
800V DC Conversion and CapEx Planning: How AI Data Centers Turn Cooling Overhead Into Compute
800 VDC cuts copper 45% and conversion losses, but the CapEx win depends on cooling. See how DRL cooling optimization defers chillers and frees megawatts for GPUs.
The next generation of AI racks will draw up to 1 MW each, and the industry has settled on 800 VDC as the way to feed them. NVIDIA plans 800 VDC power infrastructure for 1 MW IT racks starting in 2027, with Schneider Electric, Vertiv, Eaton, Infineon and more than a dozen other partners shipping reference designs ahead of it.
The case for 800V DC conversion is physics: higher voltage, lower current, less copper, fewer conversion stages. But every megawatt delivered to a rack at 800 VDC comes out the other side as a megawatt of heat. That is the part of the CapEx plan most 800 VDC roadmaps skip.
This post covers both halves. First, the CapEx benefits of 800 VDC conversion and the numbers behind them. Second, how Synapse Thermal's autonomous cooling optimization keeps the cooling plant from eating the savings, so that power moves from overhead to compute.
The CapEx planning problem: AI data centers are capped by power, not floor space
A greenfield AI data center now costs about $17.6 million per MW before a single GPU is installed, up 21% in under two years, according to Cushman & Wakefield's 2026 Development Cost Guide. JLL's 2026 outlook puts shell and core alone at $11.3 million per MW, and the GPUs add another $29 to $42 million per MW on top (figures compiled by GR1D, Oct 2026).
Power infrastructure is the single largest line in that stack at 21%, with cooling at 12% and climbing: Turner & Townsend puts the mechanical share of a liquid-cooled AI build at 33%, against 22% for an air-cooled hall. The inputs are inflating faster than the totals. Since December 2021, switchgear is up 60%, copper 46%, and cooling equipment 32%, and pad-mount transformers now carry 68 to 113 week lead times.
The constraint that actually caps revenue, though, is the utility interconnect. Gross draw equals IT load multiplied by PUE, so at a fixed feed every point of PUE is compute you cannot energize. On a 75 MW interconnect, a PUE of 1.25 supports 60 MW of IT load; a PUE of 1.15 supports 65.2 MW. That 5.2 MW gap is roughly $150 to $220 million of GPUs that either run or sit in crates, and at the $1.9 million per MW-year implied by recent hyperscale leases it is about $10 million a year in rent (derived from the sources above).
Three facts make this worse for AI halls than for cloud halls:
- Cooling consumes roughly 40% of total data center energy, and the global average PUE still sits near 1.56, far from the 1.03 that optimized plants can approach.
- Legacy power chains lose energy at every AC/DC and DC/DC hop, and all of that loss becomes heat the cooling plant must reject before any GPU runs.
- GPU workloads generate 3 to 5 times the heat of CPU racks and swing it in seconds, so static cooling setpoints are tuned to the worst case and waste capacity the rest of the time.
This is why CapEx planning for AI data centers has become a power-chain and cooling-plant exercise rather than a square-footage one. 800V DC conversion attacks the first two problems. Autonomous cooling optimization attacks the third, and decides whether the first two actually show up in the budget.
What 800V DC conversion actually changes in the power chain
800 VDC conversion moves the big AC-to-DC step from inside the rack to the edge of the building. Today's GB200 and GB300 NVL72 racks take 415 VAC from the hall, rectify it to 54 VDC in up to eight in-rack power shelves, then step it down again to 12 V and core voltage for the GPUs. Scale that to a 1 MW rack and the 54 V busbar alone needs up to 200 kg of copper and the power shelves would fill 64 U, leaving no room for compute, per NVIDIA's 800 VDC architecture post.
In the 800 VDC design, industrial rectifiers (or solid-state transformers) convert 13.8 kV grid power to 800 VDC at the data center perimeter. That DC rides busways through the hall and the row, enters the rack on two conductors, and is converted once more to the voltages GPUs use. Battery storage on the DC bus absorbs the sub-second power swings that AI training produces. The intermediate AC/DC and DC/DC hops, and the fan-cooled PSUs that did them, are gone.
The left chain converts power four times between the utility and the GPU; the right chain converts it twice, and the stages it removes are the ones that sat inside the rack generating heat.
