The Power Crunch Is Moving Inside the Data Center

Securing 100 megawatts of utility service is only the beginning of building a 100-megawatt AI data center. That electricity still has to pass through transformation, protection, backup systems, distribution equipment and power converters before it reaches the accelerators. Each stage occupies space, generates heat and imposes operating requirements.

Article Summary

0:00

/58.23093

1×

The public debate has concentrated on getting electricity to the property: generation, transmission, substations and interconnection. Those constraints remain. A second challenge is becoming harder to ignore behind the meter, where increasing rack density is forcing a redesign of the electrical system that turns an energized building into productive computing infrastructure.

Google made the scale of that change explicit in April 2025, presenting ±400 VDC power delivery intended to support racks up to 1 MW. NVIDIA’s August 2026 roadmap extends the transition from dedicated power racks in existing AC facilities to row-level and eventually facility-level DC systems. These are design targets and deployment plans, not evidence that megawatt racks have become the industry norm. ^1,2

The strategic question now extends beyond how much power a site can secure. Developers and operators also need to establish how much can reach the intended equipment, at the required density and reliability, within the project’s commissioning schedule.

Consider a representative AC power path. Utility service feeds transformers and switchgear, then an uninterruptible power supply (UPS) and AC distribution equipment. Rack power supplies convert AC to roughly 48–54 VDC; additional DC-to-DC converters reduce voltage for the processors. Batteries connect through the UPS or other power electronics, while cooling and auxiliary equipment draw from separate branches. Exact arrangements vary, and a battery energy storage system (BESS) is not necessarily a separate conversion stage in series with every IT load. ³

This architecture served earlier computing generations well. Increasing the power concentrated in each rack changes its physical demands. Conductors, connectors and power shelves must carry more current, while electrical equipment competes with computing hardware and cooling systems for space. Moving conversion into an adjacent power rack can free room inside the compute rack, although the power equipment still occupies floor space elsewhere.

The electrical arithmetic explains the pressure to raise distribution voltage. Delivering 1 MW over a 400 V DC circuit requires 2,500 amperes before losses; at 800 VDC it requires 1,250 amperes. With the same conductor resistance, halving current reduces resistive heating by 75%. That is a comparison of conductor losses at equal power and resistance, not a prediction of total facility savings. Actual copper requirements also depend on distance, temperature limits, voltage drop, redundancy and protection.

Voltage and conversion topology address different problems. Higher voltage reduces current for a given DC power transfer. Changing where AC becomes DC may remove or consolidate conversion stages. Neither change eliminates the low-voltage regulation needed near the chip, and the efficiency of the complete chain must be measured under realistic loading.

Class 4 fault-managed power belongs in this discussion because it changes how certain circuits can deliver and protect power. The 2023 National Electrical Code introduced Class 4 systems in Article 726. Their defining feature is active monitoring and limitation of energy delivered into a fault. The Class 4 framework extends to 450 V peak; UL 1400-2 specifies that voltage rating for Class 4 cables. ^4,5

Commercial evidence is concrete. UL Solutions announced its first Class 4 system certification, for Panduit, in November 2023. Cisco’s February 2025 implementation guide validates Panduit FMP with Catalyst 9300 Power over Ethernet switches, including primary, supplementary and backup-power configurations. NECA has also scheduled a September 2026 webinar comparing FMP and traditional AC distribution in high-performance data centers. ^6,7,8

These developments establish a credible product category and growing industry attention. The Cisco validation specifically concerns network power delivery. It does not establish FMP as a tested replacement for the main electrical architecture of megawatt AI racks. Distributed networking, controls, building systems and selected edge applications are more directly aligned with the demonstrated use case.

The distinction matters because Class 4 is a protection and installation category, while 800 VDC is a distribution voltage. A conventional 400 VDC circuit becomes Class 4 only through the required fault-management design and listing. Likewise, an 800 VDC system does not qualify as Class 4 simply because it has fast electronic protection. A bipolar arrangement can place each pole 400 volts from its midpoint while maintaining 800 volts between poles; that arithmetic does not itself establish Class 4 compliance.

