Meta announced on March 11, 2026 that it has developed and deployed four successive generations of its in-house Meta Training and Inference Accelerator (MTIA) chips within roughly two years, expanding the chip family’s workload coverage from ranking and recommendation inference to large language model training and GenAI inference. The company has published supporting research at ISCA’23 and ISCA’25 detailing the first two generations. Author details were not available at time of publication for the March 2026 blog post.
- Meta has deployed hundreds of thousands of MTIA chips in production across its global infrastructure, developed in partnership with Broadcom.
- The MTIA family now spans four active generations — 300, 400, 450, and 500 — each targeting distinct workload types from recommendation ranking to GenAI inference.
- MTIA 450 doubles the high-bandwidth memory (HBM) bandwidth of MTIA 400, exceeding the bandwidth of leading commercial products, with mass deployment scheduled for early 2027.
- Meta’s stated strategy is deliberate short-cadence iteration using modular chiplets, designed to keep hardware aligned with AI model evolution rather than committing to multi-year single-generation bets.
What Happened
Meta published a detailed account of its MTIA chip program on March 11, 2026, disclosing that four chip generations — MTIA 300, 400, 450, and 500 — have either already been deployed or are scheduled for deployment in 2026 or 2027. The chips are designed in close partnership with Broadcom and are a central component of Meta’s AI infrastructure alongside third-party silicon. The announcement follows earlier research disclosures at ISCA’23 and ISCA’25, which detailed the first two generations, formerly called MTIA 1 and MTIA 2i and now renamed MTIA 100 and MTIA 200.
Why It Matters
Meta’s platforms serve billions of users daily, and the company has described serving AI models at global scale as “one of the most demanding infrastructure challenges in the industry.” The MTIA program reflects a broader industry shift toward custom silicon: Google has its TPU line, Amazon has Trainium and Inferentia, and Microsoft has the Maia chip. Meta’s approach distinguishes itself through an explicit short-cadence design philosophy intended to stay ahead of rapidly shifting AI model requirements, rather than committing to long design-to-production cycles. The company also noted it tested MTIA with Llama large language models as part of validating the hardware for GenAI workloads.
Technical Details
Each chip generation targets a specific workload layer. MTIA 300 was initially optimized for ranking and recommendation (R&R) models and is currently in production for R&R training. MTIA 400 evolved from the MTIA 300 foundation to support GenAI models while retaining R&R capabilities; it features a 72-accelerator scale-up domain and has completed lab testing ahead of data center deployment. Meta described MTIA 400’s performance as “competitive with leading commercial products.”
MTIA 450 is specifically optimized for GenAI inference. Meta doubled the HBM bandwidth relative to MTIA 400, citing memory bandwidth as the dominant factor in GenAI inference performance. The resulting bandwidth figure exceeds that of existing leading commercial products, according to Meta. MTIA 450 also introduced low-precision data types co-designed for inference workloads. Mass deployment is scheduled for early 2027.
The MTIA chips use a modular chiplet architecture, which Meta says allows each generation to incorporate the latest AI workload insights and hardware technologies without a full redesign from scratch. Meta stated: “Rather than placing a bet and waiting for a long period of time, we deliberately take an iterative approach: Each MTIA generation builds on the last, using modular chiplets, incorporating the latest AI workload insights and hardware technologies, and deploying on a shorter cadence.”
Who’s Affected
The primary beneficiaries of MTIA capacity are internal Meta teams operating AI workloads across Facebook, Instagram, WhatsApp, and Meta AI products. The company’s billions of end users are served by these systems indirectly — through personalized content recommendations and AI assistant responses — without direct access to or awareness of the underlying hardware. External developers building on Meta’s Llama models are not direct users of MTIA infrastructure, but the chip program supports Meta’s ability to continue training and serving those models at scale.
What’s Next
MTIA 400 is on the path to data center deployment as of the March 2026 announcement, with MTIA 450 targeting mass deployment in early 2027. Details on MTIA 500 were disclosed in the blog post but were not fully captured in the available source excerpt; Meta confirmed it is part of the roadmap. The company stated it remains committed to a “diverse silicon portfolio” that includes both internal MTIA chips and third-party solutions, meaning MTIA is not intended to fully replace commercial GPU or accelerator procurement. No specific performance benchmarks or pricing disclosures were made public in the announcement.
