How much power does a single AI rack consume in 2026?

In 2026, AI rack power consumption ranges dramatically depending on the hardware configuration. A standard GPU server rack with 4-8 NVIDIA H100 or H200 servers typically draws 40-80 kW. An NVIDIA GB200 NVL72 rack (the compute tray plus networking and cooling infrastructure) draws approximately 120 kW per rack. The upcoming NVIDIA GB300 NVL144, which doubles the GPU count per rack, is expected to draw 240-280 kW. Custom AI accelerator deployments from Google (TPU v6), AMD (MI350), and startups like Cerebras and Groq can push individual rack power to 100-200 kW. The industry trajectory points toward 400 kW to 1 MW per rack within the next 2-3 years as GPU thermal design power continues to increase and rack-scale architectures integrate more accelerators per enclosure.

What is a busbar power distribution system?

A busbar power distribution system uses rigid copper or aluminum conductors (busbars) instead of traditional cables and whips to deliver electricity from the main switchboard to individual racks. The busbar runs overhead or under the raised floor along the row of racks, and tap-off boxes connect individual racks to the busbar at any point along its length. Busbar systems are preferred for high-density deployments (above 30-50 kW per rack) because they can carry significantly higher currents than cable-based PDUs, support hot-swappable tap-off connections for maintenance without shutting down the entire row, offer better space utilization (no cable trays full of heavy power cables), and provide more consistent voltage drop characteristics across long row runs. Major busbar manufacturers for data centers include Schneider Electric (Canalis), Siemens (SIVACON 8PS), Eaton (Powerware), and Starline.

Can existing data centers be retrofitted for megawatt-scale racks?

Retrofitting existing data centers for megawatt-scale rack power is theoretically possible but practically challenging. The primary constraints are electrical infrastructure (transformer capacity, switchgear ratings, and conductor sizing designed for 5-10 kW per rack cannot support 100+ kW without complete replacement), structural loading (a fully loaded liquid-cooled AI rack can weigh 2,000-3,000 kg, exceeding the floor loading capacity of most existing facilities designed for 1,000-1,500 kg/rack), cooling infrastructure (air cooling systems cannot remove more than approximately 30-40 kW per rack; liquid cooling requires new piping, CDUs, and potentially external heat rejection upgrades), and ceiling height (overhead busbar and liquid cooling manifolds require vertical clearance that older facilities may lack). Most operators find that purpose-built new construction or modular additions are more cost-effective than deep retrofits above the 50 kW per rack threshold.

What is the difference between rack-level and row-level power distribution?

Rack-level power distribution places the final power conversion and distribution equipment inside or immediately adjacent to the individual rack. Each rack has its own PDU (power distribution unit) that steps down voltage, provides circuit protection, and distributes power to individual servers via outlet strips. This approach works well up to approximately 30-50 kW per rack but becomes impractical at higher densities due to the physical size and heat output of the PDU itself. Row-level power distribution moves the power conversion equipment to end-of-row or overhead positions that serve an entire row of racks. Busbar systems and row-level RPPs (remote power panels) distribute power along the row, with minimal conversion hardware at each rack. This approach scales better to 100+ kW per rack because it centralizes the power conversion losses outside the rack envelope, allows higher-voltage distribution closer to the load, and simplifies rack-level cabling.

How does liquid cooling affect power delivery architecture?

Liquid cooling fundamentally changes power delivery architecture in three ways. First, it eliminates server fans as a power load: in air-cooled systems, server fans consume 10-15% of total system power; direct-to-chip liquid cooling removes this load entirely, meaning a 100 kW liquid-cooled rack delivers more compute per watt than a 100 kW air-cooled rack. Second, it relocates the cooling power load from the rack to the CDU (coolant distribution unit) and facility chiller plant, which changes the electrical distribution topology: rack-level circuits carry only IT load, while cooling infrastructure has its own dedicated power feed. Third, liquid-cooled systems typically operate at higher power densities per rack, which shifts the economic break-even point for busbar vs cable distribution: facilities that would use cable-based PDUs at 20 kW per rack almost always require busbar or overhead busway systems at 80+ kW per rack.