The transition will arrive in three forms, often in the same hall:
- Native 800 VDC racks. NVIDIA Kyber, housing 576 Rubin Ultra GPUs, ships in 2027 with 800 VDC input designed in.
- Sidecar power racks. A dedicated rack beside the compute rack takes the hall's feed, converts it, and supplies the compute rack, which lets existing halls host Rubin-class systems without re-cabling the floor; NVIDIA showed an 800 V sidecar at GTC 2025 and Schneider Electric plans to have its sidecar on the market well before Rubin Ultra ships in 2027.
- Row-level 800 VDC busways. The distribution upgrade that lets a hall carry 85% more power through the same conductor size.
The ecosystem is already lined up. At the OCP Summit in October 2025, more than 20 partners showed 800 VDC silicon, power shelves and reference designs, including Vertiv's 800 VDC MGX reference architecture and systems from Eaton, Schneider Electric, ABB, Siemens, Hitachi Energy and GE Vernova. CoreWeave, Lambda, Nebius, Oracle Cloud Infrastructure and Together AI are designing for 800 V, and Foxconn's 40 MW Kaohsiung-1 facility is already running on it.
Six CapEx benefits of 800 VDC conversion
NVIDIA's bundled claim for the architecture is up to 30% lower total cost of ownership, and the components of that number map directly onto CapEx line items. The figures below are NVIDIA's unless noted, and they assume a full 800 VDC build; a brownfield sidecar deployment captures a subset.
| CapEx line | What 800 VDC changes | Figure |
|---|---|---|
| Conductors and busbars | Higher voltage, lower current, thinner conductors | 45% less copper; 85% more power through the same conductor size vs 415 VAC |
| Rack power shelves and PSUs | In-rack AC/DC conversion removed | Frees up to 64 U per MW-class rack for compute; fewer fan-cooled PSUs to buy and replace |
| Conversion equipment | Four-plus conversion stages collapse to one rectifier plus in-rack DC/DC | Up to 5% better end-to-end efficiency, which is 5% less feed per MW of IT |
| Maintenance reserves | Fewer failure points, centralized conversion | Up to 70% lower maintenance cost |
| Electrical distribution scale | Same infrastructure spans 100 kW to over 1 MW racks | Expansion without re-cabling the hall |
| Cooling plant | Conversion losses and PSU heat leave the white space | NVIDIA lists "lower cooling expenses" from removing in-rack AC/DC PSUs |
The copper line is the one finance teams notice first. Copper prices are up 46% since December 2021, and a 1 GW campus on 54 V busbars would need up to 200,000 kg of copper for rack busbars alone. Cutting that by 45% is real money and shorter lead times.
The efficiency line compounds. Five percent less loss between the grid and the GPU means the same interconnect energizes 5% more IT load, and the rejected heat that cooling must handle drops in proportion. On a 100 MW feed that is roughly 5 MW of compute recovered without a utility upgrade.
The cooling line is the one that is easiest to get wrong. Moving rectification to the perimeter removes heat from the rack, but it does not remove it from the site; the rectifiers still reject it, just somewhere with cheaper cooling. And none of this changes the fact that a 1 MW Kyber rack is a 1 MW heat source that needs direct liquid cooling, 45°C supply loops and CDU capacity sized for it. 800 VDC lowers the cooling CapEx curve; it does not flatten it.
Where Synapse Thermal fits: the cooling side of the 800 VDC business case
800 VDC fixes the power chain. The cooling plant that sits behind it still runs on static setpoints tuned for the worst hour of the worst day, and during the transition the worst case moves every quarter. A single hall will host air-cooled legacy rows, liquid-cooled NVL72 racks with in-rack PSUs, sidecar power racks adding heat next to the compute they feed, and eventually native 800 VDC Kyber racks with their rectifiers out at the perimeter. The thermal map is uneven and it changes with every phase of the build-out.
That leaves a CapEx planner with two bad options: buy the chillers, CRAHs and CDUs for the final-state load on day one, or buy them in lurches as each phase lands and hope the lead times cooperate. Synapse Thermal exists to give you a third option.
