Summary
Meta has unveiled a four-generation roadmap for its in-house AI chip series "MTIA." The four chips — MTIA 300 (in mass production), 400 "Iris" (in lab testing), 450 "Arke," and 500 "Astrid" — are set to be released approximately every six months. Built on a RISC-V architecture in collaboration with Broadcom and TSMC, the chips are designed specifically for generative AI inference, with the goal of reducing reliance on Nvidia GPUs. However, Nvidia Blackwell will continue to be used for training workloads, making this a strategy of "diversification" rather than "replacement."
Text
On March 19, 2026, Meta announced a comprehensive roadmap for its in-house AI inference chip, the "MTIA" (Meta Training and Inference Accelerator), outlining four generations — MTIA 300, 400 "Iris," 450 "Arke," and 500 "Astrid" — to be released sequentially on approximately six-month cycles. The assignment of codenames to each generation suggests that Meta views this chip line not as a temporary experiment, but as a long-term business foundation.
MTIA 300 — First Generation Enters Mass Production
The first model, the MTIA 300, has already entered mass production and is operating in Meta's internal data centers. It handles generative AI inference workloads for flagship services including Facebook, Instagram, and WhatsApp, with Meta claiming it outperforms Nvidia GPUs in performance per watt. For Meta, which serves AI services to billions of users, optimizing inference costs is the single most critical issue directly tied to infrastructure spending.
MTIA 400 "Iris" — The First In-House Chip to Compete with Commercial Products
Currently in lab testing, the MTIA 400 "Iris" represents an important milestone for Meta. The company positions it as "the first chip capable of competing with market-leading commercial products in raw performance." In other words, whereas the MTIA 300's strength was cost efficiency through optimization for internal workloads, Iris aims to match Nvidia GPUs in absolute performance as well.
MTIA 450 "Arke" — Doubling Bandwidth
The MTIA 450 "Arke" doubles HBM (High Bandwidth Memory) bandwidth compared to the MTIA 400. Meta describes this bandwidth as "much higher than existing leading commercial products," with the goal of fundamentally eliminating the memory bandwidth bottleneck that arises during large language model inference. Generative AI inference requires rapidly loading model weight data from memory, and expanding bandwidth leads directly to faster inference speeds.
MTIA 500 "Astrid" — The Culmination of the Roadmap
Details on the MTIA 500 "Astrid," the final generation in the roadmap, remain limited, but it is said to further evolve the high-bandwidth architecture established with Arke. If Meta maintains its six-month release cycle, Astrid would arrive sometime in late 2027 to early 2028.
Technical Foundation — RISC-V, Broadcom, and TSMC
On the technical side, the MTIA series adopts the RISC-V architecture. RISC-V is an open-source instruction set architecture (ISA) that requires no licensing fees and offers high customizability, enabling proprietary designs free from the constraints of Arm-based chips or Nvidia's CUDA ecosystem. Chip manufacturing is handled by TSMC, and a deep collaborative relationship with Broadcom has been established in the design process. The aggressive six-month new-generation release cycle is faster than Nvidia's annual release cycle.
Strategic Implications of an Inference-Focused Approach — "Diversification," Not "Replacement"
Crucially, MTIA focuses on inference rather than training. Nvidia's Blackwell-generation GPUs are still used for AI model training, and Meta will continue purchasing GPUs from both Nvidia and AMD. MTIA represents a strategy of "diversification" rather than "replacing Nvidia," seeking simultaneously to reduce supplier dependency risk and optimize costs by migrating a portion of inference workloads to in-house chips.
This "diversification" approach is not unique to Meta. Google covers both training and inference with its TPU (Tensor Processing Unit), Amazon has developed its own chips with Trainium for training and Inferentia for inference, and Microsoft is developing the Maia chip. The trend of hyperscalers focusing on developing their own AI inference chips is clear, and moving away from singular dependence on Nvidia GPUs has become the direction of the industry as a whole.
Impact on the Industry
Meta's MTIA roadmap has the potential to bring structural changes to the AI semiconductor market currently dominated by Nvidia. Following Google (TPU), Amazon (Trainium/Inferentia), and Microsoft (Maia), Meta's full-scale proprietary chip strategy is expected to reduce the proportion of Nvidia GPU procurement at hyperscalers over the medium to long term.
The adoption of RISC-V is also a significant turning point for the semiconductor industry as a whole. With RISC-V being adopted in the high-volume category of AI inference chips, the architecture's ecosystem could grow at an accelerating pace.
However, risks remain. Developing proprietary chips requires massive investment, and as AI model architectures evolve rapidly, the risk of hardware obsolescence is ever-present. Whether Meta can actually sustain its ambitious six-month cycle roadmap will be the key factor determining the success or failure of this strategy.
Reference: Meta AI Semiconductor Team announcement, RISC-V International trends, AI inference market analysis report, Google TPU / Amazon Trainium comparative data