Megawatt-Scale Rack Power Delivery: 100 kW to 1 MW Per Rack for AI & HPC Data Centers [2026]

The traditional data center rack drew 5 to 8 kilowatts. A decade ago, high-density meant 15 to 20 kW. Today, a single NVIDIA GB200 NVL72 rack draws 120 kW, and the next generation of AI accelerators is pushing individual rack power demands toward 400 kW and beyond. This is not a gradual trend line -- it is a discontinuity that renders the entire electrical distribution architecture of conventional data centers obsolete.

Delivering 100 kilowatts to a single rack requires different conductors than delivering 8 kilowatts. Delivering 400 kilowatts requires a fundamentally different distribution topology. And delivering 1 megawatt to a single compute enclosure (as some custom AI training clusters now demand) requires rethinking the relationship between the utility transformer and the silicon that consumes the power.

This guide covers the engineering of megawatt-scale rack power delivery: the electrical architectures, distribution equipment, conductor systems, protection schemes, and cooling integration strategies that enable the next generation of AI and HPC infrastructure at rack densities 10 to 100 times higher than traditional IT deployments.

The Power Density Problem: Why Traditional Distribution Fails

A conventional data center distributes power through a chain of transformations: utility voltage (typically 11 kV to 33 kV) steps down through a main transformer to 480V or 415V, feeds UPS systems, then distributes through switchgear to floor-level power panels, which feed individual rack PDUs via cable whips. This architecture was designed for a world where each rack drew 5 to 15 kW, and 30 to 40 racks shared a single 200A to 400A power panel.

At 120 kW per rack, a single GB200 NVL72 rack draws more current than an entire row of traditional servers. At 480V three-phase, 120 kW requires approximately 145 amps -- roughly the capacity of a single cable whip that previously served 8 to 10 racks. At 400 kW per rack, the current draw approaches 480 amps at 480V, which exceeds the rating of most standard power cables and requires either higher-voltage distribution or parallel conductor paths.

The physical constraints compound rapidly. A 4 AWG copper cable (common for 30A rack whips) weighs approximately 0.25 kg per meter. A 500 MCM cable (required for 400A feeds) weighs approximately 3.3 kg per meter. Running 500 MCM cables from a power panel to every rack in a 20-rack row would require cable tray capacity that physically does not fit in the overhead space of most existing data centers, and the aggregate weight would exceed the structural rating of conventional cable tray systems.

This is why megawatt-scale racks cannot simply use "bigger cables." The entire distribution architecture must be reconsidered.

Power Distribution Architectures for High-Density Racks

Overhead Busbar Systems

Overhead busbar systems have emerged as the dominant power distribution method for racks drawing 50 kW and above. A busbar is a rigid copper or aluminum conductor assembly that runs the length of a rack row, typically mounted overhead at a height of 2.5 to 3 meters. Individual racks connect to the busbar via tap-off boxes (also called plug-in units) that can be installed, removed, or repositioned along the busbar without de-energizing the entire system.

The engineering advantages of busbar distribution for high-density racks are substantial. A single 4,000A busbar trunk can deliver approximately 2.7 MW at 415V three-phase along a row of 20+ racks, providing ample capacity for 120 kW per rack with significant headroom for future density increases. The voltage drop across a well-designed busbar run is typically less than 1 percent over 30 meters, compared to 2 to 3 percent for equivalent cable runs, which translates directly to improved power efficiency at the rack level.

Hot-swappable tap-off boxes are particularly valuable in AI data center operations. When a rack needs to be decommissioned, upgraded, or relocated, the tap-off box can be disconnected and reconnected without affecting power to adjacent racks. This is critical for facilities that frequently swap GPU generations or reconfigure cluster topologies, as it eliminates the need for planned outages across an entire power panel zone.

