NVIDIA's H100 accelerator draws up to 700W in its SXM configuration and 350W in its PCIe configuration, a power level NVIDIA attributes to a TSMC 4N process and 80 billion transistors packed onto a single chipCITE:E1. Vertiv reports that as rack power density rises to 20 kW and quickly approaches 50 kW, traditional room-based air cooling is pushed to its capability limits for HPC infrastructureCITE:E2. Together, these two data points mark the point at which chip-level heat output and rack-level power density both exceed what air cooling was designed to handle.
Direct liquid cooling (DLC) circulates a dielectric coolant through cold plates mounted directly on chip surfaces, while immersion cooling submerges entire servers in coolant. Vertiv states that its direct-to-chip cold-plate approach is 3,000 times more effective at removing heat than air cooling alone for HPC infrastructureCITE:E3. Immersion cooling, by contrast, absorbs 100% of the heat generated by IT components because they are fully submerged in fluid, with single-tank systems delivering cooling capacities of up to 100 kW across 42U or 52U tank configurations, using either single-phase (non-boiling circulation) or two-phase (coolant boils at the chip surface and recondenses) designsCITE:E4.
Liquid cooling is compressing both the power and water footprint of data centers relative to industry-wide averages. Uptime Institute's 2024 survey put the global average data center Power Usage Effectiveness (PUE) at 1.56, a figure it says has shown a continuing trend of inertia, while many recently built facilities — often specializing in denser IT architectures with direct liquid cooling — consistently achieve a PUE of 1.3 or betterCITE:E5. On water, Microsoft reports that a newer datacenter design using chip-level, closed-loop liquid cooling with zero operational water evaporation will avoid the need for more than 125 million liters of water per year per datacenter, and that its datacenters overall operated at an average Water Usage Effectiveness (WUE) of 0.30 L/kWh in its most recent fiscal year, a 39% improvement compared to the 0.49 L/kWh global average it reported in 2021CITE:E6.
| Metric | Prior / Baseline | Current / New Build |
|---|---|---|
| PUE (Uptime Institute, 2024) | 1.56 (industry average) | 1.3 or better (many recent builds) |
| WUE (Microsoft) | 0.49 L/kWh (2021) | 0.30 L/kWh (most recent fiscal year, -39%) |
| Water avoided (Microsoft, new design) | — | >125 million liters/year per datacenter |
| Rack cooling capacity (Vertiv immersion) | — | up to 100 kW per tank (42U/52U) |
NVIDIA's GB200 NVL72 is a rack-scale liquid-cooled platform integrating 72 Blackwell GPUs in a single rackCITE:E7. NVIDIA states that compared with H100 air-cooled infrastructure, the liquid-cooled GB200 NVL72 delivers 25 times more performance at the same power level, while also reducing water consumption and, according to NVIDIA, lowering the data center's carbon footprint and energy consumptionCITE:E7.
The chip-level power figures from NVIDIACITE:E1 and the rack-density ceiling reported by VertivCITE:E2 describe the same physical constraint from two angles — heat per chip and heat per rack — both landing liquid cooling as the response. Vertiv's own cooling-technology dataCITE:E3CITE:E4 and the facility-level metrics from Uptime InstituteCITE:E5 and MicrosoftCITE:E6 then show that constraint translating into measured outcomes: lower PUE at new builds and reduced water use at Microsoft's datacenters. NVIDIA's GB200 NVL72CITE:E7 packages both threads together in a single shipping product — a 72-GPU, liquid-cooled rack that NVIDIA positions against its own prior air-cooled generation on a same-power performance basis.
NVIDIA's H100 accelerator draws up to 700W in its SXM configuration and 350W in its PCIe configuration, a power level NVIDIA attributes to a TSMC 4N process and…
Direct liquid cooling (DLC) circulates a dielectric coolant through cold plates mounted directly on chip surfaces, while immersion cooling submerges entire serv…
Liquid cooling is compressing both the power and water footprint of data centers relative to industry-wide averages.
NVIDIA's GB200 NVL72 is a rack-scale liquid-cooled platform integrating 72 Blackwell GPUs in a single rackCITE:E7.
The two vendor-reported cooling technologies in this piece solve different problems, not the same one at different intensities: Vertiv's direct-to-chip cold plates target the 700W-class heat concentrated on a single H100 die, while its immersion tanks address whole-server heat loads up to 100 kW per tank. That split matters for how operators will likely mix the two rather than pick one. The single figure worth tracking going forward is whether more rack-scale platforms follow NVIDIA's GB200 NVL72 pattern — pairing a liquid-cooled rack design with a same-power performance claim (25x here) rather than treating cooling as a separate facilities line item. If that framing becomes standard for future GPU generations, it signals cooling architecture has moved from an operational afterthought to a stated part of the performance specification itself.
Utility-scale solar has become one of the cheapest new power sources, with costs down roughly 90% since 2010, while 2024 deployment volume outpaced every other generation technology. Paired with record-low battery prices, solar-plus-storage is already powering an AI data center in Arizona — though it still falls short of full 24-hour dispatchable baseload.
2026 marks the European Union AI Act's real enforcement start: the AI Office gains fining power over general-purpose AI on August 2, while high-risk system deadlines are pushed to 2027 and 2028. California activates two new laws on January 1 covering frontier-developer safety disclosure and training-data transparency. The federal government moves the opposite direction, ordering a Justice Department task force to challenge state AI laws, naming California's SB 53 as a target.
China's EHang (億航) is the only eVTOL maker actually flying paying passengers today, holding a full Chinese type, production, and operating certificate set and delivering 221 aircraft in 2025. US rivals Joby and Archer remain in certification or pre-launch stages, while Germany's Lilium and Volocopter both went insolvent in 2024–2025, with Volocopter absorbed by a Chinese buyer.