AI workloads are pushing rack densities from 10kW to 100kW to 300kW to 1MW — often faster than a mid-project retrofit can keep up. The busway, the breakers, the cooling, and the codes all hit limits at different points, and the standards were written when densities moved in single-digit increments.

Retrofitting a data center for AI density is rarely a story about the grid. The utility feed is the one thing most operators cannot change on AI timelines, because interconnection studies and high-voltage transformers run in years, not quarters. So, the realistic retrofit holds the service entrance fixed and deals with the harder problem: how do you force the power you already have through distribution gear that was sized for a much lower density load? Almost every binding constraint sits downstream of the utility meter, cast into the building a decade before anyone planned for higher than 100kW racks.

Same megawatts, denser racks

Density growth hasn’t been incremental; it looks nothing like it did five years ago. Data center infrastructure was built for 15 to 30kW per rack. Nvidia’s GB200 NVL72 now runs near 120kW, Vera Rubin NVL144 arrives in 2026, and the Rubin Ultra “Kyber” rack is projected near 600kW in 2027. But a retrofit usually is not stacking that load on top of what already exists. It is repacking the same contracted megawatts into far fewer, far denser racks. Take a single 250kW power cabinet: it once fed eight 30kW racks, but now has room for only two at 100kW, because a third would exceed its rating — same cabinet, a quarter of the racks, and 50kW left stranded.

That creates a problem operators recognize immediately: the building is often power-rich and distribution-poor. Contracted utility capacity frequently goes unused because the floor hit per-rack, per-busway, or cooling limits long before it ran out of megawatts. A high-density retrofit tries to reclaim that stranded capacity; to deliver power the facility already pays for to the handful of positions that now need it. But the path to those positions was never built to carry concentrated load.

The bottleneck moves downstream

Start at the rack and work up. A 100kW rack at 415V three-phase draws about 139 amps; the same rack on the 208V common in older US data halls draws roughly 278 amps. A legacy branch circuit at 30 amps, 208V three-phase, delivers under 9kW once the continuous-load derate is applied. A single 100kW position therefore needs the equivalent of twelve or more legacy circuits, or a wholesale move to higher-amperage, higher-voltage feeds.

That demand compounds at every level above the rack. Rack power distribution units (PDUs), and the floor PDUs or remote power panels feeding them, were rated for zones of light cabinets. Overhead busway sized at 800 or 1,200 amps fed dozens of 10kW racks; at 100kW per rack it is spent after a handful. Feeder breakers and PDU input breakers must grow with the load, and under the NEC continuous-load rule a 139-amp continuous draw wants a 175-amp breaker with conductors sized to match, not the nominal figure. The existing switchboards often have neither the bus rating nor the physical space to accept the larger frames, which turns a breaker change into a switchboard replacement.

You cannot just add copper

The instinct is to pull more and bigger conductors. You hit physics and code limits fast. Bundling more current-carrying conductors in a raceway triggers ampacity adjustment factors, so the tenth conductor in a conduit carries far less than the first. Existing conduits are typically near their fill limits already, so larger conductors mean new conduit and new pathways, run through a raised floor or cable tray that is already congested with the liquid-cooling manifolds these racks also require. Longer runs at higher current push voltage drop past acceptable limits, forcing yet larger copper. Past a certain point, this stops being a retrofit and starts being demolition inside an occupied building.

Voltage is the practical answer. Moving distribution from 208V to 415V roughly doubles the power delivered per amp; moving to 800V DC at the rack does far more, which is why the industry is converging on it, borrowed from EV charging, enabled by gallium-nitride and silicon-carbide electronics, and now the subject of IEEE Project P3710.1, the high-voltage-DC sub-project of IEEE’s modular-data-center effort, which addresses HVDC distribution from 300 to 1,500 volts for data centers. But raising rack voltage is its own retrofit: new PDUs, and power shelf units (PSU) input ratings, new protective-device ratings, and a workforce trained on 480V AC rather than 800V DC. It also means the topology you are converting toward is still being standardized while you build.

