Meta Teams Up with AMD to Power Next‑Gen AI Infrastructure, AI Magazine Reports

Meta’s shift toward AMD’s GPU portfolio marks a strategic pivot from its in‑house “Mosaic” ASICs to a more heterogeneous compute stack, according to AI Magazine. The partnership will initially deploy a 6‑gigawatt (GW) array of Radeon Instinct accelerators across Meta’s data centers, a scale that rivals the company’s previous AI‑specific hardware rollouts. AMD will supply the GPUs, while Meta will integrate them with its existing software stack—PyTorch‑based training pipelines, the Open Compute Project (OCP) server designs, and the proprietary “Mosaic” inference layer that has powered LLaMA‑2 and other large language models (LLMs). The move is intended to accelerate the training of next‑generation foundation models that Meta plans to release later this year, as the company seeks to close the performance gap with rivals that have already embraced third‑party silicon.

Forbes reports that the deal includes a “massive 6 GW GPU” commitment, a figure that translates to roughly 60,000 high‑end Radeon Instinct cards operating at peak power. The agreement also grants AMD a 10 percent equity stake in Meta’s AI joint venture, mirroring the equity component AMD secured in its OpenAI partnership earlier this year. The equity clause, noted by The Decoder, is intended to align the two firms’ long‑term incentives and give AMD a foothold in the rapidly expanding AI infrastructure market. The financial terms of the arrangement have not been disclosed, but The Decoder characterises the structure as “basically copy‑pasted” from the OpenAI deal, suggesting a similar valuation framework and revenue‑sharing model.

Tom’s Hardware adds that the partnership is being billed as a “$100 billion AI deal,” a headline figure that reflects the projected cumulative spend on GPU hardware, software integration, and joint research over the next decade. While the headline number is not broken down in the source material, the article implies that Meta will amortise the capital outlay across its suite of AI products, from the internal LLMs that power Facebook and Instagram content moderation to the upcoming generative AI tools slated for the metaverse platform. The hardware infusion is expected to boost Meta’s training throughput by an estimated 30 percent, according to internal benchmarks cited by AI Magazine, though the exact performance uplift remains confidential.

Technical analysts note that AMD’s Radeon Instinct GPUs bring a different architecture than Meta’s custom ASICs, emphasizing higher memory bandwidth and broader support for mixed‑precision workloads. This aligns with Meta’s reported focus on scaling model size beyond 100 billion parameters, where memory‑bound operations become a bottleneck. AMD’s RDNA 3‑based compute units, combined with the company’s Infinity Fabric interconnect, enable tighter scaling across multiple nodes, a capability that Meta’s internal teams plan to exploit through a new “distributed training fabric” built on top of the Open Compute Server (OCS) specifications. The partnership also opens the door for AMD to contribute its ROCm open‑source software stack, which could streamline driver compatibility and reduce the engineering overhead of integrating third‑party GPUs into Meta’s proprietary pipelines.

The collaboration arrives at a moment when the AI hardware market is consolidating around a few dominant players. By securing a multi‑gigawatt supply of AMD GPUs and tying the deal to equity participation, Meta is hedging against supply‑chain volatility that has plagued custom‑silicon programs in recent years. As AI Magazine concludes, the move “signals Meta’s willingness to blend its own silicon expertise with best‑in‑class off‑the‑shelf accelerators to sustain its AI ambitions.” Whether the mixed‑hardware approach will translate into a measurable competitive edge remains to be seen, but the scale of the AMD commitment suggests Meta is preparing for a sustained, high‑intensity AI research agenda.