The ‘Is AC or DC power better suited for tomorrow’s high-density needs?’ panel at Datacloud Global Congress 2026 produced a more nuanced verdict than the session title implies. The transition to 800V DC power distribution is real and, for hyperscale AI applications, ultimately unavoidable-but panellists were pointedly sceptical about whether the industry is ready to execute it safely, at scale, and without a fragmented standards landscape that creates more problems than it solves.

The physics of high-density AI computing are driving a structural rethink of power delivery inside data centres. As GPU and networking requirements grow and rack power densities rise beyond what traditional alternating current (AC) distribution systems can efficiently handle, the case for higher-voltage direct current (DC) architectures has become technically compelling.

Moving to higher voltage DC allows for reduced conductor size-less copper-and improved energy efficiency, which matters both economically and in the context of increasingly stringent sustainability requirements.

The Data Centre Power Revolution session at DCGC 2026 identified 2028 as a likely milestone for broader implementation of these architectures. But the 800V transition panel introduced a note of caution that is worth taking seriously: technical direction and delivery readiness are not the same thing.

Why 800V DC is inevitable for hyperscale AI

The technical drivers are clear. At rack densities above 100 kilowatts-which the NVIDIA GB300 at 141kW has already made real, with future projections reaching 600kW-traditional AC distribution creates challenges around cable size, heat generation, and the sheer volume of copper required.

Moving to 800V DC addresses these constraints by reducing current for a given power level, which proportionally reduces conductor size and resistive losses. The efficiency gains, while incremental at the individual facility level, compound significantly across the portfolio of a large hyperscaler.

Successful deployments in Asia were cited as evidence that the architecture can work at scale, and chip manufacturers and hyperscalers were identified as the primary drivers of this transition-not equipment vendors or standards bodies.

Where the scepticism lies: safety, standards, and supply chain

The panel’s caution centred on three interconnected concerns. First, safety: high-voltage DC systems present arc flash and fault management challenges that are significantly different from AC systems, and the panel called for new standards and rigorous testing protocols before widespread deployment.

The training implications for operations and maintenance personnel are substantial-this is not a system that can be handed over to a workforce trained on conventional AC infrastructure. Second, standards: the absence of established, harmonised standards for 800V DC systems creates a risk of fragmented, incompatible implementations that would disadvantage the entire market.

The panel was explicit that chip manufacturers, hyperscalers, vendors, and standards bodies need to coordinate now to avoid this outcome. Third, supply chain maturity: the panel expressed genuine scepticism about the supply chain’s current ability to deliver reliable, safe products at the required scale.

The role of batteries and UPS in an 800V world

One underappreciated dimension of the 800V transition, raised in the BESS panel, is the changing role of uninterruptible power supply (UPS) systems. In conventional data centres, UPS provides backup power during grid interruptions. In AI-intensive facilities, the power demand profile is highly dynamic-large language model inference generates significant load swings that can stress both facility power systems and grid stability.

The BESS panel heard that without adequate damping of these fluctuations, AI-scale data centres could cause grid-level oscillations, and that some countries are already imposing restrictions on large-scale AI deployments on precisely these grounds. In an 800V DC architecture, the UPS must handle not just backup functions but active load management and grid support-a fundamentally different operating requirement that demands both different equipment and different operational expertise.

Key Takeaways

• The transition to 800V DC power architecture is technically inevitable for hyperscale AI applications, but 2028 is the most credible near-term milestone for broader implementation-operators making design decisions now should plan for both.
• Safety is the most underappreciated risk in the 800V transition: arc flash protection, fault management standards, and workforce training must be resolved before widespread deployment, not after.
• The absence of harmonised standards for high-voltage DC systems creates a fragmentation risk that would disadvantage the entire market-coordination between chip manufacturers, hyperscalers, and standards bodies is an urgent requirement.
• UPS and BESS systems in AI-scale facilities must manage active load swings, not just provide backup power-a requirements shift that demands new equipment specifications and operational capabilities.
• The 800V transition is unlikely to be universal across all data centre segments in the near term; traditional cloud facilities operating below 100kW per rack face a different cost-benefit calculation from AI factory operators.

The 800V DC transition represents a genuine engineering inflection point for the data centre industry, comparable in significance to the shift from diesel-only UPS to lithium-ion battery systems. The operators who will execute it well are those who treat the safety, standards, and workforce dimensions with the same rigour as the efficiency argument. The efficiency case is compelling. The readiness case, as DCGC 2026 made clear, is not yet made.