Cooling becomes an electrical problem

Power is only half of the density problem; the heat is the other half, and it feeds directly back into the electrical system. Racks at 15 to 30kW were air-cooled, and air handling units (AHUs) pushing cold air through a raised floor or a contained aisle were enough. Air runs out of headroom well before 100kW per rack; there is no practical way to move enough of it to carry that heat off a cold plate. Direct liquid cooling becomes mandatory. The retrofit problem is getting there without rebuilding the central plant on day one.

The simplest first step is liquid-to-air. An in-row or in-rack coolant distribution unit (CDU) pulls heat off the chips with a liquid loop and rejects it into the room, where the existing air handlers carry it away. It reuses the cooling plant already in place, but it is a stopgap: it is less efficient than liquid-to-liquid, and it works only while the room air still has capacity to absorb the rejected heat. Pushing further means liquid-to-liquid, a CDU that hands the technology loop’s heat to a facility water loop and on to the chillers or towers. That is no longer a drop-in. It means new facility piping, new CDUs, and often new chillers, dry coolers, and pumps, and every one of those is an electrical load. Variable-frequency pump drives, chiller compressors, and tower fans need their own power panels and feeder breakers, added to switchgear that was sized for IT and few AHUs only. In a fixed-utility retrofit, the cooling system does not only compete with the racks for space and water; it competes with them for megawatts and for room in the same lineups, which sends the engineer back to the switchgear, breaker, and coordination problems the power retrofit began with.

The constraints that remain

A fixed utility connection does not mean the internal distribution has room. Concentrating load unbalances the internal distribution. A step-down transformer that comfortably carried a zone of light racks can be driven past its rating when that zone goes high-density, even though total facility load is unchanged. The dominant stressor is less the classic harmonic content, which active power-factor-corrected supplies largely broken, than sustained high loading and the violent, sub-second load swings of training clusters that move 70 to 80 percent from idle to peak. Legacy UPS units were never sized and designed to handle these AI load swings; they get replaced whole, and the battery system behind them often must be reconfigured with them.

Often all of it happens in a building that cannot go dark. Revenue racks run a few feet from the work. The hard engineering is in the intermediate states, not the target design, and the proof that each transition holds the availability tier the contract promised. In the retrofits I have seen, the cutover window, when old protection comes out and new protection picks up the load, is where you lose availability. One miscalculated transfer, one spike during the swap, and you spend years buying back the SLA.

The code lags the hardware

Every one of these changes trips a standards problem. Under NFPA 70E, swapping a transformer, resizing a breaker, or altering a protective setting invalidates the existing arc-flash study. The 2026 NEC widens the exposure further: Section 110.16, now titled “Arc-Flash Hazard Marking,” drops the amperage qualifier, so switchboards, switchgear, panelboards, and motor control centers that were previously exempt must now carry arc-flash labels. The same edition reorganizes the over-1,000-volt rules, adding a new Article 270 for medium-voltage grounding and bonding. For a phased retrofit, the arc-flash study is not a deliverable but a recurring obligation at every phase boundary, with single-line diagrams that must track each interim configuration. IEEE authorized the parent project, P3710, in 2025 because existing frameworks were written for fixed, permanent installations, not for systems that are reconfigured as density blocks change.

That’s the real gap. We have codes for components and new architectures being defined in real time, but nothing that governs what operators do: reconfigure a live distribution system to concentrate fixed power into high-density racks, in phases, without going dark. Tier definitions do not describe AI’s transient load behavior. Power-quality limits predate this load profile. No one has published a recognized methodology for this.

Standards committees won’t write this first. Practitioners will—by publishing what actually happened, the sequencing, the derates that bit, the intermediate states nobody planned for, and from standards bodies treating phased live conversion as a first-class scenario rather than an edge case. The cheapest path is overbuilding upfront: distribution voltage chosen for where density is heading, conductor and conduit pathways sized with headroom, breaker frames that can grow. Reopening finished work always costs more than the original overbuild would have. The racks won’t wait. The infrastructure has to be rebuilt while it’s still running.