Major busbar systems designed for data center deployment include Schneider Electric Canalis KT (rated up to 5,000A), Siemens SIVACON 8PS (up to 6,300A), Eaton Powerware Busway (up to 5,000A), and Starline Track Busway (up to 1,250A for lower-density applications). The choice between systems depends on the current capacity required, the mechanical mounting constraints of the facility, and the vendor ecosystem already present in the electrical infrastructure.

Under-Floor Busbar Systems

In facilities with raised-floor construction, busbar systems can run beneath the floor instead of overhead. Under-floor busbar preserves overhead clearance for liquid cooling manifolds and network cabling, and keeps the high-current conductors physically separated from the low-voltage data cabling that runs overhead in most modern designs.

The trade-off is accessibility: under-floor busbars require floor tile removal for tap-off box installation and maintenance, which is more disruptive than overhead systems where tap-off boxes are accessible from standing position. For this reason, overhead busbar is generally preferred for new construction, while under-floor busbar is more common in retrofits of existing raised-floor facilities.

Direct-Feed Architecture (Transformer to Rack)

For the highest power densities (400 kW to 1 MW per rack), some operators are moving toward architectures that minimize the number of conversion steps between the utility transformer and the rack. In a direct-feed architecture, a dedicated step-down transformer feeds a small number of racks (4 to 8) through a short busbar or direct cable run, bypassing the traditional switchboard and power panel hierarchy entirely.

This approach reduces distribution losses, eliminates intermediate protection equipment, and simplifies the power path. The trade-off is granularity: each transformer serves a fixed number of racks, and unused capacity on one transformer cannot be shared with racks on a different transformer feed. This architecture works best for homogeneous deployments (every rack draws the same power) but is inflexible for mixed-density environments.

NVIDIA's reference architecture for GB200 NVL72 clusters recommends a variation of this approach: a dedicated 1 MVA transformer per cluster of 8 NVL72 racks, with overhead busbar distribution from the transformer's secondary to each rack position. This delivers approximately 960 kW of IT capacity (120 kW x 8 racks) from a single transformer, with the remaining 40 kW of transformer capacity allocated to cooling infrastructure and distribution losses.

Power Distribution Architecture Comparison

Architecture	Capacity Per Row	Best For	Voltage Drop	Flexibility
Cable Whip PDU	Up to 500 kW	5-30 kW/rack	2-3%	Low (fixed circuits)
Overhead Busbar	Up to 3 MW	50-200 kW/rack	0.5-1%	High (hot-swap taps)
Under-Floor Busbar	Up to 2 MW	50-150 kW/rack	0.5-1.5%	Medium (floor access)
Direct Transformer Feed	Up to 1 MW	200 kW-1 MW/rack	Less than 0.5%	Low (fixed allocation)

Voltage Distribution Strategy: 415V, 480V, or Higher

The distribution voltage directly affects conductor sizing, distribution losses, and equipment compatibility. Higher voltages deliver the same power with lower current, enabling smaller conductors and lower I2R losses. The trade-off is equipment compatibility and safety considerations.

480V Three-Phase (North American Standard)

In North American facilities, 480V three-phase is the standard distribution voltage for data center power. At 480V, a 120 kW rack draws approximately 145A, which is within the capacity of a single 200A tap-off box. This voltage level is directly compatible with most server PSUs through rack-level step-down transformers or auto-transformers, and the extensive availability of 480V-rated components simplifies procurement and maintenance.

415V Three-Phase (International Standard)

International facilities (including the UAE) typically distribute at 415V three-phase (240V line-to-neutral). At 415V, the same 120 kW rack draws approximately 167A, requiring a larger tap-off box or paralleled connections. While the higher current at lower voltage increases conductor sizing requirements by approximately 15 percent compared to 480V, the voltage difference does not fundamentally change the distribution architecture.

