Meta unveiled four new generations of its in-house AI accelerator and extended a chip-design partnership with Broadcom through at least 2029, pushing deeper into custom silicon to cut its reliance on Nvidia's GPUs. Meta still makes significant GPU purchases from both Nvidia and AMD, but the new chips target inference workloads where tailored hardware can beat a general-purpose GPU on cost and power. Nvidia's moat has always been CUDA, its software ecosystem.
Meta rolled out four new generations of its Meta Training and Inference Accelerator on March 11, 2026: the MTIA 300, 400, 450, and 500. Each targets AI workloads that Nvidia's general-purpose GPUs handle today, but at lower cost and with tighter integration into Meta's infrastructure. The rivalry could reshape industry economics if each company ends up dominating a different layer of the market rather than one displacing the other.
Chips built for Meta's own workloads
The chiplet-based MTIA design lets Meta iterate roughly every six months, faster than a typical GPU development cycle. Earlier MTIA generations, the 100 and 200 series, already run in production, with hundreds of thousands of those chips handling ranking and recommendation work behind Instagram's feed and Facebook's ad-targeting engine. Some have even been tested against Meta's Llama models.
Broadcom deal extends to 2029
An internal memo that surfaced in July 2026 revealed that production of Meta's newest chip, code-named Iris, would begin in September 2026 after six weeks of successful testing. Meta also formalized an expanded partnership with Broadcom on April 14, 2026 to co-develop MTIA chips through at least 2029. Its compute needs keep climbing: the company plans to scale computing capacity from 7 gigawatts in 2026 to 14 gigawatts by 2027.
Nvidia's moat still holds, for now
Meta isn't cutting Nvidia loose. It continues to make significant GPU purchases from both Nvidia and AMD, positioning MTIA chips as complements for inference rather than outright replacements. Google, Amazon, and Microsoft are chasing the same strategy with their own TPU, Trainium and Inferentia, and Maia chips.
Nvidia's advantage has always rested on CUDA, the software layer that makes switching away from its hardware painful for developers.
Sources: Fool (snippet-based), Crypto Briefing
Trading involves risk.
Risk Disclaimer
