Overview: From a FLOPS Race to a Physics Fight

The competition in AI computing infrastructure is no longer confined to chip-level throughput. As Nvidia's latest-generation GPUs and specialized AI ASICs push single-die thermal design power beyond 1000W and toward 1500W, and per-rack power density climbs from the traditional 8kW–10kW to 100kW–120kW and above, conventional data center architectures are being hit with what amounts to a structural shock. The binding constraint is shifting from transistor counts to two physical limits: delivering kilowatt-scale electricity into microscopic structures, and removing kilowatt-scale heat from them. The response is coalescing around 800V high-voltage DC (HVDC) distribution, vertical power delivery (VPD), and cold-plate and phase-change immersion liquid cooling — with the United States and China occupying markedly different industrial positions.

The Power Wall and the Thermal Wall

Two hard boundaries define the current inflection point. The first is the power wall: traditional 12V/48V low-voltage distribution must carry kiloampere-level currents, generating enormous I²R copper losses, while board-level power routing crowds out precious die area and high-density interconnect space. The second is the thermal wall: air's specific heat capacity and thermal conductivity can no longer satisfy kilowatt-class chip cooling, meaning air cooling has failed at the physical level and intelligent computing centers must move from air to liquid.

Vertical Power Delivery: Rebuilding Power at the Micro Level

When chip operating voltage falls to 0.8V–1.0V while single-chip power reaches 1000W–1500W, the supply bus must carry 1000A–1500A. Under Joule's law (P = I²R), resistive heating and IR drop rise quadratically with current. In conventional lateral PCB architectures, complex power traces occupy more than 50% of the board area, competing directly with the micron-scale routing between the GPU and HBM memory, and 10%–15% of energy is dissipated as useless heat along the transmission path.

Vertical power delivery addresses this by stacking miniature DC-DC converters and inductor modules directly beneath the GPU, on the backside of the package substrate, then injecting current vertically into the transistor core through high-density through-silicon vias (TSVs) or micro-bumps. The delivery path shrinks from tens of millimeters to hundreds of microns, board-level resistance falls by more than 90%, IR drop drops to millivolt levels, and overall system efficiency improves qualitatively.

800V HVDC: Rewriting the Macro Power Architecture

The macro-level distribution architecture is being rebuilt in parallel. Traditional AC distribution requires repeated AC-DC-AC-DC conversion, with total link losses of 15%–20%. 800V HVDC is emerging as a hard requirement for 100,000-GPU clusters: utility or green power is converted directly to 800V DC and delivered to the rack, eliminating three unnecessary conversion stages and lifting end-to-end transmission efficiency above 95%. Because higher voltage means lower current, transformer and switchgear footprint shrinks by roughly 40% and machine-room copper cable usage drops by 60%.

The Thermodynamics of Cooling: Why Air Must Go

Cooling efficiency depends on the medium's heat capacity and thermal conductivity. Air conducts heat at roughly 0.026 W/(m·K); water at about 0.6 W/(m·K); specialty fluorinated liquids reach hundreds of times air's capability. Once chip heat density exceeds 100 W/cm², air-cooled heatsinks and fans must expand dramatically, fan power alone can exceed 20% of total server power, and noise passes 90 decibels. Liquid flowing through a metal cold plate removes kilowatt-scale heat precisely, pulling PUE from roughly 1.5 under air cooling to below 1.15.

  • Cold-plate cooling: sealed copper or aluminum microchannel plates sit against the GPU while coolant circulates inside, never contacting electronics. Lower retrofit cost and strong compatibility with existing servers make it the dominant commercial route for 2025–2026 deployments.
  • Immersion cooling: entire servers — GPU, motherboard and power supply — are submerged in insulating fluorinated liquid or hydrocarbon coolant. Boiling phase change carries away large latent heat, and a condenser restores the liquid. With no fans or vibration and no cooling blind spots, PUE can approach the physical limit of 1.02–1.05.

The US–China Contest Across Four Tracks

  • Micro-level power (VPD): the US holds high-frequency, high-power-density DC-DC topology patents, GaN driver ICs and soft-switching know-how through Vicor, MPS and Texas Instruments, while Intel and Nvidia help shape VPD standards inside 2.5D/3D advanced packaging. China leads in module power engineering and manufacturing (Huawei Digital Power, Sungrow, Megmeet) and is accelerating domestic substitution of high-phase-count PWM controllers, DrMOS and high-frequency miniature inductors.
  • 800V HVDC distribution: China leverages its ultra-high-voltage grid and green power base, with Huawei, Inspur and Delta in the first tier for HVDC cabinets, high-power rectifier modules and megawatt modular data centers; the "East Data, West Computing" hubs are demonstrating solar-wind-storage direct DC coupling. The US is pursuing microgrid AI scheduling algorithms, with Microsoft, Amazon and Google investing in small modular reactors (SMRs) and geothermal for continuous off-grid zero-carbon power.
  • Cold-plate liquid cooling: China's Envicool, Sugon and Shenling bring strong cost and delivery advantages in cold plates, CDUs, quick connectors and fluid piping. The US leads in rack-level standardization — as in Nvidia's GB200/NVL72 liquid-cooling systems — and thermal interface materials.
  • Immersion cooling: Alibaba and Baidu have accumulated deep operational data from long-term phase-change immersion testing. The US controls CFD and phase-change thermodynamic modeling standards and high-end electronic fluorinated liquid patents held by chemical majors such as 3M and Chemours.

Implications for AI Data Centers

Together, these shifts mean the data center is being redefined as an integrated electromechanical and thermodynamic system rather than a building full of servers. Power delivery migrates into the package, distribution migrates to 800V DC, and heat rejection migrates into sealed liquid loops. Operators that treat power and cooling as afterthoughts will find their expansion plans capped by grid interconnection, copper consumption and floor space — precisely the constraints that 800V HVDC and liquid cooling are designed to relax.

Future Outlook

The contest over AI computing power has become a contest in electromagnetism, fluid dynamics and materials physics. The US continues to hold infrastructure standards and top-tier intellectual property in power topologies, chip-level packaging integration, thermodynamic simulation and advanced chemicals. China is translating construction scale, an efficient green grid and full-chain manufacturing into deployment speed and cost reduction. In the coming era of 10,000- and 100,000-GPU clusters, only companies and countries that break through both the micro- and macro-level power and thermodynamic limits will hold the high ground of AI computing.