The recent strategic partnership between AMD and Meta Platforms mirrors the earlier deal AMD struck with OpenAI, sharing key similarities: both involve 6 gigawatts of datacenter capacity across compute, storage, and networking, and each includes a warrant for 160 million AMD shares as part of the incentive package. While the idea of ten companies owning 100% of AMD within five years is clearly hyperbolic, it underscores the extraordinary scale of these commitments. The real significance lies in the fact that both OpenAI and Meta are using stock warrants—convertible into shares upon achieving milestones—to finance massive hardware purchases, effectively turning future equity into a form of payment. Unlike OpenAI, which faces scrutiny over its ability to generate hundreds of billions in cash annually, Meta has the financial muscle to back its commitments. The company is already investing heavily in AI infrastructure, and its ability to deploy such vast amounts of capital makes this deal far more credible than the OpenAI arrangement. The AMD-Meta deal is also comparable in scale to Meta’s recent agreement with Nvidia, though the specifics differ. While Nvidia’s deal includes access to millions of Blackwell and Rubin GPUs, much of it is tied to cloud capacity rather than on-premises hardware. In contrast, AMD’s deal is centered on a five-year, multigenerational partnership beginning in the second half of 2026, with the first 1 gigawatt of systems based on a custom MI450 GPU and the co-designed Helios Open Rack Wide v3 system. At 500,000 to 600,000 GPUs per gigawatt, the 6 gigawatts equate to roughly 3.3 million MI400-series equivalents. At an average GPU price of $35,000, that’s $115.5 billion in GPU costs alone—about $23.1 billion annually. When factoring in racks, networking (including AMD’s Pensando DPUs), and storage, the total hardware cost per gigawatt could reach $35 billion, with additional expenses for power and cooling. AMD CEO Lisa Su emphasized that Meta is an early adopter of the MI300X and MI350X, and that the upcoming MI450 series was designed with Meta’s needs in mind. The custom MI450X GPU, tailored for Meta’s inference workloads, does not require a new tapeout, suggesting it’s a variant of the existing MI400 architecture. It may feature adjustments in HBM memory stacking, clock speeds, or a reconfiguration of vector and tensor cores—potentially optimizing for performance per watt or raw throughput depending on Meta’s priorities. Meta is also set to adopt AMD’s upcoming Venice Zen 6 Epyc 9006 and Verrano Zen 7 Epyc 9007 CPUs, which will power both AI workloads and general applications across its platforms. The deal’s structure provides AMD with significant manufacturing certainty. With 2 gigawatts already committed by OpenAI and Meta, and the remaining 4 gigawatts to be contracted incrementally through 2030, AMD can plan production with far less risk. Assuming the warrants are exercised linearly and AMD’s stock reaches $600 by 2030, the 160 million shares could be worth around $69 billion—equivalent to roughly 2 gigawatts of hardware at current pricing. This effectively reduces the cash outlay for Meta, using stock as a form of deferred payment. If successful, AMD could capture about 40% of Meta’s AI accelerator revenue, second only to Nvidia’s 50%, with the rest split between Meta’s own MTIA chips and potential Google TPU deals. Together, these systems represent over half of Meta’s projected $600 billion in datacenter spending through the end of the decade. As Su concluded, this is a mutual bet: AMD is betting big on Meta, and Meta is betting big on AMD. The outcome could redefine the AI infrastructure landscape.