Rack-Scale AI System Market - By Platform Generation (GB200/GB300 NVL72, Vera Rubin NVL144, Kyber-Class 576 GPU, Custom ASIC Racks), By Rack Power (Below 150kW, 150-300kW, Above 300kW), By Buyer (Hyperscaler, Neocloud, Sovereign AI, Enterprise), By Region

The global rack-scale AI system market size was USD 96.28 Billion in 2025 and is expected to register a revenue CAGR of 18.2% during the forecast period. A rack-scale AI system integrates accelerators, CPUs, scale-up interconnect, memory, storage, power shelves, and liquid cooling into a single engineered rack sold and deployed as one unit, with NVIDIA's GB200 and GB300 NVL72 defining the category at 72 GPUs per rack drawing 120 to 130 kilowatts. The supply chain concentrates in Taiwan ODMs including Foxconn, Quanta, and Wistron building trays and racks against NVIDIA MGX reference designs supported by more than 50 partners, while AMD's Helios rack-scale platform and hyperscaler custom ASIC racks broaden the architecture base. NVIDIA's Vera Rubin platform entered full production in January 2026 with the VR200 NVL144 rack specified at roughly 190 to 230 kilowatts, and the 2027 Kyber-class rack connects 576 Rubin Ultra GPUs at approximately 600 kilowatts.

Hyperscaler and neocloud capacity contracting is the primary revenue growth driver, because AI infrastructure operators purchase compute in rack and multi-rack blocks against multi-year model training and inference commitments, and the deployment cadence is now annual. NVIDIA opened GTC 2026 citing approximately one trillion dollars in orders through 2027 across its rack-scale platforms, and CoreWeave, Lambda, Nebius, Oracle Cloud Infrastructure, and Together AI are designing facilities for the 800 volt DC power architecture that Kyber-class racks require. Foxconn detailed its 40 megawatt Kaohsiung-1 facility in Taiwan built specifically for 800 VDC rack deployment, and the Vera Rubin NVL144 design introduced 45 degree Celsius liquid cooling, a liquid-cooled busbar, twenty times more rack energy storage, and a printed circuit board midplane replacing cable harnesses for faster assembly and serviceability. For instance, in January 2026, NVIDIA Corp., USA, confirmed at CES 2026 that Vera Rubin had entered full production with volume rack shipments scheduled for the second half of 2026 across more than 50 MGX ecosystem partners. These are some of the key factors driving revenue growth of the market.

However, rack-scale systems have outrun the facilities that must host them, because a VR200 NVL144 rack at 190 to 230 kilowatts exceeds the electrical and thermal envelope of the vast majority of existing data centres, and retrofit costs for cooling alone run tens of thousands of dollars per rack before electrical plant upgrades. The annual platform cadence, Rubin in 2026, Rubin Ultra Kyber in 2027, Feynman in 2028, invalidates facility designs faster than construction cycles complete, creating stranded asset risk for operators who build to a single generation's specification. Component concentration compounds the exposure, since TSMC advanced packaging, HBM supply sold out through 2026, and Taiwan-centred ODM assembly mean a single regional disruption propagates through the entire category, and rack prices in the millions of dollars concentrate buyer credit risk in a small set of heavily leveraged neocloud operators. These factors substantially limit rack-scale AI system market growth over the forecast period.