Synapse Thermal is an autonomous cooling control layer for AI data centers. A deep reinforcement learning agent reads server temperatures, ambient conditions, workload metrics and power draw, forecasts thermal load 15 to 30 minutes ahead with LSTM models, and continuously adjusts fan speeds, chilled-water flow and liquid cooling loops to hold thermal constraints at the lowest energy cost. It is hardware-agnostic across HVAC, CRAC, direct-to-chip liquid and immersion systems, runs inference on-premise with no cloud dependency, and deploys in four to six weeks without replacing equipment. In simulation and deployment it runs 35 to 40% more efficiently than PID and rule-based control, cutting cooling energy by up to 40%.
For an 800 VDC program, that translates into four planning levers:
1. Recover stranded cooling capacity before you order new plant
Static setpoints and worst-case provisioning lock up headroom in equipment you have already paid for. Matching cooling to real thermal load, row by row, frees that headroom and tells you how much more IT load the existing plant can safely carry.
2. Defer the next chiller, CRAH or CDU
With a measured headroom figure instead of a nameplate one, major cooling purchases move later in the phasing plan, which at 2026 equipment prices and lead times is the single easiest CapEx deferral available. Customers could potentially recover the full cost of the platform in 8 to 14 months on energy savings alone, before counting deferred equipment.
3. Shift megawatts from cooling to compute
At a fixed interconnect, every kilowatt the cooling plant stops using is a kilowatt available to IT load. Every 0.1 of PUE on a 75 MW feed is roughly 4 to 5 MW of GPUs that can be energized. Stack that on top of 800 VDC's 5% power-chain gain and the same utility feed hosts materially more compute.
4. Phase the build-out on forecasts, not guesswork
Because the agent learns the site's behavior over time, facilities and finance teams get data-backed headroom estimates for each rack addition, density upgrade and architecture change. That is what de-risks running mixed AC, 48 V, sidecar and 800 VDC architectures in the same hall: the cooling adapts to the mix instead of the plan being hostage to it.
The outcome is a cooling plant sized to what the site actually does, and an 800 VDC CapEx plan that keeps the savings the power architecture promised.
A six-step CapEx planning framework for the 800 VDC transition
The order matters: cooling headroom is measured first, because it changes how much of the power upgrade you need and when.
- Baseline the plant you have. Measure real PUE by hall and by season, chiller and CDU utilization against nameplate, and how much of the cooling overhead is worst-case provisioning rather than load. A cooling audit should produce a headroom number in kW, not an opinion.
- Model the power chain per MW, both ways. Build the grid-to-GPU loss budget for your current 415 VAC path and for the 800 VDC path (perimeter rectifier, busway, in-rack DC/DC). The difference, up to 5%, is IT load you can add at the existing interconnect.
- Pick the deployment form per phase. Sidecar racks let an AC hall host Rubin-class systems in 2026 and 2027; native 800 VDC rows make sense where you control the hall from the busway up. Most sites will run both, so plan the cooling for the mix, not the end state.
- Size cooling to measured load, with optimization in the baseline. Run the headroom math assuming autonomous control is in place, because that is the cheaper plant. Only the load that still exceeds measured headroom justifies new chillers, CRAHs or CDUs.
- Sequence purchases against lead times. Transformers at 68 to 113 weeks and switchgear at 38 to 63 weeks set the critical path. Cooling purchases that optimization pushes later should come off that path entirely.
- Price the stranded megawatt. Delay costs roughly $0.23 million per MW per month in lost rent and carrying cost, so every MW of IT load that optimization or 800 VDC lets you energize early has a hard dollar value in the model.
Two numbers belong at the top of the resulting business case: the IT megawatts recovered at the existing interconnect, and the cooling CapEx moved out of the current budget year. Both are measured, not estimated, once a site has run under autonomous control for a few weeks.
Frequently asked questions about 800 VDC and data center CapEx
What is 800 VDC in a data center?
800 VDC is a power distribution architecture that converts medium-voltage AC grid power to 800 volts DC at the data center perimeter and distributes DC through the hall to the rack, where it is stepped down for the GPUs. It replaces the chain of AC/DC and DC/DC conversions in today's 415 VAC to 54 VDC designs.
