April 14, 2026, the company formalized an expanded partnership with Broadcom to co-develop MTIA chips through at least 2029.
Meta’s compute appetite is staggering. The company plans to scale its computing capacity from 7 gigawatts in 2026 to 14 gigawatts by 2027.
What this means for Nvidia #
The important caveat: Meta isn’t dumping Nvidia. The company continues to make significant GPU purchases from both Nvidia and AMD. Custom MTIA chips are positioned as complements, not replacements, targeting inference workloads where specialized hardware can deliver better performance per watt than a general-purpose GPU.
Meta isn’t the only tech giant thinking this way. Google has its TPU chips, now in their sixth generation. Amazon has Trainium and Inferentia for AWS customers. Microsoft has Maia.
Nvidia’s moat has always been CUDA, its proprietary software ecosystem that makes it painful for developers to switch to competing hardware.
What to watch next: whether Meta begins deploying MTIA chips for training workloads, not just inference. That would represent a far more direct challenge to Nvidia’s core business, and it’s exactly where the MTIA 500’s improved HBM bandwidth could prove decisive.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our