Advancing AI 2026: AMD and Schneider Electric's Helios Reference Design Signals a New Bar for AI Rack Deployment

AMD first unveiled Helios as a unified rack-scale architecture for frontier AI training and inference at its Advancing AI conference in 2025. At Advancing AI 2026 in San Francisco, AMD moved Helios from architecture to product, detailing the full Helios rack-scale solution built on Meta's new Open Rack Wide (ORW) standard submitted to the Open Compute Project (OCP), combining 72 Instinct MI455X Graphics Processing Units (GPUs), EPYC "Venice" Central Processing Units (CPUs), and Pensando Digital Processing Units (DPUs)/Network Interface Controllers (NICs) in an open, double-wide rack, with engineering samples set to ship in 2H 2026 and mass production beginning in 2Q 2027. Alongside the specs and timeline, Schneider Electric and AMD released their first jointly engineered reference design for deploying Helios, giving operators a pre-validated blueprint for high-density Artificial Intelligence (AI) clusters, supporting racks up to 246 Kilowatts (kW) and clusters up to 10.4 Megawatts (MW) of Information Technology (IT) capacity.

The event is notable less for the spec sheet than for whom AMD brought on stage with it. A chip vendor pairing directly with a power and cooling incumbent to co-engineer a deployment blueprint is a sign that compute and passive data center infrastructure are converging into a single design process, rather than being specified separately.

What makes the Schneider Electric and AMD announcement significant is less the rack itself than what it turns Helios into: not just a high-density compute concept, but an infrastructure blueprint that operators can actually build against. The jointly validated reference design supports 246 kW AI racks and modular clusters up to 10.4 MW of IT load, with documentation covering power distribution, liquid and air cooling, electrical infrastructure, floor layouts, and thermal and electrical simulation. That level of detail matters because AI capacity growth is running up against a facility problem as much as a chip supply one.

ABI Research forecasts that global active data center IT capacity will increase from 24.4 Gigawatts (GW) in 2025 to 147.1 GW by 2035, an addition of roughly 122 GW, with AI workloads growing from 38% of data center demand in 2026 to 51% of active capacity by 2031. A rack with Helios' footprint does not sell itself on paper: the OCP ORW format is nearly twice the size of NVIDIA's NVL72, and every operator now has to translate that footprint into floor loading, power architecture, and cooling performance before a single unit goes in. Schneider Electric's design compresses that translation work into a pre-validated facility blueprint. That changes the deployment decision: operators now choose, finance, and build around an entire facility design, rather than just a GPU rack.

The timing also lines up with a broader shift that ABI Research is tracking in who is building AI capacity. Tier One hyperscalers still hold the majority of data center capacity in 2026 at around 55%, but neoclouds are the fastest-growing class of operator, with AI-dedicated cloud capacity expected to climb from 832 MW in 2025 to nearly 31 GW by 2035 and neocloud capacity expanding at a 46.3% Compound Annual Growth Rate (CAGR). ABI Research also expects reference design adoption itself to become close to standard practice, projecting 61% of new data centers built in 2026 to use at least one reference design, increasing to 86% by 2030, concentrated most heavily in AI-first and hyperscale facilities.

Against that backdrop, the Helios blueprint reads less like a one-off product launch and more like a recognition that deployment certainty is now a competitive lever. Particularly, as the landscape of data center operators is diversifying (think about more and more neocloud providers moving to build their own data centers), validated reference designs are becoming even more important as they provide reliability and reassurance for these new operators with a limited set of skills and expertise when it comes to data center management and operation.

Motivair’s cooling technology underpins the design, and states that the system can remove up to 84% of heat through liquid cooling and hit a modeled full-load Power Usage Effectiveness (PUE) between 1.121 and 1.137, while supporting both greenfield builds and retrofits in existing facilities. For neoclouds and other newer self-build operators without deep in-house data center engineering, a validated architecture like this offers a fast path to credibility they would otherwise have to earn deployment by deployment. For colocation providers, it raises the stakes: customers wanting more control inside the shell will increasingly expect their colocation partner to offer not just space and power, but an environment already proven to handle validated, double-wide, and ultra-dense deployments like Helios.

The Schneider Electric AMD reference design sets a new baseline for what "launch-ready" means for rack-scale AI platforms, and it puts pressure on every other player in the value chain, including competing infrastructure vendors, Original Equipment Manufacturers (OEMs) building on Helios, and the operators actually deploying it, to react rather than wait and see. ABI Research expects this pattern to repeat with every major rack-scale launch going forward, which means the ecosystem's response now matters as much as the announcement itself.

For competing power and cooling vendors, the response should be immediate parity, not differentiation for its own sake. The move parallels rival Vertiv's NVIDIA reference designs, as infrastructure vendors race to support AMD's growing GPU share, and any vendor without a published, engineering-validated reference design for a major rack-scale platform is now visibly behind on a criterion that operators are starting to shortlist against. Vertiv, Eaton, and regional System Integrators (SIs) should expect Requests for Proposal (RFPs) to start asking for Helios-specific or equivalent blueprints, not just generic high-density capability statements

For OEMs and SIs such as HPE and Supermicro building Helios-based systems, the Schneider Electric design should be treated as the reference architecture to build compliant Stock-Keeping Units (SKUs) around, reducing the risk of shipping racks that customers then struggle to site. For colocation and hyperscale operators, the practical takeaway is to request the reference design documentation directly rather than relying on AMD or Schneider Electric's marketing claims, and to weight retrofit-readiness heavily given how much near-term AI capacity growth is happening inside existing shells rather than new builds. Finally, ABI Research expects NVIDIA's ecosystem partners, including Schneider Electric itself, given its existing collaboration with NVIDIA on cooling, power, and rack systems across Europe, to respond with their own updated or expanded reference designs to keep pace, reinforcing reference design publication as a recurring competitive cycle rather than a one-off announcement.