Meta New AI Chips Set for 2027 Deployment as Inference Takes Center Stage

Mirror Review
Date: September 16, 2026

Meta is preparing to deploy its next generation of custom AI chips in data centers during 2027. The MTIA 450, internally known as Arke, is expected to enter data centers in the first half of 2027, while the MTIA 500, known as Astrid, is expected to enter data centers toward the end of 2027. Meta is focusing the two chips primarily on AI inference as the company works to improve the efficiency and cost of running AI workloads at scale.

Why Meta New AI Chips Are Designed for AI Inference

Meta new AI chips are being developed primarily to handle high-volume AI inference efficiently. Meta says the MTIA 300, MTIA 400, MTIA 450, and MTIA 500 support multiple AI workloads, while Meta is optimizing the MTIA 450 and MTIA 500 primarily for generative AI inference.

AI inference is the stage in which a trained AI model processes an input and produces a prediction, recommendation, or response. Meta can design inference-focused hardware around the computing and memory requirements involved in serving AI models at scale.

Meta says the MTIA roadmap is designed to improve efficiency while allowing the chips to fit into Meta’s existing rack infrastructure. Meta has already deployed hundreds of thousands of MTIA chips for inference across workloads such as recommendations and advertising.

When Will Meta MTIA 450 Arke and MTIA 500 Astrid Launch?

The MTIA 450, also called the Arke AI chip, is currently being tested and is expected to enter Meta’s data centers in the first half of 2027. The MTIA 500, also called the Astrid AI chip, is expected to complete its current design work in about a month and begin entering AI data centers toward the end of 2027, according to September 15 reporting.

Meta’s March 2026 MTIA roadmap had already identified MTIA 450 for mass deployment in early 2027 and MTIA 500 for mass deployment during 2027. Meta is developing four new MTIA generations within two years, covering MTIA 300, MTIA 400, MTIA 450, and MTIA 500.

Meta designed MTIA 450 with increased high-bandwidth memory capacity and bandwidth for generative AI inference. MTIA 500 extends that approach with further increases in memory and compute capability.

How Much Faster Is MTIA 500 Than MTIA 450?

The MTIA 500 provides higher memory bandwidth, greater HBM capacity, and higher low-precision compute capability than the MTIA 450. Meta reports 50% higher HBM bandwidth, up to 80% higher HBM capacity, and 43% higher MX4 FLOPS for MTIA 500 compared with MTIA 450.

The MTIA 500 increases HBM bandwidth from 18.4 TB/s to 27.6 TB/s and raises HBM capacity from 288 GB to as much as 512 GB. The MTIA 500 also increases MX4 compute from 21 PFLOPS to 30 PFLOPS. These figures describe hardware specifications rather than independent benchmark results.

Across the MTIA 300-to-500 roadmap, Meta reports a 4.5x increase in HBM bandwidth and a 25x increase in compute FLOPS. Meta presents these figures as comparisons across its chip generations, not as third-party benchmark results.

Why Did Meta Cancel the Olympus AI Chip?

Meta canceled Olympus, a planned AI accelerator designed to support both AI training and AI inference. September 2026 reporting says Meta canceled the dual-purpose accelerator after determining that a chip designed for both workloads could cost about 30% more than an inference-focused design.

Meta engineering executive Yee Jiun Song said the cost difference becomes significant when Meta deploys computing capacity at gigawatt scale. Meta is therefore emphasizing workload-specific chip designs for large-scale inference rather than using one accelerator for both training and inference.

Meta’s custom-silicon strategy supports the workload-specific approach. Meta says different AI workloads have different requirements and that no single chip can efficiently address every workload, which is why Meta is developing a portfolio of MTIA accelerators.

How Meta Uses Custom AI Chips Alongside NVIDIA and AMD

Meta custom silicon is not intended to replace every external accelerator across Meta AI infrastructure. Meta says the company uses a diversified silicon strategy and matches different processors to different AI workloads.

Meta also has a long-term agreement with AMD for up to 6 gigawatts of AMD Instinct GPUs. AMD says the first gigawatt of the agreement is expected to begin shipping in the second half of 2026, with additional deployments planned across subsequent generations.

Meta can therefore combine internally developed MTIA accelerators with NVIDIA, AMD, and other external computing hardware. Meta custom silicon gives the company additional control over hardware optimized for workloads such as recommendations and generative AI inference.

How Broadcom and TSMC Are Involved in Meta’s 2027 AI Chips

Meta is working with Broadcom to co-develop multiple generations of next-generation MTIA chips. Meta expanded its Broadcom partnership in April 2026 as part of the company’s longer-term custom-silicon strategy. Meta said the partnership covers chip design, advanced packaging, and networking technologies.

TSMC is manufacturing the latest MTIA chips, according to September 2026 reporting. Twelve Arke test chips reportedly arrived at Meta from TSMC on September 1, with early performance measurements within approximately 2% to 3% of pre-production simulations. Meta also used the test chips to run Meta AI models as well as models from DeepSeek and Alibaba.

Meta’s Arke testing provides an early validation stage before broader 2027 deployment. Meta still needs additional testing, tuning, and production work before the chips reach large-scale deployment.

What Meta New AI Chips Mean for Meta AI Infrastructure

Meta says each MTIA generation is designed to improve AI workload efficiency. Meta engineering executive Yee Jiun Song described the target in terms of better performance per watt and performance per dollar, which are important measures for operating AI computing infrastructure at large scale.

Meta has committed to deploying more than one gigawatt of custom chips over 12 months, according to September 2026 reporting.

For Meta AI infrastructure, the 2027 deployment of Arke and Astrid represents an expansion of workload-specific computing. Meta will use the MTIA 450 and MTIA 500 primarily for generative AI inference while continuing to use external accelerators for other computing requirements.

Meta new AI chips are therefore part of a broader portfolio strategy rather than a plan to rely on one processor design. Arke and Astrid extend Meta’s custom-silicon program toward higher-volume AI inference, while Meta continues to develop and deploy different processors for the varying demands of its AI systems.

Gurushanth S Jatti