Why are AI data centers moving to 800V DC?
Rack power is heading toward 1 MW, and a 54 V busbar cannot carry that current without roughly 200 kg of copper and 64 U of power shelves per rack. Higher voltage means lower current, so 800 VDC delivers 85% more power through the same conductor, cuts copper by 45% and removes conversion losses, according to NVIDIA.
When will 800 VDC data centers be available?
NVIDIA plans full-scale 800 VDC production alongside its Kyber rack systems in 2027. Sidecar power racks and row-level 800 VDC busways from Vertiv, Schneider Electric, Eaton and others let existing halls start the transition before then, and Foxconn's 40 MW Kaohsiung-1 site is already operating on 800 VDC.
Does 800 VDC reduce data center cooling requirements?
Partly. Removing in-rack AC/DC PSUs and up to 5% of conversion loss takes heat out of the white space, which NVIDIA cites as lower cooling expense. The GPUs themselves still reject every watt they draw, so a 1 MW rack needs 1 MW of liquid cooling capacity regardless of how the power arrives.
Can an existing data center convert to 800 VDC?
Yes, in phases. Brownfield halls typically add sidecar power racks or 800 VDC busways to specific rows while the rest of the floor stays on AC, which is why most facilities will run mixed AC, 48 V, sidecar and 800 VDC architectures for several years.
How does cooling optimization defer CapEx?
Static cooling control provisions for the worst case and leaves headroom unused. Autonomous DRL control matches cooling output to measured load, which recovers that headroom, lowers PUE, and shows how much more IT load the existing chillers and CDUs can carry before new equipment is justified. Synapse Thermal customers typically see payback in 8 to 14 months from energy savings alone.
How much compute does a 0.1 improvement in PUE unlock?
At a fixed utility interconnect, IT capacity equals feed capacity divided by PUE. On a 75 MW feed, moving from 1.25 to 1.15 adds about 5.2 MW of IT load, which is on the order of $150 to $220 million of GPUs at 2026 prices.
Plan the power upgrade and the cooling plant together
800 VDC conversion is the right call for any site that expects to host Rubin-class racks, and the CapEx benefits are large: 45% less copper, up to 5% better end-to-end efficiency, up to 70% lower maintenance cost and up to 30% lower TCO. But those numbers are only as good as the cooling plant behind them. A hall that still over-provisions cooling for the worst case will spend the 800 VDC savings on chillers and CDUs it did not need yet, and leave megawatts of interconnect stranded.
Synapse Thermal gives facilities and finance teams the measured headroom, the deferred equipment and the recovered megawatts that make the 800 VDC business case hold. Deployment takes four to six weeks on the equipment you already own, and the platform pays for itself in 8 to 14 months.
Find out how much more load your cooling plant can carry before you buy another chiller. Request a capacity assessment or contact info@synapsethermal.com to see the headroom at your site.
Related reading: Why PUE Alone Won't Save the Next Generation of AI Data Centers and How to Reduce Data Center Cooling Energy Without Replacing Existing Equipment.
Sources
- NVIDIA 800 VDC Architecture Will Power the Next Generation of AI Factories, NVIDIA Technical Blog, May 20, 2025 (updated Jan 2026)
- Nvidia prepares data center industry for 1MW racks and 800-volt DC power architectures, DCD, Oct 14, 2025
- The 1 MW AI IT rack is coming, and it needs 800 VDC power, Schneider Electric, Oct 16, 2025
- Data Center Cost per MW in 2026: The Full Stack, GR1D, Oct 15, 2026 (compiles JLL Jan 2026, Cushman & Wakefield Sep 2026, Turner & Townsend Nov 2025, CRE Daily Aug 2026)
- Cushman & Wakefield 2026 Data Center Development Cost Guide, Sep 2026
- JLL 2026 Global Data Center Outlook, Jan 2026
- Synapse Thermal: platform capabilities, CapEx and capacity planning, and 800 VDC transition positioning
Synapse Thermal performance figures (up to 40% cooling energy reduction, 35 to 40% gain over PID control, 4 to 6 week deployment, 8 to 14 month payback) are the company's published claims. Interconnect and GPU-value calculations in this post are illustrative arithmetic on the cited benchmarks.