Medium Voltage Distribution (1 kV to 3.3 kV)

For the densest deployments approaching 400 kW to 1 MW per rack, some advanced facilities are exploring medium voltage distribution within the data hall. By distributing power at 1 kV to 3.3 kV and stepping down to IT voltage at the rack or row level, conductor sizes are reduced dramatically. A 400 kW load at 3.3 kV draws only 70A, compared to 480A at 480V.

The challenge is safety: medium voltage distribution requires qualified high-voltage personnel, arc-flash-rated enclosures, and specialized protection equipment. The installed cost per point of connection is significantly higher than low-voltage distribution. Current adoption is limited to hyperscale operators (Google, Microsoft, Meta) building custom facilities where the long-term operational savings justify the higher upfront engineering cost.

GPU and AI Accelerator Rack Power Profiles

Understanding the power profile of specific AI hardware is essential for designing the distribution system. Unlike traditional servers that draw relatively constant power, GPU-accelerated workloads exhibit significant power variability depending on whether the GPUs are idle, training, or running inference.

NVIDIA GB200 NVL72

The GB200 NVL72 is NVIDIA's current flagship rack-scale AI training system. It integrates 36 Grace CPUs and 72 Blackwell GPUs into a single liquid-cooled rack consuming approximately 120 kW at full training load. The power profile is relatively stable during continuous training workloads, with variations of approximately 10 to 15 percent between idle and full load.

The NVL72 requires 480V three-phase input through redundant (A+B) power feeds, each capable of carrying the full rack load for failover. This means the distribution system must provision 240 kW of capacity per rack (2x 120 kW) even though steady-state consumption is 120 kW. Static transfer switches at the rack level handle failover between the A and B feeds.

NVIDIA GB300 NVL144 (Next Generation)

The upcoming GB300 NVL144 doubles the GPU count per rack to 144 Blackwell Ultra GPUs, with projected power consumption of 240 to 280 kW per rack. This effectively doubles the per-rack power requirement and pushes the distribution architecture into territory where overhead busbar becomes essential rather than optional.

At 280 kW and 480V, a single GB300 NVL144 rack draws approximately 337A at full load, exceeding the capacity of a standard 200A tap-off box. Facilities planning for NVL144 deployments must either use parallel tap-off connections (2x 200A), higher-rated tap-off boxes (400A+), or transition to direct-feed architectures for these specific rack positions.

Custom AI Accelerator Racks (400 kW to 1 MW)

Several AI hardware companies are designing rack-scale systems that exceed 400 kW. Cerebras' wafer-scale systems, Google's TPU pods, and custom ASIC-based training clusters for large language models can approach or exceed 500 kW per rack equivalent. At these power levels, the rack itself becomes essentially a substation, requiring its own dedicated protection, metering, and monitoring equipment.

The electrical design for a 1 MW rack position typically includes a dedicated dry-type transformer (1 MVA or larger), a dedicated circuit breaker panel with individual feeder protection, current transformers for per-phase metering, and a local HMI (human-machine interface) for monitoring and control. The total footprint of the electrical support equipment often equals or exceeds the footprint of the compute rack itself, which must be factored into floor space planning.

Liquid Cooling Integration with Power Delivery

At power densities above 30 to 40 kW per rack, air cooling becomes physically insufficient. Every megawatt-scale rack is liquid cooled, and the cooling system's power requirements and physical routing must be coordinated with the electrical distribution design.

Cooling Power Budget

Liquid-cooled racks eliminate server-level fans but introduce coolant distribution units (CDUs), pumps, and heat rejection equipment that draw their own power. A typical CDU serving 4 to 8 high-density racks consumes 5 to 15 kW, and the facility-level chiller plant adds additional load. The total cooling overhead for a well-designed liquid-cooled facility achieves a PUE of 1.03 to 1.10, meaning 3 to 10 percent of total facility power goes to cooling.