Efficiency claims require equal care. Berkeley Lab reported in 2006 that DC data-center architectures could reduce overall electricity use by roughly 10–20%, with preliminary demonstration measurements supporting that estimate. The comparison involved the equipment and conversion chains of that period, including 480-to-208 VAC distribution and server power supplies. It was a DC-architecture demonstration, not a trial of the Class 4 systems introduced much later. ⁹

That historical result cannot be carried forward as a demonstrated FMP saving for modern AI facilities. Cisco’s documented FMP system itself includes AC-to-DC conversion at the supply and conversion at the receiver. Whether a proposed arrangement saves energy depends on the stages it replaces, conductor losses, equipment efficiency and the load profile. ⁷

A more recent, narrower reference comes from Google, which reported approximately 3% improvement in end-to-end efficiency for its AC-to-DC sidecar approach. That is a company-reported result for a particular architecture. It should not be restated as a universal facility-energy saving or silently converted into a three-percentage-point improvement. The starting design and measurement boundary matter. ¹

The emerging transition is therefore likely to be staged. NVIDIA said in August that an MGX-compatible 800 VDC power rack was planned for the second half of 2026, followed by a row power center supporting up to 2 MW per row with availability expected in 2027. A row rating is an aggregate design capacity, distinct from the rating of an individual compute rack. ²

OCP describes two practical routes: side power racks that convert existing 480 VAC to ±400 or 0–800 VDC near the load, and future systems that convert medium-voltage AC into 800 VDC for a data hall. Its August 2026 update distinguishes the published Mt Diablo ±400 V sidecar specification from a later Mt Diablo 2.0 intended for native 800 V. Common direction is emerging, while interfaces and deployment maturity continue to evolve. ¹⁰

For an existing facility, the sidecar route can preserve upstream investment where sufficient AC capacity and row space are available. Cooling, structural loading, protection and maintenance access still require project-specific assessment. For a new campus, earlier decisions about conversion, storage and DC distribution may support a more integrated design. The technologies below occupy different positions in that system and can coexist. ¹⁰

Where the power architectures fit

These are complementary layers and design choices, not mutually exclusive whole-facility alternatives. Status reflects public evidence available September 7, 2026.

Source: AIxEnergy synthesis of the linked primary sources. No generic efficiency advantage is assigned to an architecture; system boundaries and operating conditions determine results.

Solid-state transformers extend the redesign toward the utility interface. Using power electronics and higher-frequency magnetic components, they can combine controlled medium-voltage conversion and DC output in a compact system. OCP’s 2026 specification, developed by Google, Microsoft and NVIDIA, defines a framework for this equipment. It describes interfaces and performance requirements; publication of a specification does not establish fleet operating reliability. ¹²

An 800 VDC facility does not require a solid-state transformer. Conventional transformer-rectifier systems offer another conversion route. The procurement choice should turn on efficiency across the operating range, installed footprint, protection, replacement time, spares and service capability. Consolidating equipment can reduce some failure points while increasing the consequence of a failure in a shared unit.

DC distribution also changes the integration options for batteries, fuel cells and solar. These sources provide DC internally, potentially avoiding an inverter followed by a rectifier when connected to a compatible DC system. Their voltages and operating characteristics still have to be reconciled. DC-to-DC conversion, isolation, protection and coordinated controls may remain necessary. Electrical compatibility says little by itself about fuel availability, emissions, storage duration or firm capacity. ¹³

Storage has a second role beyond outage backup. NVIDIA’s technical work identifies rapid, synchronized AI load changes and describes capacitors and batteries operating across different timescales to buffer them. Higher distribution voltage alone does not smooth the load seen by the utility. ³

A system engineered to absorb short power swings is also not automatically capable of reducing demand for several hours. That requires sufficient stored energy, discharge capability, controls and an operating agreement that leaves adequate reserve for the facility’s reliability needs. Utilities should evaluate the resulting behavior at the point of connection, including ramps and disturbance response, rather than infer flexibility from the presence of a DC bus.

The economic comparison should follow electricity all the way to the relevant computing boundary. Power usage effectiveness (PUE) divides total facility energy by IT-equipment energy. It captures facility overhead, including cooling and upstream electrical losses, but it does not separate accelerator consumption from other IT loads or establish useful computing output. ¹⁴

A grid-to-compute accounting approach can extend that view. For a grid-only facility, define an accelerator delivery fraction as energy entering the accelerator modules divided by energy entering at the utility meter, measured over the same period. This is a descriptive accounting measure, not a new industry standard. The boundary should be explicit: module-input energy still includes any conversion losses and supporting components inside the module; it is not a measurement of energy reaching the processor silicon alone.

The accounting should separately identify cooling and auxiliary demand; transformer, UPS and other upstream conversion losses; distribution losses; IT power-conversion losses outside the accelerator modules; and other IT consumption. CPUs, networking and storage perform necessary work and should not be classified as electrical waste. Backup losses should reflect standby and charging behavior, plus actual discharge cycles, rather than treating battery round-trip losses as if they applied continuously to every utility kilowatt-hour.

