Meta and AMD have signed a multi-year agreement to deploy up to 6 gigawatts of AI-focused compute for the group’s data centers, powered by Instinct GPUs and AI-optimized CPUs. This capacity will be rolled out progressively over several years, starting in the second half of 2026, making it one of the largest AI infrastructure projects in the world.
The initial deployments will rely on the Helios rack architecture, co-developed by Meta and AMD within the Open Compute Project, designed to optimize density, energy efficiency and software integration.
AMD will supply next-generation GPUs (MI450) and 6th-generation EPYC CPUs, sized for large-scale inference workloads, complementing Meta’s existing training capabilities.
Objective: Diversify and Optimize the AI Infrastructure
This ramp-up of AMD’s involvement comes as Meta remains one of Nvidia’s biggest customers, with a fleet equivalent to hundreds of thousands of H100 GPUs and future Blackwell GPUs for foundation-model training. By massively introducing AMD Instinct GPUs into its data centers, Meta aims to diversify suppliers, reduce exposure to the CUDA ecosystem alone and optimize the total cost of ownership of its AI clusters.
Instinct GPUs are already used by Meta for certain workloads, and AMD highlights gains in performance per watt and total cost of ownership for broadly deployed models, whether for recommendations, vision, or LLMs.
The agreement with Meta also includes customized GPUs for the group’s workloads, a feature presented as a differentiator versus Nvidia, which does not emphasize that degree of customization in its major contracts.
A Foundation for the “Personal Superintelligence”
Meta says it intends to build an infrastructure capable of supporting a “personal superintelligence” accessible at scale, which implies massive capabilities for training and inference of multimodal models for billions of users.
The targeted 6 GW of AI power is meant to feed future generations of recommendation algorithms, conversational assistants, creative tools and content moderation across Facebook, Instagram, WhatsApp and mixed-reality services.
To meet this ambition, Meta plans to build or convert 30 data centers, including 26 located in the United States, designed natively for GPU-heavy AI workloads and energy optimization. Standardization around common architectures like Helios and the Open Compute Project should facilitate rapid scaling, easier maintenance and the co-design of new generations of racks with AMD.
A Warrants Mechanism Aligning Meta and AMD
Beyond the hardware, the agreement includes a performance warrants mechanism giving Meta the right to acquire up to 160 million AMD shares, roughly 10% of the company’s equity, subject to conditions. The warrant tranches unlock as Meta receives Instinct deliveries (starting at 1 GW, then up to 6 GW), as certain price hurdles are crossed by AMD, and as technical and commercial milestones are achieved by both partners.
This structure creates strong alignment: the better AMD’s infrastructure performs for Meta, the more Meta benefits from becoming a meaningful AMD shareholder, and the more AMD benefits from having a large, recurring customer. In a context where AMD has already put in place a similar arrangement with OpenAI for its future MI450 GPUs, this equity-based model strengthens AMD’s standing as a preferred alternative to Nvidia in the AI cloud.
A Message to the AI Semiconductor Market
For AMD, this multi‑billion‑dollar, multi-year contract marks a major milestone in its push for market share against Nvidia in the high-end AI GPU segment. The announcement has also driven AMD’s stock higher in pre-market trading, while Meta’s reaction has been more measured, as investors weigh the scale of AI capex against long-term value creation.
