Rackspace Technology and AMD signed a definitive agreement on June 16, 2026 to deploy an initial 30 MW of AMD-based AI compute across Rackspace data centers from late 2026 through 2028, according to Barron's coverage of the announcement. RXT shares moved roughly 25 percent in premarket trading on the news, and the company paired the deal with a 15 percent workforce cut to fund its pivot toward what it calls a governed AI cloud for regulated industries. The agreement operationalizes a memorandum of understanding the two companies disclosed on May 7.
The footprint runs on AMD Instinct MI355X and MI350P GPUs paired with EPYC CPUs. The MI355X carries a 1,400W TDP, roughly 40 percent above the air-cooled MI350 variant, and AMD has positioned it for direct liquid cooled deployments. A standard eight-GPU node draws around 11 kW at full load. At that density, air cooling stops being a serious option. Rackspace is not buying chips here so much as committing to a heat rejection architecture, and the 30 MW figure is a thermal budget before it is a compute budget.
Liquid cooling is what lets the deal scale. With direct-to-chip cold plates, AMD rates the MI355X platform for as many as 128 accelerators per rack, double the 64-GPU ceiling of the lower-TDP air-cooled configuration. That doubling is the entire economic case for going wet, and it mirrors the broader shift where liquid cooling has become the default rather than an upgrade for current-generation accelerators. Rackspace's phased rollout through 2028 implies a multi-year liquid plumbing commitment across its sites, not a one-time retrofit.
Thirty megawatts of GPU compute concentrated into high-density racks reshapes the heat rejection load at every site that takes the deployment. The chips drink the watts, and the cooling loop has to move that energy out continuously. That is the recurring tension operators keep getting wrong on the water-versus-power axis: a closed liquid loop trims water draw relative to evaporative towers but raises the stakes on coolant distribution units, manifolds, and leak management. For a managed provider targeting healthcare and finance, cooling uptime becomes a compliance variable, not just an efficiency metric.
The regulated-industry angle sharpens the cooling requirement. Clinical AI inference and financial workloads carry availability guarantees, and a CDU failure or a thermal throttle event reads as a service-level breach in those contracts. Rackspace is selling accountability from silicon to outcome, which means it owns the thermal envelope end to end. The vendor selection for cold plates, CDUs, and rack manifolds now sits on the critical path of the company's repositioning.
The market reaction tracked the financing and the layoffs, but the durable signal is architectural. Rackspace committed to a fleet of 1,400W accelerators that cannot run on legacy air-cooled halls, which pulls forward liquid infrastructure spend across its footprint through 2028. That capex pattern is showing up across the sector, where liquid cooling buildouts are becoming the investable thesis behind AI infrastructure deals. For Rackspace, the AMD partnership is a bet that owning the heat is the same as owning the workload.