Bottom line: Most 800 VDC coverage focuses on the IT side: sidecars, busways, and rack-level DC-DC. But an AI facility also spends a big chunk of its power budget on non-IT loads such as chillers, pumps, and fans. Trane's new demo asks a simple question: if the hall already has an 800 V DC bus, why is the cooling plant still on AC?
What happened
- On September 30, 2026, Trane Technologies announced a laboratory proof-of-concept of an 800 V DC chiller, built with Eaton and Danfoss by modifying an existing high-efficiency chiller to accept a DC feed.
- The unit delivered more than 1,000 tons (about 3.5 MW) of cooling and showed potential for up to 2% higher system efficiency versus conventional AC equivalents.
- Trane estimates that in a 200 MW data center, the gain could free up to 1.8 MW for compute: up to fifteen 120 kW racks or three 600 kW racks.
- Trane says the architecture could extend to pumps and fans. The same week, Delta showed an 800 VDC in-row power system (660 kW) and an HVDC 3.6 MW in-row CDU at DCW Asia 2026 in Singapore, so DC-fed thermal hardware is becoming a trend rather than a one-off.
Quick sanity check on why 800 V: 1 MW at 54 V needs about 18,500 A. At 800 V it needs 1,250 A, roughly one-fifteenth. Lower current means smaller copper and lower I²R loss, which is the whole premise of 800 VDC distribution.
Engineering point 1: which conversion stage disappears
A modern high-efficiency chiller compressor is almost always on a variable-frequency drive: three-phase AC, then a rectifier (diode bridge or active front end), then the DC link, then the inverter, then the motor. In a facility where an SST or central rectifier already turns medium voltage into 800 V DC for the GPUs, keeping the chiller on AC means a parallel AC path and a second rectification inside the drive, effectively DC to AC to DC. Feeding 800 V straight into the DC link can, in principle, remove the front-end rectifier and much of its input filtering and harmonic mitigation. Trane's release doesn't itemize which stages were removed, so treat this as the likely mechanism, not a confirmed bill of materials.
Engineering point 2: why 2% is worth a press release
1.8 MW out of 200 MW is only 0.9% of the facility. But when grid interconnection is the bottleneck, every recovered megawatt is deployable compute. Trane's own conversions check out: 15 × 120 kW = 1.8 MW and 3 × 600 kW = 1.8 MW. Keep in mind that "up to 2%" is a lab-condition ceiling, not a guaranteed field number.
Engineering point 3: the hard part is protection, not delivery
AC crosses zero every half-cycle, and breakers rely on that to extinguish the arc. DC never crosses zero, so a fault arc keeps burning. A deployable 800 VDC chiller therefore depends on DC-rated or solid-state circuit breakers (SSCBs), a proper pre-charge circuit for the DC-link capacitors (otherwise closing onto the bus causes a massive inrush), and a clearly defined grounding scheme, for example a ±400 V midpoint-grounded bus. With compressors in the hundreds of kilowatts, these details matter more than the chiller itself.
Engineering point 4: EVs already work this way
An EV's electric A/C compressor draws power straight from the 400 V or 800 V traction battery bus through its own inverter, with no rectifier in between. We covered how that compressor loads the battery in our EV air-conditioning episode, and why higher voltage cuts current in our 400 V to 800 V SiC episode. The data center is now applying the same high-voltage DC logic at megawatt scale.
Open questions
- What is the 2% measured against: the cooling plant, total non-IT load, or the full power chain? The release doesn't say.
- This is a lab proof-of-concept. There's no long-term field data from a live facility yet.
- The release doesn't break down what Eaton and Danfoss each contributed.
- How compressor drives ride through 800 V bus swings caused by bursty GPU loads, and which safety standards will govern DC chillers, are still unpublished.
Watch on The Engineering Core
- AI's Power Crisis: SOFC vs SMR Nuclear, The Ultimate Energy Showdown
- Efficiency vs. Power Factor: The Real Reason AI Data Centers Need SSTs
- Does EV Air Conditioning Kill Your Range? How Much Range Do You Really Lose?
Your take: what's the real blocker for an all-DC data center: protection hardware, standards, or the cost of retrofitting existing plants? Tell me in the comments.
Sources: Trane Technologies press release (2026-09-30), Refindustry (2026-10-01), Data Center Digest (2026-10-02), Pokde: Delta at DCW Asia 2026. Current calculations are our own.
The Automation Core | theautomationcore.blogspot.com | Videos on The Engineering Core YouTube channel