Data center · Vendors
Who sells what, and where to read their own numbers
Each lineup below is drawn from the same dataset as the fit guide. Product names open the full record with sources; the link lists go to the vendors' own pages.
NVIDIA
The broadest range, from a 72 W inference card to 240 kW liquid-only racks. NVIDIA publishes maximum system draw and airflow for its own DGX servers, which makes those numbers the easiest to plan from; it publishes no rack power for Vera Rubin. Its software stack, CUDA, is what most AI software is written against.
- Data center productsProduct pages and positioning
- DGX documentationHub for DGX system guides
- DGX B300 user guide14.5 kW, 10U, air-cooled
- DGX GB200 hardware guideRack layout and what is liquid-cooled
- GB300 NVL72Blackwell Ultra rack
- Vera Rubin NVL72Rubin rack, 45 °C coolant inlet
AMD
Instinct accelerators match or exceed NVIDIA's memory per device at each tier, and the MI350P card gives air-cooled sites a current-generation option. AMD does not publish system power: the server figures here are power supply capacity, which overstates real draw. Its software stack is ROCm, which is open source and not CUDA-compatible without porting.
- Instinct acceleratorsProduct family overview
- MI350 seriesMI350X and MI355X overview
- MI400 seriesMI455X and roadmap
- Helios rack72-GPU liquid-cooled rack design
- ROCm documentationSoftware stack and framework support
Intel
Intel's lineup is the thinnest and the most in flux. Gaudi 3 ships today in air-cooled servers at roughly half the price of comparable NVIDIA parts on 2024 figures, but Intel has archived its open-source user-space Gaudi software, and the reported successor is not expected before 2027. The Arc Pro B70 is a low-cost inference card; Crescent Island, the next data center GPU, is only sampling.
- Gaudi AI acceleratorsCurrent Gaudi 3 products and partners
- Gaudi 3 white paperMemory, power and cooling options
- Gaudi software documentationFrameworks and supported models
- Arc Pro B70Specifications
- Crescent Island announcementIntel newsroom
Server makers
Most accelerators reach a buyer inside another company's server, and the server maker's documents are often the only place system power and cooling are written down.
- Supermicro GPU systemsNVIDIA, AMD and Intel platforms
- Supermicro GB200 NVL72 datasheetRack power and cooling unit options
- HPE GB300 NVL72 briefNominal, peak and provisioning power
- Dell PowerEdge XE spec sheetDell's AI server line, air and liquid
- Lenovo PressDetailed product guides per accelerator
- Gigabyte Vera Rubin NVL72Source of the 240 kW rack figure
Independent references
- MLCommons MLPerfAudited training and inference benchmarks across vendors
- Open Compute ProjectOpen rack and module standards used by AMD and Intel systems
- ASHRAEData center thermal guidelines for air and liquid cooling
- US Energy Information AdministrationAverage household electricity use, the basis for the "homes" comparison
Figures come from public vendor, server-maker and press sources, last checked on October 7, 2026. How the numbers were sourced