A simple illustration shows why the distinction is useful. Assume a grid-only facility draws a steady 100 MW, operates at PUE 1.20, and delivers 80% of its IT input power to accelerator-module inputs. IT power is then 100 ÷ 1.20, or 83.3 MW; accelerator-module power is about 66.7 MW. If an alternative design operates at PUE 1.15 with the same 80% allocation, the figures become 87.0 MW and 69.6 MW. The roughly 2.9 MW difference is additional power available at the defined module boundary under those assumptions.

Those values are illustrative assumptions and arithmetic, not measured AC, FMP or 800 VDC performance. They do not establish an equivalent increase in useful compute. That requires a controlled workload comparison with the same hardware, performance target, utilization and reliability conditions. Annual assessments should use energy over matching periods, and facilities with onsite generation must also count that supply and changes in stored energy. Otherwise, reducing grid imports can make the ratio improve without improving the electrical system.

For a power-constrained developer, this shifts attention from nominal service capacity to the combination of usable IT power, achievable rack density and time to operation. A higher-efficiency architecture may make more computing equipment supportable within an existing connection. A denser architecture may let that equipment fit in a viable building layout. A technically superior design that delays commissioning can still be the wrong commercial choice.

The comparison therefore belongs in project underwriting and procurement. Bidders should show the complete electrical path, efficiency at expected loading, achievable output with the specified redundancy, commissioning dependencies and maintenance strategy. One party should own the integrated performance guarantee across conversion, distribution, storage and rack interfaces. Individual component certifications cannot establish the availability of the assembled system.

The incentives also deserve scrutiny. If an operator can fill every released megawatt with additional accelerators, efficiency may raise compute output while leaving utility demand unchanged. The benefit is then more output per constrained connection, rather than a reduction in site demand. Equipment providers, utilities and investors need to know which outcome the project actually intends to deliver.

For site intelligence, these technologies are best represented as documented configuration assumptions attached to a project: equipment choices, useful power at defined boundaries, density limits and upgrade requirements. They should inform usable-power and site-economics calculations when engineering evidence supports them. An automatic efficiency uplift assigned to every site would imply precision the evidence does not justify.

The next competitive advantage in AI infrastructure may emerge in the engineering between the meter and the accelerator: selecting a power architecture that can be procured, installed, maintained and expanded on schedule. Secured megawatts establish the opportunity. The electrical system determines how much of that opportunity can become operating compute.

Sources and measurement notes

Source links identify standards, technical documentation and attributed company statements. Roadmap dates are prospective. The 100 MW comparison is an AIxEnergy illustration using stated assumptions,

1.    Google Cloud, “AI infrastructure is hot. New power distribution and liquid cooling infrastructure can help.” April 30, 2025.

2.    NVIDIA, “Why Scaling AI Compute Performance Requires a New Power Architecture.” August 11, 2026.

3.    NVIDIA, “Building the 800 VDC Ecosystem for Efficient, Scalable AI Factories.” October 13, 2025.

4.    NECA, “What Is Class 4 Fault Managed Power?” June 26, 2023.

5.    UL Standards & Engagement, UL 1400-2, “Cables in Fault-Managed Power Systems,” scope §1.2. September 19, 2025.

6.    UL Solutions, first Class 4 system certification for Panduit. November 21, 2023.

7.    Cisco, “Panduit FMPS and Cisco Implementation Guide.” February 7, 2025.

8.    NECA, FMP and traditional AC distribution in high-performance data centers; webinar scheduled for September 17, 2026.

9.    Berkeley Lab, DC data-center demonstration and preliminary energy-savings estimates. June 21, 2006.

10.  Open Compute Project, Google, Microsoft and NVIDIA collaboration on LVDC power architecture. August 11, 2026.

11.  ETSI EN 300 132-3 V2.3.1, power-supply interface for ICT equipment, up to 400 VDC. January 2023.

12.  Open Compute Project, “Solid State Transformer Specification,” revision 0.3.0, effective June 22, 2026; especially §§3–5.

13.  Open Compute Project, “Data Center Facility – Low Voltage Direct Current Power Distribution,” version 1.0. March 30, 2026; protection, storage and DC integration sections.

14.  Vertiv, “What Is PUE (Power Usage Effectiveness) and What Does It Measure?” Definition and measurement boundaries.