Wich inverters are used for the new H200 datacenters? How much percent of total cost?

Inverter Models Used in H200 Datacenters:

For NVIDIA H200-based datacenters (2024 generation), the most common inverters used are high-efficiency, three-phase, uninterruptible power supply (UPS) inverters designed for hyperscale and AI workloads. The specific models depend on the integrator and region, but the following are industry standards for Tier III/IV facilities hosting H200 clusters:

• Eaton 93PM G2 (208/400/480V, 50/60Hz, 96.7% efficiency, modular up to 1200 kW)
• Vertiv Liebert EXL S1 (400/480V, 97% efficiency, up to 1200 kW)
• Schneider Electric Galaxy VX (400/480V, 97% efficiency, up to 1500 kW)
• Huawei UPS5000-E (modular, up to 800 kW per unit, 97% efficiency)

These inverters are selected for their high conversion efficiency, modularity, and ability to handle the dense, high-draw AI loads typical of H200 deployments (where racks can exceed 60 kW per rack).

Percent of Total Datacenter Cost:

• Inverter/UPS systems typically account for 3–6% of total datacenter capex.
  • For hyperscale AI datacenters built for H200, inverter/UPS cost is generally 4–5% of total build cost (including batteries, switchgear, and installation).
  • Example: For a $250M facility, inverter/UPS systems would cost approximately$10M–$12.5M.

Edge Cases and Considerations:

• For colocation or retrofitted sites, inverter costs may be higher (up to 7%) due to integration complexity.
• In facilities using lithium-ion battery UPS systems, initial capex may increase, but opex decreases due to higher efficiency and lower maintenance.
• Some hyperscalers (e.g., Google, Microsoft) are piloting DC bus architectures, but the vast majority of H200 deployments in 2024 use AC-coupled, modular, double-conversion inverters as listed above.

Summary Table:

References:

• Uptime Institute 2023/2024 datacenter cost breakdowns
• Manufacturer datasheets (Eaton, Vertiv, Schneider, Huawei)
• Public hyperscale datacenter build disclosures (Meta, Microsoft, Google, NVIDIA partners)