Data centre design reference

Generic reference designs for power, cooling, controls and operations

Power for AI halls

What changes in the power system when the load is a hall of AI accelerators: density, load swings and growth.

On this page
  1. Density
  2. Load swings from training jobs
  3. Mitigation at three levels
  4. How much UPS an AI hall needs
  5. 800 V DC
  6. Sources

Density

AI racks draw an order of magnitude more than the racks most halls were designed for. A current rack-scale GPU system is specified at 135 kW, with up to 155 kW at peak.1 A later generation, announced for 2027, was designed to draw 600 kW per rack.2 A 2026 write-up says the design shown that year appeared to have lost half its GPU density, and it is unclear whether that rack is now a 300 kW rack,3 so treat 600 kW as an upper planning case. A hyperscaler has published a ±400 V DC rack power design that can support up to 1 MW per rack, saying machine learning will need more than 500 kW per rack before 2030.4

At those densities almost all heat goes to liquid, the rack's power feed becomes a design item of its own, and a power block holds far fewer racks than before.

Load swings from training jobs

A large training job runs thousands of GPUs in lock-step: they compute, then wait together for data to be exchanged, and the hall's power swings with every iteration. One study reports that in extreme cases a data centre's power can oscillate by tens of megawatts, recurring from about once a second to once every tens of seconds, with most of the energy between 0.2 and 3 Hz.5 A newer preprint notes that iteration periods of one to ten seconds put the forcing frequency in the band of the grid's inter-area oscillation modes, where a training schedule can drive a mode at resonance.6

The North American reliability body, NERC, recorded a ramp of 1.9 per unit per second, lasting about 250 ms, when a 50 MW block of a 200 MW AI training data centre started a training run.7 It also describes a transmission fault after which about 1,500 MW of load, mostly data centres, was lost at once: the circuit breakers at the data-centre substations did not trip, but several data centres moved part of their supply to backup systems.7 NERC also notes that such forced oscillations may age generators faster, while their reliability risks are not always well defined.7

Mitigation at three levels

The research names three families of mitigation: a secondary workload that soaks up power in the troughs, GPU firmware that smooths power above a set floor, and energy storage.5 Smoothing has a price: in one simulation it cost 10.5 % extra energy with the floor at 90 % of the GPUs' thermal design power.5 On the facility side, a grid-interactive UPS running in grid-forming mode has been proposed to support the grid against these oscillations,8 and battery storage alongside the UPS needs a smoothing reserve sized in, so that smoothing does not eat into the guaranteed backup time.9

Design position. Specify mitigation at all three levels and test it with a recorded training profile: power smoothing in the IT, a UPS or storage system with a defined smoothing duty at the hall, and generators that meet their transient class under the real load profile. Agree the limits with the grid company before the connection agreement is signed, and design ride-through so that a short grid disturbance does not send the whole hall to generators.

How much UPS an AI hall needs

Here the industry disagrees, and the disagreement is worth stating plainly. NERC reports that for AI training loads there may be no UPS on the IT equipment at all: training clusters save checkpoints at regular intervals and restart from the last one after a power loss.7 At the same time, a data-centre operator writes that for AI data centres with zero downtime tolerance, 2N or higher is increasingly the standard.10

Between the two sits block redundancy: one shared reserve block instead of a duplicate of every block, with 33 % spare capacity at 4N/3 or 50 % at 3N/2, plus transfer switches.11

Alternative: block-redundant UPS with a shared reserve system and static transfer switches N block UPS systems each carry one block. One reserve UPS system, rated like the others, feeds a reserve bus. A static transfer switch per block moves the block to the reserve bus within a fraction of a cycle when its own UPS fails or is taken out for maintenance. Installed UPS capacity is (N+1)/N of the load instead of 2N. Reserve bus AC/DCDC/AC UPS 1 block 1 AC/DCDC/AC UPS 2 block 2 AC/DCDC/AC UPS N block N AC/DCDC/AC Reserve UPS rated as one block STS STS 1 STS STS 2 STS STS N Block 1 racks Block 2 racks Block N racks 1. Each UPS input comes from the block's LV switchboard (not drawn). Racks can stay single-corded through the STS, or be dual-corded with the second cord on a separate path. 2. Fewer UPS units and batteries than 2N, at the cost of a transfer through the STS and a reserve system shared by all blocks: fit it where the IT load tolerates that and the availability target allows it. 3. Lines that cross without a dot are not connected.
Alternative: block-redundant UPS with a shared reserve system and static transfer switches
ApproachInstalled UPS for N blocksFitsWatch
2N per block2NMixed or inference loads, contracts that promise no interruptionCapital cost, each side lightly loaded
Block redundant (reserve)N+1Large halls of similar blocksTransfer through static switches, the shared reserve
No UPS on training ITUPS for cooling and controls onlyTraining clusters that checkpointLost work per outage, ride-through, the load swing on generators

Design position. Decide per power block (and so per tenant), and write the decision into the design basis, with the checkpoint interval and the step-load of a synchronised restart. Where blocks of different classes share switchgear, bypass or controls, keep their fault domains separate. Whatever the IT gets, critical cooling and the control systems stay on protected supply: a training cluster can restart from a checkpoint, but a CDU that stops takes the racks with it.

