The next generation of Meta’s AI infrastructure will be partly AMD-powered. 6 Gigawatts of AMD Instinct GPUs will be rolled out, with MI450 variants being the first chips to be delivered in the second half of 2026.
For AMD, the deal is an important one, especially since the hyperscale AI expansion consists of only a few huge companies. Meta is one of them and is now opting for a more diverse approach than before. Previously, Nvidia, for many the standard choice for AI chips, was the sole supplier. The collaboration also goes beyond hardware alone. Roadmaps for “silicon, systems, and software” are being aligned, according to both parties.
The first Gigawatt implementation runs on a custom AMD Instinct GPU based on the MI450 architecture. This is combined with 6th generation AMD EPYC CPUs, codenamed “Venice,” and runs on ROCm software. The whole thing is built on the AMD Helios rack-scale architecture, a development that AMD and Meta realized through the Open Compute Project. Exactly how the MI450 differs from a normal variant remains unclear.
Strategic diversification for Meta
Meta founder Mark Zuckerberg emphasizes the importance of the deal. He talks about AMD’s strong inference capabilities. Training therefore seems to remain an Nvidia party. Since building AI models is more intensive, Nvidia’s performance advantage and software offering still seem to lead the way over AMD’s.
The collaboration has another striking aspect. Meta is entitled to up to 160 million AMD shares, depending on the number of GPUs purchased by the former company. The first tranche will be made available after the initial Gigawatt of deliveries. It is yet another interesting financial tool in an AI ecosystem that is full of unique monetary constructs. The criticism is that the deals, particularly those involving Nvidia, are circular in nature, with GPU suppliers contributing financially to AI players who then use that money to buy GPUs.
Financial structure with shares
Additional conditions link the vesting to AMD’s share price and to technical and commercial milestones that Meta must achieve. AMD’s CFO Jean Hu expects the deal to deliver “substantial multi-year revenue growth” and increase earnings per share.
Meta and AMD are also expanding their EPYC collaboration, the data center CPU that is gaining ground against Intel Xeon. Meta has already deployed millions of EPYC CPUs and significant numbers of Instinct MI300 and MI350 series GPUs worldwide. Meta will be the lead customer for the 6th generation EPYC CPUs (“Venice” and “Verano”). These processors will receive workload-specific optimizations for better performance per dollar per watt.
Nvidia remains the number one competitor
The announcement puts AMD in a stronger position in the AI race, even though Nvidia remains dominant. Since both parties have their chips manufactured by TSMC, the Taiwanese chip manufacturer’s capacity is the deciding factor. As long as supply is limited, margins will remain high.
Read also: Meta CTO: Superintelligence Labs delivers first high-performance models