If you are buying or building colocation for AI and HPC workloads, the power architecture conversation eventually comes down to 415V three-phase versus the traditional 208V. Most existing colocation facilities are built around 208V. Most new high-density builds — and most retrofits chasing the AI workload — are moving to 415V. This article walks the actual electrical engineering, not the marketing.
The short version
At low rack densities — under about 15kW per rack — 208V is fine. The cost of the electrical infrastructure works out, conversion losses are tolerable, and the entire industry has been built around it.
At high rack densities — 50kW+ per rack, where AI training and inference clusters now routinely operate — 208V starts to lose. The current required gets large, the copper required to deliver that current gets expensive, the conversion losses become a real PUE penalty, and the rack-level distribution gets messy. 415V is the cleaner answer at this density.
In between — roughly 30 to 40kW per rack — it’s a real engineering trade-off. Above that, the math tilts hard to 415V.
How 415V and 208V actually work
Both voltages are three-phase AC systems with a neutral. The difference is in the voltage between hot legs and the voltage from leg to neutral:
- 208V system — 208 volts between any two of the three hot legs, 120V from each leg to neutral. Standard North American commercial power
- 415V system — 415 volts between any two of the three hot legs, 240V from each leg to neutral. Common in Europe, increasingly common in modern data centers in North America
Server power supplies (PSUs) in the AI/HPC class accept a wide voltage range — typically 100V to 240V at the PSU input. A modern GPU server PSU runs equally well off 208V (one leg-to-leg from a 208V system) or 240V (one leg-to-neutral from a 415V system). The server doesn’t care. The electrical infrastructure upstream of the server cares a lot.
Current is the enemy
The fundamental physics: power = voltage × current. To deliver 50kW to a rack, you can use more voltage and less current, or less voltage and more current. The math:
- At 208V, a 50kW rack draws roughly 139 amps per phase (three-phase, balanced)
- At 415V (240V leg-to-neutral), the same 50kW rack draws roughly 70 amps per phase
Half the current. And current is what drives everything expensive downstream — bigger conductors, bigger breakers, more copper, more cable trays, more heat in the cables, and tighter constraints on rack-level cable routing.
Conversion losses and PUE
Inside a data center, the path from utility supply to the server PSU usually involves multiple voltage transformations. At each transformation, you lose a small percentage to heat. The fewer transformations, the better.
Traditional 208V architecture in many older facilities runs roughly: utility → transformer to ~480V → transformer to 208V → PDU → rack. Each transformation costs ~1–3% efficiency depending on equipment.
Modern 415V architecture can run: utility → transformer to 415V (delta-wye providing 240V leg-to-neutral) → PDU → rack. One fewer transformation step, less heat dissipated as loss, lower PUE (Power Usage Effectiveness) impact.
At facility scale — 10, 50, 100 racks running 50kW each — a 1–2% PUE improvement compounds into hundreds of thousands or millions of dollars per year of power cost reduction. The transformer pad and the upstream electrical may cost more to install, but the lifetime efficiency advantage is real.
Copper cost — the part that surprises people
Doubling the current requires roughly square-root-of-two more copper at constant voltage drop (the actual math depends on conductor sizing rules and ampacity tables, but the principle holds). At rack-level scale, that’s tolerable. At facility scale, copper is one of the biggest line items in the electrical buildout.
For a 50-rack high-density deployment:
- 208V — copper feeders sized for ~139A per rack × 50 racks. Significant copper, large conduit, large trays
- 415V — copper feeders sized for ~70A per rack × 50 racks. Substantially less copper, smaller conduit, smaller trays
Copper prices fluctuate, but the directional cost difference is meaningful — often 15–25% less copper required in the rack-level distribution for the 415V architecture at high densities. Combined with the smaller conductor sizes in the rack PDUs themselves, the buildout cost gap closes the headline price difference of going to 415V transformers and switchgear.
Why most colocation hasn’t transitioned
If 415V is so clearly better at high density, why isn’t it everywhere? Three reasons:
1. Existing facilities are built for 208V. A retrofit isn’t trivial. The transformers, the switchgear, the bus, the PDUs — many of these are sized and specified for 208V architecture. Replacing them is expensive and disruptive to existing tenants.
2. Most tenants didn’t need it. For years, the average colocation rack ran 5–15kW. At that density, 208V is fine, sometimes preferable, and tenant demand for 415V was minimal. Operators built what tenants asked for.
3. The industry inertia is real. Trained electricians, established design patterns, off-the-shelf PDU products, internal SOPs — all built around 208V. Changing to 415V requires retraining and re-tooling.
The shift is happening now because AI workloads broke the assumption. A 50kW+ rack is not a niche request anymore. It’s the baseline for modern GPU clusters. And operators building net-new capacity for those workloads are choosing 415V because the math obviously wins at that density.
Server PSU compatibility
A real concern customers raise: “Does this break my servers?”
No, for modern equipment. AI- and HPC-class server PSUs are designed for global deployment and accept 100–240V at the input. Whether you feed them 208V or 240V, they run equally well. Some PSUs are technically more efficient at the higher voltage — typically a fraction of a percent.
Maybe, for older equipment. Some legacy gear is 120V-only or has narrow input ranges. If you’re standing up a mixed environment with legacy servers, confirm input voltage compatibility before committing the rack to 415V/240V distribution.
For a greenfield AI deployment with current-generation servers, PSU compatibility is not the bottleneck. The bottleneck is the upstream electrical infrastructure of the facility.
The TCO math at high density
Pull the numbers together for a 50-rack, 50kW-per-rack deployment:
208V scenario:
- Higher copper cost in distribution
- Higher current means larger conductors, breakers, busway
- Additional voltage transformation upstream means 1–2% higher PUE
- Lower CapEx for switchgear (often) but higher operating power loss
- Acceptable at this density, but increasingly painful
415V scenario:
- Lower copper cost in distribution
- Smaller conductors, breakers, busway
- One fewer voltage transformation step, lower PUE
- Higher CapEx for some upstream electrical components
- Cleaner architecture, more headroom for future density growth
At 10 racks, the TCO math is roughly a wash. At 50 racks, 415V starts to clearly win on a multi-year basis. At 100+ racks, 415V is hard to argue against.
Where Data Suites fits
Data Suites’ Middle Tennessee facility is architected for the modern AI workload — 50kW+ per rack, 415V power distribution, modular suite expansion that doesn’t force you to overbuy on day one. The infrastructure was designed assuming high-density compute is the default, not the exception, so the electrical architecture is built around the math that actually serves AI and HPC workloads rather than being retrofitted from a legacy 5kW-per-rack design.
For buyers running GPU clusters or other density-driven workloads, the 415V architecture isn’t a marketing line — it’s a measurable advantage in operating cost and PUE at the density you’re actually running.
Bottom Line
For low-density colocation, 208V is fine and probably what you have. For modern AI and HPC workloads at 50kW+ per rack, 415V wins on copper cost, on PUE, on rack-level distribution complexity, and on headroom for the next density jump. The industry shift is real, the math is well-understood, and the buyers asking colocation operators about 415V availability are the buyers who have run the numbers.
If your workload is dense — or about to be — the power architecture conversation is one of the highest-leverage technical decisions in your colocation contract. Don’t let it be an afterthought.
Data Suites is a Murfreesboro-based Tier 2 colocation operator serving AI, HPC, and high-density compute workloads across Middle Tennessee.
