AMD and Meta Deepen Chip Deal: Helios Racks and MI450 GPUs

AMD and Meta have expanded their chip partnership: AMD confirmed on October 27, 2026 that its Helios rack-scale AI systems are now installed and being tested inside Meta's data centers, that AMD EPYC processors already run the compute behind Meta's Muse AI agent, and that Meta will deploy a custom AMD Instinct GPU built from the MI450 architecture. AMD is also investing $1 billion in Meta's equity offering as part of the deal. The update, posted to AMD's investor-relations newsroom, builds on the two companies' February 2026 agreement to deploy up to 6 gigawatts of AMD Instinct GPUs, and it is the clearest signal yet that AMD's silicon — not just Nvidia's — is now running in production inside one of the world's largest AI infrastructure buildouts.

What AMD and Meta actually announced

The news comes from a press release AMD filed to its investor-relations site, titled "AMD and Meta Advance Strategic Partnership to Power Meta's Superintelligence Infrastructure." It isn't a brand-new deal so much as a progress report and expansion of the companies' existing relationship: it confirms that hardware promised earlier in 2026 is now physically running at Meta, names the next generation of chips the two companies are co-designing, and adds a financial component — AMD putting $1 billion into Meta's own stock.

AMD frames the collaboration as spanning "CPUs, GPUs, networking, systems and software," which matches the five product lines named in the release: EPYC server processors, Instinct GPUs, Pensando networking silicon, Helios rack-scale systems, and the ROCm software stack that ties them together. That is a broader, more vertically integrated pitch than a single chip sale — AMD is positioning itself as a full-stack infrastructure supplier for Meta's AI buildout, not just a GPU vendor.

Helios rack-scale systems are now running at Meta

The headline hardware claim is that AMD Helios — the company's rack-scale AI system, which bundles Instinct GPUs, EPYC CPUs and Pensando networking into a single liquid-cooled rack built to Meta's Open Rack Wide mechanical standard — is installed and actively being tested inside Meta's infrastructure. AMD says broader deployment is expected "in the near future," though it did not give a rack count, a site location, or a firm date for that wider rollout in this release.

Helios itself was co-developed with Meta through the Open Compute Project, the open-hardware consortium Meta helped found, and was first shown publicly at the OCP Global Summit. Having Meta run early production units is a meaningful proof point for AMD: Helios is the system AMD is counting on to compete with Nvidia's rack-scale platforms for the largest AI training and inference clusters, and an early, named hyperscale customer validates that the rack-level engineering — power, cooling, networking — works at scale, not just on a slide.

EPYC CPUs power Meta's Muse agent, with "Verano" next

On the CPU side, AMD says Meta has deployed millions of EPYC processors across multiple generations and is a lead customer for 6th Gen EPYC, the generation AMD has previously said carries the codename "Venice" (see AMD's EPYC server processor lineup). The release adds that EPYC capacity underpins the computing environment behind Muse, Meta's AI agent, and that AMD and Meta are now collaborating on a future EPYC generation code-named "Verano" — the first time that codename has surfaced publicly in an AMD release.

Server CPUs rarely make headlines the way GPUs do, but they still do the unglamorous work of orchestrating storage, networking and data pipelines around every GPU cluster, and "lead customer" status on 6th Gen EPYC means Meta is shaping how that chip performs in its own racks before it ships broadly. For background on where AMD's CPU roadmap is headed next, see our coverage of AMD's Zen 6 architecture plans, which underpin EPYC's client-side counterpart.

Instinct MI450 today, MI500-series on the roadmap

AMD Instinct GPUs are already in production use at Meta for training, ranking and recommendation workloads — the release doesn't specify which Instinct generation handles which workload today, but it does say Meta was an "early definition customer" for the Instinct MI450 architecture, the generation AMD has built around its next-gen CDNA-based compute and HBM4 memory for rack-scale AI training (part of AMD's broader Instinct accelerator lineup). As part of that work, Meta will deploy a custom Instinct GPU tailored to its own workloads, and AMD says the collaboration is already informing the design of future MI500-series GPUs.

That's a notable sequencing: Meta isn't just buying MI450 parts off a roadmap AMD built in isolation — it helped define the architecture, gets a workload-specific variant of it, and is feeding requirements into the generation after that. AMD has used nearly identical language ("early definition customer," custom silicon) to describe its relationship with OpenAI, which suggests this kind of co-design arrangement is becoming AMD's standard model for its largest AI customers, not a one-off for Meta.

Pensando networking, plus ROCm and PyTorch

The release also calls out AMD Pensando, the networking silicon AMD acquired in 2022, as part of the Meta collaboration — Pensando's data processing units (DPUs) and switch silicon are the networking layer that ties Helios racks together and to the rest of a data center. On the software side, AMD and Meta say they've worked to make Instinct GPUs usable "out of the box" through ROCm, AMD's open GPU computing stack, and that the companies have contributed improvements to PyTorch and Triton — the two most widely used open-source frameworks for training and running AI models — as well as to ROCm itself. The release specifically cites joint work optimizing Meta's recommendation models on Instinct MI350X GPUs, AMD's current-generation data-center accelerator.