The key design decision is whether to power CDUs from the same distribution system as the IT racks or from a separate utility feed. Powering CDUs from the IT distribution simplifies metering (cooling power is included in the per-rack measurement) but reduces the effective IT power available from each busbar run. Separating CDU power onto dedicated feeds provides more accurate IT power accounting and enables independent UPS protection for cooling systems, which is critical because a cooling failure at 120 kW per rack can cause thermal shutdown within minutes.

Physical Coordination

In overhead busbar configurations, the liquid cooling manifold (supply and return piping) runs parallel to the busbar along the rack row. The standard layout places the busbar at 2.8 to 3.0 meters height and the cooling manifold at 2.4 to 2.6 meters, leaving clearance for rack-top connections and maintenance access. This dual-overhead routing eliminates the raised-floor requirement that complicated earlier high-density designs but demands ceiling heights of 3.5 meters or more from slab to underside of structural members.

For facilities using immersion cooling (full tank immersion rather than direct-to-chip), the power delivery connections are made below the tank level through sealed penetrations. The electrical design must accommodate the thermal mass of the cooling fluid, which provides some inherent thermal buffering but also creates thermal inertia that affects power management strategies during load transitions.

Protection, Monitoring, and Power Quality

Megawatt-scale rack power delivery amplifies the consequences of electrical faults and power quality issues. A short circuit at a 120 kW rack can produce fault currents exceeding 10,000 amps, creating arc flash hazards and equipment damage risks that are orders of magnitude more severe than traditional IT environments.

Arc Flash Mitigation

Arc flash incident energy increases with available fault current and clearing time. At megawatt-scale power densities, arc flash analysis per NFPA 70E and IEEE 1584 is mandatory for every work location, including tap-off box positions, busbar maintenance points, and rack-level connection points. Typical mitigation strategies include zone-selective interlocking (ZSI) between upstream and downstream breakers to minimize clearing time, current-limiting fuses at tap-off points, and arc-resistant busbar enclosures that direct arc energy away from personnel.

Compliance requirements for arc flash labeling, personal protective equipment (PPE) categories, and approach boundaries become significantly more stringent at higher power densities. Facilities operating busbar systems above 2,000A typically require Category 3 or Category 4 arc flash PPE for any work within the arc flash boundary, which can extend several feet from the busbar enclosure.

Power Quality and Harmonics

GPU power supplies are switching-mode converters that draw non-sinusoidal current, generating harmonic distortion on the facility's electrical system. A single GPU server produces relatively modest harmonics, but a row of 20 high-density GPU racks drawing 2+ MW collectively can generate harmonic current levels that cause overheating in transformers, nuisance tripping of protection devices, and interference with monitoring equipment.

IEEE 519 (Standard for Harmonic Control) sets limits for harmonic current injection at the point of common coupling. Facilities must verify that their total harmonic distortion (THD) stays within IEEE 519 limits, which typically requires active harmonic filters, K-rated transformers designed for non-linear loads, or delta-wye transformer configurations that cancel third-order harmonics. Power factor correction is equally important: uncorrected power factor in GPU-heavy environments can drop below 0.85, triggering utility penalties and reducing effective transformer capacity.

Per-Rack Monitoring

At 120 kW per rack and higher, real-time power monitoring moves from a nice-to-have to an operational necessity. Each tap-off box or rack feed should include per-phase current and voltage measurement with at least 1-second granularity, feeding into the facility's DCIM system. This data serves multiple purposes: capacity planning (identifying headroom on each circuit), anomaly detection (detecting ground faults, phase imbalance, or unexpected load changes), energy accounting (metering per-rack consumption for tenant billing or internal chargeback), and predictive maintenance (trending conductor temperature and connection resistance over time).

Advanced monitoring systems also track power quality metrics at the rack level, including THD, power factor, crest factor, and inrush current during server power-on sequences. A single GB200 NVL72 rack can draw 2 to 3 times its steady-state current during initial power-on, and the distribution system must be designed to accommodate this inrush without tripping upstream protection.

