Four MTIA Chips in Two Years: Scaling AI Experiences for Billions

Meta has been quietly building its own AI chips for years, but the company just laid out how aggressively it's speeding up that effort. After publishing research on two earlier chip generations, MTIA 100 and 200, at academic conferences in 2023 and 2025, Meta says it has already moved through four more designs: MTIA 300, 400, 450, and 500. Some are already running in production, others are slated for 2026 or 2027, and the whole arc covers everything from the recommendation systems that quietly power Meta's apps to the generative AI models now driving chatbots and assistants.

What's notable is the reasoning behind the speed. Meta argues that by the time a chip finishes its roughly two-year design cycle, the AI workloads it was built for have often already changed shape. Instead of betting big on one design and hoping it ages well, Meta is leaning into a chiplet-based, modular approach where compute, networking, and memory components can be swapped and upgraded independently. That's how MTIA 300, originally tuned for ranking and recommendation training, could evolve into MTIA 400 for broader generative AI work, then into MTIA 450 and 500, both aimed squarely at generative AI inference.

The specs show the trajectory clearly. Going from MTIA 300 to MTIA 500, Meta says high-bandwidth memory bandwidth grows 4.5 times and compute throughput jumps 25 times, thanks to lower-precision number formats. MTIA 450 doubles memory bandwidth over MTIA 400 and boosts certain compute operations by 75%, while MTIA 500 adds another 50% bandwidth gain and up to 80% more memory capacity on top of that. MTIA 400 itself has finished lab testing and is headed toward data center deployment, with a 72-accelerator setup that Meta claims is competitive with established commercial hardware.

Meta frames its overall approach around three ideas: moving fast, by aiming to ship a new chip roughly every six months, building an inference-first design rather than repurposing chips built for massive pretraining runs, and staying deeply tied to PyTorch, the machine learning framework Meta itself created. The software stack lets the same models run on MTIA or on GPUs without rewrites, and tools like vLLM and Triton are baked in rather than bolted on. Meta describes this as reducing friction for developers, which matters more the more chip generations you're cycling through.

None of this means Meta is walking away from GPUs. The company is explicit that it still wants a mix of hardware, including outside options, alongside its homegrown chips. But the sheer number of MTIA generations moving through the pipeline in such a short window signals how central custom silicon has become to running AI at Meta's scale, and how much the company is willing to bet on iteration speed over any single perfect design.