Abstract
This paper examines a specific decision facing owners of artificial intelligence data centers: whether the K-rated isolation transformer can be removed from the power train as facilities move to 800 volt high-voltage direct current distribution, and if it can, who inside the owner’s organization is entitled to make that call. The question arrives on projects as a value-engineering proposal from an equipment supplier or an integrator, usually late, usually with a schedule and cost argument attached, and usually without a written account of what the transformer was doing in the chain before it was removed.
The analysis reframes the proposition. Removing the isolation transformer does not remove the work the transformer performed. It re-allocates six distinct functions, galvanic isolation, harmonic control, voltage transformation, fault current limiting, common-mode noise attenuation, and the establishment of a separately derived system for grounding reference, to other equipment in the chain: the active front end rectifier, its high-frequency transformer, the rack power supply unit, the protection scheme, and the grounding design. Each re-allocation is defensible on its own. The aggregate is defensible only when the owner can show, function by function, where each one now lives, what evidence demonstrates it works, and which party warrants it.
The paper sets that re-allocation against the Chinese 240 volt direct current precedent, where high-voltage direct current distribution with direct mains supply has operated at scale in Tencent, Alibaba, and Huawei FusionPower deployments for years. That precedent is real and instructive, and it does not transfer without translation. Bus fault energy, arc-flash incident energy, protection speed demand, and common-mode coupling all scale unfavourably from 240 volts to 800 volts, and the regulatory environment, the listing regime, and the supporting product ecosystem differ substantially between the two markets.
Compliance is treated as a live constraint rather than a closing formality. National Electrical Code Article 706 keys several requirements to a 100 volt direct current threshold, Article 712 contemplates direct current microgrids at voltages below the operating point of these systems, and the Underwriters Laboratories listing regime addresses converters, energy storage, and prefabricated modular data center assemblies without yet covering 800 volt direct current distribution as a whole. The May 2026 North American Electric Reliability Corporation Level 3 Essential Actions alert, issued after events in which more than one thousand megawatts of computational load left the bulk power system within seconds, moved several grid-side questions from advisory guidance into reporting obligation.
The framework the paper offers is owner-side and procedural. It sets out the decision rights that must be assigned before topology selection, the function re-allocation statement that belongs in the basis of design, the evidence each of the three bankability parties requires, the authority having jurisdiction, the insurer, and the lender or investor, the witness testing that converts vendor claims into owner evidence, and the contract positions an owner should hold on a first-of-kind topology. Recommendations are stated with a named decision owner and a trigger point for each.
Executive Summary
The 800 volt high-voltage direct current power train is the architecture that artificial intelligence data centers are converging on, and the convergence is far enough along that owners are now making irreversible commitments against it. Silicon vendors have published the architecture, the Open Compute Project has specified the rack and power interfaces, and equipment manufacturers have products in qualification. The decision in front of most owners is no longer whether to plan for direct current distribution. It is how much of the inherited alternating current protective apparatus to carry forward, and on whose authority.
Finding one. The isolation transformer question is being decided by default on live projects. On the projects where the K-rated unit has come out of the chain, the removal has usually been proposed by a supplier or an integrator as a cost and space saving, accepted by a project team under schedule pressure, and documented nowhere in a form that an insurer or a plan reviewer can evaluate. The engineering may well be sound. The absence of a written record of the decision is what creates the exposure.
Finding two. Six functions, not one device. The transformer provided galvanic isolation, harmonic control, voltage transformation, fault current limiting, common-mode attenuation, and a separately derived system reference for grounding. In the 800 volt architecture each of these can be re-allocated, and the strongest candidates are the active front end rectifier and its high-frequency transformer, the isolated rack power supply, the protection scheme, and a deliberate grounding design. Common-mode attenuation and grounding reference are the two functions that re-allocate least cleanly and deserve the most scrutiny.
Finding three. The Chinese 240 volt precedent supports the direction and does not settle the case. Direct current distribution with direct mains supply has run at scale for years and reports supply efficiency in the range of 94 to 95 percent. Moving the same topology to 800 volts increases bus fault energy, arc-flash incident energy, and the speed demanded of protection by roughly an order of magnitude in the author’s estimation, and the listing regime and product ecosystem that support the Chinese deployments do not exist in the same form in the United States.
Finding four. The code does not yet describe the architecture. Article 706 keys obligations to a 100 volt direct current threshold, Article 712 contemplates microgrid voltages well below the operating bus, and no Underwriters Laboratories listing covers 800 volt direct current distribution as an assembled system, although the first system-level certifications for data center equipment have begun to appear. This is a documented gap rather than a prohibition. It shifts cost and schedule onto the owner in the form of engineering evidence, test reports, and plan review cycles that a conventional alternating current line-up would not require.
Finding five. The grid side hardened in May 2026. The North American Electric Reliability Corporation issued a rare Level 3 Essential Actions alert after repeated events in which more than one thousand megawatts of computational load dropped off the bulk power system within seconds. Ride-through behaviour, load modelling, and commissioning practice for large computational loads now sit inside a reporting obligation directed at transmission owners, planners, and balancing authorities. Owners who treated interconnection as a procurement milestone will find it is a design constraint.
Recommendation one. Assign the decision rights before topology selection, not after. Name the individual who owns the isolation transformer decision, the individual who owns the protection coordination study that spans the skid boundary, and the individual who owns the composite grounding design. On the projects that go badly, these three owners are discovered during commissioning rather than assigned during design.
Recommendation two. Require a function re-allocation statement in the basis of design. One page, six rows, one per function: where the function now lives, what evidence demonstrates it, and which party warrants it. This single artifact resolves most of what the authority having jurisdiction, the insurer, and the lender will ask, and it converts a value-engineering conversation into a documented engineering position.
Recommendation three. Buy the evidence, not the assurance. Independent witness testing at factory acceptance and site acceptance, with hold points tied to payment milestones, is the mechanism that turns a vendor performance claim into something an owner can present to a third party. On a first-of-kind topology it is the cheapest risk retirement available.
Recommendation four. Phase the migration and keep the option open. Owners with a multi-block campus programme can carry the transformer on the first block, instrument it heavily, and remove it on later blocks once the evidence base exists. The cost of that optionality is small against the cost of discovering a coordination problem across an entire energised campus.
This paper is written for the executive whose campus is in the commitment phase, the engineering organisation that must produce the evidence, the capital sponsor underwriting the programme, and the operating team that will hold the result for fifteen years.
Full White Paper Below