800 V DC

A GPU maker has announced an 800 V DC power infrastructure to support 1 MW racks starting in 2027,12 and says 800 V busways instead of 415 V AC distribution enable 85 % more power through the same conductor size.12 It names both conventional transformer-based and solid-state transformer approaches for the change.12 A power-equipment maker expects full electrical-code support for 800 V DC only around 2029, and says the missing provisions concern hall-level DC distribution rather than DC power racks beside the IT racks.13

Direction of travel: medium voltage converted once to an 800 V DC bus that feeds the racks, with storage on the DC bus A solid-state transformer or rectifier converts medium-voltage AC to 800 V DC at the edge of the hall. A DC bus feeds the compute racks, which convert down on board. Energy storage on the DC bus rides through fast load swings. Shown as a direction to prepare for, not a design to build today. 800 V DC bus MV supply from the site switchboard MV breaker ACDC Solid-state transformer or transformer and rectifier, MV AC to 800 V DC DC breaker DC breaker Energy storage rides through swings Compute rack 800 V DC in Compute rack 800 V DC in Compute rack 800 V DC in 1. Fewer conversion stages and less copper than 400 V AC distribution when racks draw several hundred kilowatts. 2. Protection, earthing and arc-flash practice for DC at this voltage are still settling: check current standards and the platform maker's specification.
Direction of travel: medium voltage converted once to an 800 V DC bus that feeds the racks, with storage on the DC bus

Design position. Halls designed now stay AC at facility level. Where a rack platform needs DC, feed it from a row-level DC power rack on the AC hall; one platform maker announces an 800 VDC power rack for the second half of 2026.14 The published open rack specification is ±400 V, and a later revision is to add native 800 V.15 Reserve space, medium-voltage capacity and cable routes for a later facility-level DC conversion, and do not build hall-level 800 V DC distribution ahead of the codes and a published specification.

Sources

  1. Lenovo NVIDIA GB300 NVL72 Rack Scale AI Product Guide. Lenovo Press, 2026-08-30. Read 2026-10-18.https://lenovopress.lenovo.com/lp2357-lenovo-nvidia-gb300-nvl72-rack-scale-ai
  2. Nvidia's Vera Rubin CPU, GPUs chart course for 600kW racks. The Register, 2025-03-19. Read 2026-10-18.https://www.theregister.com/2025/03/19/nvidia_charts_course_for_600kw/
  3. NVIDIA Kyber. Glenn K. Lockwood (personal site), 2026-04-02. Read 2026-10-18.https://www.glennklockwood.com/garden/kyber
  4. Enabling 1 MW IT racks and liquid cooling at OCP EMEA Summit. Google Cloud, 2025-04-30. Read 2026-10-18.https://cloud.google.com/blog/topics/systems/enabling-1-mw-it-racks-and-liquid-cooling-at-ocp-emea-summit
  5. Power Stabilization for AI Training Datacenters. arXiv (preprint), 2025-08-21. Read 2026-10-18.https://arxiv.org/html/2508.14318
  6. A Pre-Dispatch Resonance Safety Criterion for AI Training Clusters. arXiv (preprint), 2026-06-20. Read 2026-10-18.https://arxiv.org/abs/2606.22096
  7. Characteristics and Risks of Emerging Large Loads: Large Loads Task Force White Paper. North American Electric Reliability Corporation (NERC), 2025-07. Read 2026-10-18.https://www.nerc.com/globalassets/who-we-are/standing-committees/rstc/whitepaper-characteristics-and-risks-of-emerging-large-loads.pdf
  8. Operational Risks in Grid Integration of Large Data Center Loads: Characteristics, Stability Assessments, and Sensitivity Studies. arXiv (preprint), 2025-10-24. Read 2026-10-18.https://arxiv.org/html/2510.05437v2
  9. Grid Integration of AI Data Centers: A Critical Review of Energy Storage Solutions. arXiv (preprint, review), 2026-02-28. Read 2026-10-18.https://arxiv.org/html/2603.00415v1
  10. Redundant Data Center Power for AI: Why It's Non-Negotiable. Hanwha Data Centers, 2026-03-20. Read 2026-10-18.https://www.hanwhadatacenters.com/blog/redundant-data-center-power-for-ai-why-its-non-negotiable/
  11. UPS & Energy Storage: From Ride-Through to Transient Absorption. The Definitive Guide to AI Data Centers (aidatacenterguide.com). Read 2026-10-18.https://aidatacenterguide.com/part-4-electrical-and-energy-infrastructure/4-5-ups-and-energy-storage-from-ride-through-to-transient-absorption
  12. NVIDIA 800 VDC Architecture Will Power the Next Generation of AI Factories. NVIDIA Developer Blog, 2025-05-20. Read 2026-10-18.https://developer.nvidia.com/blog/nvidia-800-v-hvdc-architecture-will-power-the-next-generation-of-ai-factories/
  13. The 800 VDC decision: A practical guide for AI power architecture. Vertiv, 2026-09-14. Read 2026-10-18.https://www.vertiv.com/en-us/insights/articles/educational-articles/the-800-vdc-decision-a-practical-guide-for-ai-power-architecture/
  14. Why Scaling AI Compute Performance Requires a New Power Architecture. NVIDIA, 2026-08-11. Read 2026-10-18.https://blogs.nvidia.com/blog/800-vdc-power-architecture-ai-factory/
  15. Powering the Next Era of AI: How Google, Microsoft and Nvidia Are Standardizing and Accelerating the Industry Transition to LVDC. Open Compute Project Foundation, 2026-08-11. Read 2026-10-18.https://www.opencompute.org/news/powering-the-next-era-of-ai-how-google-microsoft-and-nvidia-are-standardizing-and-accelerating-the-industry-transition-to-lvdc/