Meta's MTIA Chip Roadmap: Four Custom AI Chips in Two Years to Challenge Nvidia
Meta unveiled four MTIA chip generations — 300, 400, 450, and 500 — launching every six months through end of 2027. The custom silicon strategy, paired with a $27B infrastructure deal, signals Meta's serious commitment to the global AI race.
Photo: Meta AI Blog — Meta's next-generation MTIA chip
Four Chip Generations in Two Years
Meta announced a detailed chip roadmap for the first time with a six-month release cadence — a rare pace in the custom semiconductor industry. This strategy reflects the ambition to reduce reliance on Nvidia and AMD before the end of 2027.[1]
MTIA 300
LiveThe first generation in Meta's new roadmap, MTIA 300 is already deployed across Meta's production infrastructure, serving AI models for Facebook, Instagram, and WhatsApp.
MTIA 400
72 AcceleratorsKey feature: 72-accelerator scale-up domain, directly competitive with commercial products like Nvidia H100. Marks Meta's biggest leap forward in AI inference performance.
MTIA 450
Efficiency+An optimized iteration of the MTIA 400 with improved power efficiency and compute density. Expected to reduce inference costs by 40% compared to MTIA 300.
MTIA 500
Next-GenThe flagship generation in the 2027 roadmap, MTIA 500 is designed to support next-generation AI models at trillion-parameter scale. Aims to position Meta alongside the world's leading AI compute providers.
MTIA 300 vs. MTIA 400
The MTIA 400 with its 72-accelerator scale-up domain represents the most significant leap, bringing Meta's in-house chip to commercial-product levels.[3]
The $27 Billion AI Infrastructure Commitment
Alongside revealing the MTIA chip roadmap, Meta confirmed a $27 billion AI infrastructure deal. This massive investment covers new data center construction, large-scale MTIA chip deployment, and global expansion of AI compute capacity.[2]
Photo: CNBC — Meta's large-scale AI data center
New Data Centers
Multiple SitesBuilding specialized hyperscale data centers for AI workloads, optimized for MTIA chips rather than commercial GPUs.
Large-scale MTIA Deployment
Deployment ActiveMTIA 300 is already fully deployed across production infrastructure. MTIA 400 will replace most Nvidia GPUs in new data centers.
AI Compute Capacity
ExaFLOPS ScaleExpanding Meta's total AI compute capacity to tens of exaFLOPS, capable of serving over 3.2 billion daily active users.
Meta's largest ever AI infrastructure deal, announced March 2026
Why Custom Silicon Matters for Meta's AI
Meta's multi-billion-dollar investment in custom AI chips is no accident. It's a long-term strategy to cut costs, gain control, and better serve over 3.2 billion daily users.
Reduce Nvidia Dependence
Nvidia controls over 80% of the AI chip market. Dependence on a single vendor creates supply and cost risks. MTIA chips allow Meta to actively control its hardware roadmap.
Optimized for Meta's Models
Commercial chips must serve many different workloads. MTIA is purpose-built for Meta's own AI models like Llama, content recommendation systems, and automated content moderation.
Long-term Cost Savings
The $27B AI infrastructure investment combined with custom chips could save tens of billions in GPU costs long-term. This converts operating costs into strategic proprietary assets.
AI for Billions of Users
With over 3.2 billion daily users across Facebook, Instagram, WhatsApp, and Threads, Meta needs AI infrastructure at unprecedented scale. Custom chips are key to meeting that demand efficiently.
MTIA vs. Competing AI Silicon
The race to develop custom AI silicon is intensifying. Beyond Nvidia's dominant H100/H200, tech giants like Google, Amazon, and Microsoft are all developing proprietary silicon.
Nvidia H100
Dominant AI GPU with peak performance. MTIA 400 is designed to compete directly in the AI inference segment.
Google TPU v5
Google's custom tensor processor with advantages in large model training. Meta drew lessons from the TPU model when building the MTIA roadmap.
Amazon Trainium 2
AWS AI chip focused on LLM training in the cloud. Similar strategy to Meta but serving external customers.
Microsoft Maia 100
Microsoft's first AI chip, developed to serve Azure AI workloads. Also building independence from Nvidia like Meta.
Photo: Tom's Hardware — Close-up of the next-generation MTIA chip
▸ Meta building MTIA chips reduces NVIDIA dependence -- potentially saving billions in AI infrastructure costs annually
Related: Perplexity AI Search and Morgan Stanley AI bet.