Meta, AMD Reportedly Seal $100B Chip Deal
Meta has reportedly sealed a custom AI chip deal with AMD valued at over $100 billion. The massive commitment goes beyond buying off-the-shelf Instinct accelerators, indicating a deep co-design partnership for tailored datacenter solutions. This move solidifies a second-source strategy as a critical pillar for hyperscalers, ensuring they aren't solely dependent on Nvidia for their AI roadmaps.
The AMD-Meta deal is structured around a multi-year, multi-generation partnership to deploy up to 6 gigawatts of AMD Instinct GPUs, with the first 1 GW shipment of custom MI450-based accelerators beginning in the second half of 2026. This collaboration goes beyond just GPUs, aligning silicon roadmaps for AMD's EPYC CPUs and ROCm software, all integrated into the AMD Helios rack-scale architecture that was co-developed with Meta through the Open Compute Project. A key feature of the agreement is a performance-based warrant that could give Meta up to a 10% stake in AMD, representing 160 million shares. This warrant vests in tranches as Meta achieves specific GPU shipment milestones, with the final tranche tied to AMD's stock price reaching $600 per share, creating a tightly coupled financial incentive for both companies. This deal structure mirrors a similar agreement AMD made with OpenAI in October 2025. This partnership is a cornerstone of Meta's broader multi-vendor silicon strategy, aimed at mitigating supply chain risks and reducing over-dependence on Nvidia. Just days before the AMD announcement, Meta also signed a significant multi-year deal with Nvidia for millions of its Blackwell and Rubin GPUs. Meta's approach now strategically leverages Nvidia for frontier model training, AMD for large-scale inference, and its own in-house MTIA (Meta Training and Inference Accelerator) for specialized recommendation algorithms. Meta's custom silicon journey began with the MTIA v1, an ASIC (Application-Specific Integrated Circuit) designed specifically to accelerate its critical deep learning recommendation models. While these initial chips, fabricated on a 7nm process, were more efficient for simpler workloads, they highlighted Meta's long-term goal: gaining full-stack control from silicon to software to optimize performance and cost for their specific needs. This in-house development provides crucial negotiation leverage with external vendors. The move by hyperscalers like Meta, Google (TPU), and Amazon (Trainium/Inferentia) to design their own chips is reshaping the semiconductor landscape. By creating custom ASICs optimized for their primary workloads, these tech giants can significantly reduce total cost of ownership (TCO) and power consumption compared to general-purpose GPUs. This industry-wide pivot from relying solely on merchant silicon to developing bespoke hardware is driven by the need for control, efficiency, and a competitive edge in the AI arms race. AMD's recent wins with Meta and OpenAI signal a significant challenge to Nvidia's market dominance, which currently accounts for an estimated 80-95% of the AI chip market. While AMD's market share is still below 10%, its Instinct MI300 series, which boasts a memory advantage over Nvidia's H100, is gaining traction. This deal, focused on custom MI450 GPUs, validates AMD's rack-scale platform as a credible alternative for hyperscale AI deployments.