Design Checklist for Megawatt-Scale Rack Facilities

Operators planning new construction or major retrofits for high-density AI racks should address these critical design parameters during the facility design phase, not after construction begins:

Design Parameter	Traditional (5-15 kW/rack)	High Density (50-120 kW/rack)	Megawatt (200 kW-1 MW/rack)
Distribution Method	Cable whips from panel	Overhead busbar	Direct transformer feed + busbar
Transformer Sizing	1 per 200-400 racks	1 per 20-40 racks	1 per 4-8 racks
Cooling Method	Air (CRAC/CRAH)	Direct-to-chip liquid	Direct-to-chip + immersion hybrid
Floor Loading (kg/m2)	1,000-1,500	2,000-3,000	3,000-5,000
Ceiling Height (m)	3.0-3.3	3.5-4.0	4.0-5.0
UPS Architecture	Centralized room-level	Distributed row-level	Rack-level or BESS bypass
Redundancy	2N UPS + generator	2N UPS + generator + STS	2N + dedicated transformer + STS
Arc Flash Category	Category 1-2	Category 2-3	Category 3-4

Implications for High-Density Colocation Customers

For organizations procuring colocation space for AI workloads rather than building their own facilities, the power delivery architecture has direct implications for contract negotiation, deployment timelines, and operational flexibility.

Verify provisioned vs available power. A facility may have 50 MW of total capacity but only 5 MW provisioned for high-density racks with appropriate busbar distribution. Ask specifically about the power delivery method for your target rack density, not just the facility's aggregate capacity.

Understand the power path. Request a single-line diagram showing the electrical path from utility to your rack, including transformer, ATS/STS, UPS, switchgear, and distribution method. Count the number of conversion steps and points of failure. Fewer conversions mean higher efficiency and lower failure probability.

Negotiate power density guarantees. Standard colocation SLAs specify uptime percentages but rarely guarantee power density. For GB200 NVL72 or similar deployments, the contract should explicitly guarantee sustained power delivery at the per-rack level (e.g., "120 kW continuous per rack position") rather than relying on aggregate facility capacity that might be shared with lower-density tenants.

Plan for growth. If your current deployment requires 120 kW/rack but your hardware roadmap includes NVL144 at 280 kW/rack, verify that the facility's busbar capacity and transformer headroom can accommodate the increase without a facility retrofit. Retrofitting busbar systems while adjacent racks are energized is complex, expensive, and often requires partial shutdowns that affect other tenants.

The Road to 1 MW Per Rack: What Comes Next

The trajectory of GPU power consumption shows no signs of slowing. NVIDIA's roadmap suggests that each successive GPU generation will maintain or increase per-chip TDP, while rack-scale integration packs more GPUs into each enclosure. The industry consensus is that 400 kW to 500 kW per rack will become common for training clusters by 2028, with individual enclosures approaching 1 MW by 2030.

This trajectory will force several architectural shifts in power delivery:

Medium voltage within the data hall will become standard rather than exotic. Distributing at 1 kV or higher reduces conductor copper mass by 4 to 10 times compared to 480V, which is not merely an economic consideration but a physical necessity when each rack draws 500+ amps at low voltage.

48V DC rack-level distribution (already adopted by Google's OCP designs) will expand as GPU power supply architectures evolve. Eliminating the AC-to-DC conversion at the rack PDU improves efficiency by 2 to 4 percent and removes the power factor and harmonic issues associated with hundreds of individual switching PSUs.

Battery energy storage systems (BESS) will increasingly replace or supplement centralized UPS for high-density rows, providing both ride-through capability and peak shaving functionality that reduces maximum demand charges from the utility.

The operators and colocation providers who invest in flexible, high-capacity power distribution infrastructure today will have a significant competitive advantage as rack densities continue to climb. Those who build for 30 kW per rack will find themselves either stranding capacity as customers' needs outgrow their infrastructure or facing expensive, disruptive retrofit projects. The time to design for megawatt-scale rack power is now.