Meta forced AMD to halve its top AI chip, analysts say

Meta Platforms ↗ 0.00% is deploying a custom version of AMD's Advanced Micro Devices, Inc. ↗ 1.30% Instinct MI450 accelerator that halves the compute silicon and slashes memory capacity by roughly two-thirds compared to the standard MI455X, according to semiconductor research firm SemiAnalysis. The decision, optimized for Meta's recommendation systems rather than generative AI, has prompted an unusual public rebuke from the analyst firm, which warned the configuration will leave Meta's AI research teams at a disadvantage against rivals.

A Stripped-Down Design

SemiAnalysis disclosed on July 21 that Meta's custom MI400-series chip uses four compute dies instead of the standard eight and replaces the MI455X's twelve 12-Hi HBM4 stacks — totaling 432GB — with six 8-Hi stacks providing roughly 144GB of memory. Meta's infrastructure team justified the configuration as tuned for the CPU-to-GPU ratio its recommendation-system workloads demand, optimizing cost per unit of memory bandwidth.FuTu News+1

The standard MI455X, built on TSMC's 2nm process with hybrid bonding packaging, represents the current ceiling in GPU packaging and memory density. The custom variant sacrifices that headroom to serve workloads that Meta runs at enormous scale but that differ sharply from the large language model training and inference tasks now central to the AI industry.FuTu News

SemiAnalysis Sounds the Alarm

SemiAnalysis labeled the decision "catastrophic" and took the rare step of publicly urging AMD to work directly with TBD Lab — Meta's superintelligence research division, led by Alexandr Wang — rather than Meta's infrastructure organization. The firm argued the cut-down chip holds little appeal for TBD Lab engineers, who would "strongly favor" Nvidia's ↘ 5.47% Vera Rubin architecture instead.LinkedIn+1

The criticism fits a broader pattern SemiAnalysis identified in Meta's hardware choices. The firm pointed to the company's Ariel server — a custom GB200 configuration with a one-to-one CPU-to-GPU ratio that carried a 14 percent higher total cost of ownership than the standard NVL72 design — as another instance of recommendation-system priorities overriding the needs of generative AI teams. Meta has since reverted to a standard configuration for its GB300 servers, an implicit acknowledgment the custom approach failed.FuTu News

Stakes for the AMD-Meta Partnership

The custom chip sits within a broader five-year agreement, announced in February, under which AMD will supply up to six gigawatts of Instinct GPUs to Meta in a deal valued at up to $60 billion. AMD issued Meta a performance-based warrant for 160 million shares — potentially 10 percent of the chipmaker's equity — with vesting tied to shipment milestones. The first gigawatt of MI450-based hardware is scheduled to begin shipping in the second half of 2026.Jon Peddie Research+1

SemiAnalysis warned that if the cut-down design remains Meta's primary AMD deployment, it could undermine AMD's volume at the company and limit Mark Zuckerberg's ability to rent spare capacity to outside customers.LinkedIn