Meta has installed the first of AMD's newest AI servers, called Helios, in its data centers. Each one is a refrigerator-size cabinet packed with AI chips that work together as a single machine. AMD announced on October 27 that Meta, the company behind Facebook and Instagram, is testing these cabinets ahead of a wider rollout.

Meta already runs important parts of its revenue-producing work on AMD chips. Meta uses AMD's specialized AI chips, called Instinct, to train AI models and to rank the posts and ads people see, and advertising brings in most of Meta's revenue. Meta also runs millions of AMD's everyday server chips, called EPYC, including the ones behind Muse, its AI assistant (AMD, 2026).

AMD engineers also had Meta's newest AI models running on AMD hardware, including Ryzen AI PCs, on the day Meta released them, so developers could start building without waiting for a port.

Meta Gets Chips Built Around Its Own Workloads

In February, Meta agreed to buy AMD AI chips over several years, enough to draw 6 gigawatts of electricity, about the output of six large nuclear reactors (AMD, 2026). I wrote about that deal the day it was announced. AMD scheduled the first deliveries for the second half of 2026.

Meta told AMD what it needed before AMD finished its next AI chip, and AMD is building a custom version tuned to Meta's work. The two companies also designed the Helios cabinets together through the Open Compute Project, a group where tech companies share data center designs. A company that gets chips built around its own work gets more done for each dollar it spends on hardware and electricity.

Buying from both AMD and NVIDIA also gives Meta more bargaining power on price and delivery dates.

Switching Chip Brands Pays Off After Engineers Adapt the Software

Meta published a research paper on October 15 about training its recommendation systems, the software that picks what shows up in your feed, on AMD chips instead of NVIDIA chips. Meta's engineers reshaped the software to suit AMD's chips, and training ran up to 1.5 times faster than a version they had moved over unchanged (Meta, 2026).

Moving AI software to a new chip brand resembles moving a favorite recipe from a gas stove to an induction cooktop. The dish cooks on both, and you get the best results after you adjust the timing and the pans. Meta has a large engineering staff to make those adjustments, a resource most companies lack. Meta and AMD publish many of their fixes in free, open software tools that other companies use, which lowers the cost of switching for everyone else.

AI Assistants Need Ordinary Computer Chips Too

Meta runs Muse on AMD's everyday server chips, the same kind companies use for websites and databases. An AI assistant that searches and writes on behalf of millions of people does a lot of ordinary computing between answers, and Meta handles that work on regular servers.

If your company starts using AI assistants at scale, plan for your regular server costs to rise along with your AI chip costs.

Stock Ties Give Buyers a Reason to Lock In Delivery Dates

AMD also said it will invest $1 billion in Meta's stock (AMD, 2026). On October 26, Meta said it plans to sell $95 billion of new stock to pay for AI data centers, and NVIDIA, AMD and Broadcom said they were interested in buying up to $5 billion of it between them (CNBC, 2026).

Meta already had a path to owning AMD stock. In the February chip deal, AMD gave Meta the right to collect up to 160 million AMD shares (AMD, 2026). Meta earns those shares in stages as AMD delivers chips and as AMD's stock price climbs, according to Latham & Watkins, the law firm that advised AMD.

A restaurant chain and its main food supplier with the same setup would look like this: every big delivery earns the restaurant a few shares of the supplier, and the supplier buys shares in the restaurant, which helps the restaurant open more locations that will order more ingredients.

AMD and Meta now each profit from the other's growth.

AMD signed a deal of the same size with OpenAI, the maker of ChatGPT, in October 2025, which also gave OpenAI a warrant for up to 160 million AMD shares (AMD, 2025). OpenAI plans to start using Helios in the second half of 2026. AI chips are in short supply, and AMD keeps the split of its supply among customers private, as chipmakers do, so pin down delivery dates in your contract. Earlier this month I wrote about Vultr ordering $1.2 billion of AMD servers, a sign that buyers are reserving AI capacity years before they need it.

AMD keeps its supply split private, so pin down delivery dates in your contract.