This software layer matters because it's historically been Nvidia's biggest moat: CUDA's maturity, not raw chip performance, is the main reason many AI teams default to Nvidia GPUs. Every improvement AMD and Meta make to ROCm, PyTorch and Triton in production — rather than in a lab benchmark — chips away at that advantage for AMD's next customer, too, since most of that code is open source.

Day-0 support for Meta's Muse models

AMD says it provided "day-0" support for two of Meta's newer open AI models on its own hardware: Muse Spark Open runs on Instinct GPUs, and Muse Glimmer runs on Ryzen AI PCs and Radeon GPUs — meaning both models were tuned to run on AMD silicon from the moment Meta released them, rather than getting AMD support added later. From a hardware-desk perspective, the detail that matters is the client-side half of that claim: AMD is positioning ordinary Ryzen AI laptops and Radeon graphics cards, not just data-center Instinct GPUs, as the on-device companion to Meta's cloud AI stack. That's a second avenue for AMD silicon into Meta's AI ecosystem beyond the data center, even as the model work itself sits with the AI desk.

AMD's $1 billion stake in Meta

Beyond chips, AMD says in its release that it will invest $1 billion in Meta's equity offering — AMD buying a stake in Meta, structured through an equity offering rather than an open-market purchase. The release does not disclose the price, share count, or vesting terms of that investment, and AMD's own cautionary language around the announcement notes that actual outcomes (including around that investment and the pace of the broader rollout) could differ from what's described. It's worth noting this mirrors a pattern already visible elsewhere in Meta's AI buildout: Meta has also taken on hardware-tied equity arrangements with other chip and compute partners. Our earlier look at Nvidia's own multi-billion-dollar Muse investment covers a similar, separately announced equity-and-compute arrangement from the GPU side.

How this fits the February 2026 deal

None of this is happening in a vacuum. In February 2026, AMD and Meta signed a definitive multi-year, multi-generation agreement for Meta to deploy up to 6 gigawatts of AMD Instinct GPUs, with the first gigawatt of shipments expected in the second half of 2026 using the Helios rack architecture. Meta described that deal, announced on its own newsroom, as part of a "portfolio-based approach" to compute that mixes outside silicon — AMD and others — with Meta's in-house MTIA chips.

The October 27 update doesn't restate that 6-gigawatt figure; it uses the vaguer phrase "gigawatt scale" and focuses on confirming that hardware from the February agreement is now actually installed and running, plus adding the new EPYC "Verano" codename, the MI450 custom-GPU detail, and the $1 billion investment. Read together, the two releases describe a deal that started as a capacity commitment on paper in February and has, by late October, turned into physical Helios racks running real Meta workloads.

AMD's product stack inside Meta's infrastructure

Here's how the individual pieces AMD named in the October 27 release map onto Meta's infrastructure, based on the details AMD disclosed:

Why this matters for AMD's AI chip business

AMD has spent the last two years trying to convince the market that it can be a credible, large-scale alternative to Nvidia in AI data-center chips, and investors have been watching customer commitments as the clearest evidence of that. We covered how that strategy has started showing up in the numbers in our report on AMD's Q3 2026 earnings and data-center revenue growth; a named hyperscaler running production Helios racks, deploying a custom Instinct GPU, and taking an equity stake in the customer itself is exactly the kind of commitment that supports that narrative, since it's harder to walk back than a non-binding capacity pledge.

It's also a signal about where AMD sees its competitive edge: not just "faster chip," but a full rack-to-software stack — CPU, GPU, networking and open-source frameworks — co-designed with a customer big enough to actually stress-test all of it in production. That's the same playbook Nvidia has run with its own rack-scale systems and CUDA ecosystem, and AMD running even a slice of Meta's real AI infrastructure on Helios, EPYC and Instinct is evidence the company can execute that playbook at hyperscale, not just describe it in a roadmap slide.

What's next

AMD did not attach a firm calendar date to the "broader deployment" of Helios at Meta, nor to the ship date for the custom MI450-based GPU, the EPYC "Verano" generation, or the MI500-series chips the collaboration is informing — all of those remain roadmap items rather than confirmed shipping products. The terms of AMD's $1 billion investment in Meta's equity offering also haven't been detailed beyond the headline figure. Expect more specifics, including possible rack counts, additional data-center sites, or firmer EPYC and Instinct ship dates, at AMD's next quarterly earnings call or its next AI-focused event; we'll update this story as AMD or Meta disclose more.

Written byMara Lindqvist

Mara Lindqvist edits the hardware desk. She covers graphics cards, processors, memory and storage, the foundries and chip designers behind them, and what the numbers on a spec sheet mean for people choosing a PC. Specifications in her stories come from manufacturer spec pages and datasheets.

More from Mara LindqvistAll Hardware coverage