AMD Meta GPU deal: up to 6 gigawatts and a share warrant
The AMD Meta GPU deal, announced on February 24, 2026, commits Meta to deploying up to 6 gigawatts of AMD Instinct GPUs over several hardware generations, starting with custom MI450 chips in the second half of 2026. In return, AMD issued Meta a warrant for up to 160 million AMD shares that vests as shipments arrive. For companies buying AI compute, it shows how the largest buyers now spread their GPU spending across suppliers and win hardware built for their own workloads.
Why Meta wanted a second GPU supplier
Meta needs far more AI compute than one chip vendor can comfortably supply, and it wants options. Mark Zuckerberg said the AMD agreement is about deploying efficient inference compute and diversifying where that compute comes from. The week before, Meta had announced a multi-year agreement with Nvidia to deploy its CPUs alongside "millions" of Blackwell and Rubin GPUs, according to DCD.
"This is an important step for Meta as we diversify our compute." — Mark Zuckerberg, founder and CEO of Meta
Meta is not a new AMD customer. It already runs AMD EPYC CPUs across its global infrastructure, in the millions, and AMD said on its conference call that those chips power many of Meta's core services.
What Meta is buying from AMD
Meta has agreed to deploy up to 6 gigawatts of AMD Instinct GPUs over several hardware generations. The first deployment uses custom Instinct GPUs based on AMD's MI450 architecture, paired with sixth-generation EPYC CPUs codenamed "Venice" and built on the AMD Helios rack-scale architecture. AMD says Helios was developed with Meta through the Open Compute Project. When AMD first showed the rack in June 2025, it combined Venice CPUs, MI455X GPUs and Pensando Vulcano NICs and delivered 2.9 exaflops of FP4 performance per rack, DCD reports.
The systems run ROCm, AMD's open-source GPU software. Meta will also be a lead customer for the Venice CPUs and for "Verano," the next EPYC generation, which is targeted for 2027.
How the 160 million share warrant works
AMD pays part of the price for this business in its own stock. Under an 8-K filed with the SEC, AMD issued Meta a performance-based warrant on February 23, 2026, to buy up to 160 million AMD shares at $0.01 each. Tom's Hardware puts that at roughly 10% of AMD's stock.
The shares vest in tranches. The first vests when the initial 1 gigawatt equivalent of Instinct GPU products ships, and full vesting requires Meta to buy 6 gigawatt equivalent. Each tranche also depends on AMD's share price clearing thresholds that rise to $600 for the final tranche, and Meta must meet certain technical and commercial conditions before it can exercise. The warrant can be exercised until February 23, 2031. Alongside it, Meta made a binding commitment to buy the first gigawatt under an amended Master Purchase Agreement dated May 23, 2023.
What the AMD Meta GPU deal is worth, and what is still a forecast
No hardware has shipped yet, so there is no measured result. AMD says shipments for the first gigawatt begin in the second half of 2026, and that each gigawatt will bring it significant double-digit billions of dollars in revenue. The Wall Street Journal, cited by Tom's Hardware, valued the deal at more than $100 billion. Those are projections, not results.
The structure copies AMD's October 2025 agreement with OpenAI, which also covers 6 gigawatts and 160 million shares. Neither company has said where its first gigawatt will be deployed.
What the deal means for companies buying AI compute
For any business that depends on AI capacity, the deal shows three things. First, the largest buyers no longer bet on one GPU vendor; Meta now has long-term agreements with both Nvidia and AMD. Second, a big enough buyer can get hardware shaped around its own workloads: the MI450 parts for Meta are custom, not off the shelf, and the two companies say they will align roadmaps across silicon, systems and software. Third, suppliers will share upside to lock in demand, here through equity that only pays out if the volume and the share price both arrive.
For smaller companies, the practical point is software. ServeTheHome argues that with Meta and OpenAI both deploying at this scale, the question of whether ROCm is ready for production is fading. That widens the choice of hardware for anyone running inference.
Questions and answers
How many gigawatts of AMD GPUs will Meta deploy?
Up to 6 gigawatts of AMD Instinct GPUs across multiple generations. Meta has made a binding commitment to the first gigawatt, with shipments scheduled to start in the second half of 2026.
Why does Meta get AMD shares in the deal?
AMD issued Meta a warrant for up to 160 million shares at $0.01 each. The shares vest only as GPU shipments reach set milestones and AMD's stock clears price thresholds up to $600, which ties Meta's reward to buying the full volume.
Which chips are in the first deployment?
Custom Instinct GPUs based on the MI450 architecture, sixth-generation EPYC "Venice" CPUs and the Helios rack-scale architecture, running ROCm software.
MOZGII view
How we would build it
A 6 gigawatt GPU order is a power and operations project before it is a chip project. We would start with the plan for where each gigawatt lands: the site, the grid connection, the cooling and the date power is actually available. Hardware that arrives before the building is ready earns nothing.
The second need is a software layer that does not care whose GPU is underneath. A company running both Nvidia and AMD systems would want one scheduler, one set of model-serving tools and one test suite that every model passes on both. We would keep a fixed benchmark of real inference jobs and run it on each new rack before it takes traffic. That is where a two-vendor fleet usually breaks: a model that runs well on one stack slows down or gives slightly different answers on the other, and nobody notices until costs rise.
The one metric we would watch is cost per million tokens served, measured the same way on every rack. It folds in power, utilization and software efficiency, and it shows whether a second supplier is saving money or only adding work.
Next, we would build placement rules that send each inference job to the cheapest rack that meets its latency target, so the mix of vendors becomes a lever on cost rather than a burden.
Sources
Where the facts in this story come from.
- 01AMD and Meta Announce a Massive 6GW Deal - ServeTheHomeservethehome.com
- 02AMD and Meta strike $100 billion AI deal that includes 10% stock deal — 6 gigawatt agreement includes up to 160 million AMD shares | Tom's HardwareTom's Hardwaretomshardware.com
- 03Meta and AMD sign multi-year agreement for 6GW GPU deployment - DCDdatacenterdynamics.com
- 04amd-20260223sec.gov