AI 數據中心基建同時遭遇三大物理瓶頸，複合效應更嚴重

Here is a trajectory that should focus the mind of anyone designing AI infrastructure right now. In 2023, a typical accelerator rack consumed around 20 kW and each chip dissipated roughly 350 watts. By 2024, rack power climbed to 50 kW; chip power hit 700 watts. In 2025, we crossed the kilowatt-per-chip threshold - racks reached 120 kW, individual processors exceeded 1,000 watts, and heat flux density landed between 50 and 100 W/cm². The current generation, shipping in 2026, pushes racks to 240 kW, chips past 2,000 watts, and heat flux to 120 W/cm².

The next generation - already in validation at multiple hyperscalers - targets megawatt-class racks, 3,000-plus watt chips, and heat flux densities above 200 W/cm².

That is not a gentle curve. It is a step function that breaks three things at the same time: cooling, power delivery, and electrical interconnect. At the 2026 OCP Global Summit in San Jose, xFusion's Director of Product Management Arthur Wang walked through the engineering detail behind all three in a keynote titled "Next-Generation Power & Cooling Solutions." What made the talk notable was not any single product announcement but the framing: these three constraints are not independent. They interact, they compound, and solving any two while ignoring the third still results in stranded infrastructure.

xFusion calls this "computing-power synergy." The less catchy, more precise description: co-optimizing the entire energy conversion chain from grid to transistor - cooling, power delivery, and interconnect treated as coupled variables rather than separate procurement categories.

The thermal wall: materials science versus Moore's Law

Air cooling is finished for high-density AI racks. That is no longer a controversial statement. The question is how much headroom the liquid cooling solutions being deployed today actually provide - and whether they will survive the next chip generation without a forklift upgrade.

The physics are straightforward. At 200-plus W/cm², convective air cooling cannot extract heat fast enough to prevent thermal throttling, regardless of plenum geometry or fan configuration. Direct liquid cooling - specifically cold plate-based DLC - is the only viable path. But a cold plate is only as good as the materials and micro-channel geometries inside it.

xFusion's approach is to use diamond-copper composite cold plates. The numbers: equivalent thermal conductivity exceeding 650 W/m·K, roughly twice that of pure copper (385–400 W/m·K). The cold plate design uses non-uniform fin geometries mapped to chip hot spots, yielding an average heat flux handling capacity of 200 W/cm² across the full cold plate area and over 400 W/cm² at localized hot spots. Total single-chip cooling capacity exceeds 4,000 watts.

That 4,000-watt figure matters because it provides roughly 33 percent headroom above the incoming 3,000-watt chip generation. In liquid cooling, headroom is not luxury - it is insurance against the silicon roadmap moving faster than the mechanical refresh cycle.

The material matrix behind this is broader than a single cold plate. xFusion presented five core thermal interface and heat-spreading materials at OCP, each targeting a different cost-performance point: graphene thermal pads (130 W/m·K, filling irregular gaps at 0.3–5 mm thickness), high-density copper skived cold plates (385–400 W/m·K, the workhorse for current-generation chip cooling), liquid metal TIM (30 W/m·K but with interface thermal resistance as low as 0.02 °C·cm²/W, eliminating microscopic air gaps), pure diamond cold plates (800–2,500 W/m·K, for extreme heat spreading), and the diamond-copper composites described above (550–800 W/m·K, combining diamond's conductivity with copper's machinability).

Looking further out, the company is developing two-phase jet cooling - using the latent heat of liquid-to-gas phase change rather than single-phase sensible heat to absorb energy from the chip surface. The advantage is dramatically reduced coolant flow rates for the same thermal load. The engineering challenges - sealing at scale, pressure management, blind-mate fluid connectors for serviceability - are substantial, but so is the thermal runway if single-phase DLC is the only option and chip heat flux keeps climbing.

For current deployments, xFusion's FusionPod for AI V8 - a 42U modular liquid-cooled rack rated at up to 155 kW - ships with a distributed CDU architecture requiring zero secondary piping modifications. That last detail matters more than it sounds. Retrofitting liquid cooling into an existing facility typically means ripping up floors to install secondary coolant loops - a process measured in quarters, not weeks. Eliminating that dependency compresses deployment timelines from years to months.

The electrical wall: why 48V-class distribution cannot survive the megawatt era

Cooling and power delivery are often treated as separate disciplines. They are not. Every watt delivered to a chip must leave as heat. The path that watt travels - from utility transformer through conversion stages to the voltage regulator on the board - bleeds energy at every step.

At today's standard 48V-class bus voltages (54V nominal), the current required for megawatt-scale racks is staggering. A 10 kW rack draws 185 amps. A 40 kW rack: 741 amps. A 120 kW rack: 2,222 amps. A 240 kW rack: 4,444 amps. A 500 kW rack: 9,259 amps. And a 1 MW rack: 18,519 amps.

The I²R loss implications are brutal. Copper busbar losses scale with the square of current. When rack power increases 25× (from 40 kW to 1 MW), current increases 25× - but resistive losses increase 625×. At 1 MW per rack with 54V distribution, busbar copper losses alone can exceed 50 kW. That is the power consumption of an entire server, wasted as heat in the busbar before a single floating-point operation is performed. And the busbars themselves become impractical objects - copper conductors weighing upwards of 200 kg per rack, creating structural and thermal challenges of their own.

The solution is straightforward in principle and formidable in execution: raise the distribution voltage. Moving from 54V to 800V DC - a factor of roughly 14.8× - reduces current by approximately 15× and I²R losses by approximately 220×. Simultaneously, the number of power conversion stages drops from four (in legacy AC/54V architectures) to three, pushing end-to-end system efficiency from roughly 88 percent to above 90 percent, and compressing line losses from 5 to 8 percent to below 1 percent.

xFusion is building toward this with its 800V HVDC Sidecar - a next-generation power supply prototype engineered for 300 kW to 1.2 MW configurations per rack. The roadmap also includes a direct 800V HVDC architecture that eliminates additional conversion stages entirely and natively supports green power integration - a detail that becomes increasingly important as grid-tied renewable capacity and behind-the-meter generation become standard features of new AI campuses.

Underpinning the voltage transition is a shift in semiconductor devices. xFusion is deploying wide-bandgap materials across the power conversion stack: SiC MOSFETs for AC-to-HVDC rectification (PFC + LLC topology, 100–300 kHz switching), GaN HEMTs for HVDC-to-DC conversion (LLC resonant, 800 kHz–1 MHz), and GaN-based multi-phase voltage regulators for the final 54V-to-12V/6V stage at 500 kHz to 1 MHz. Higher switching frequencies translate directly to smaller magnetics and higher power density - critical when every cubic centimeter of rack space is contested.

The company is also advancing vertical power delivery - placing power conversion modules directly behind the chip package rather than routing current laterally across the motherboard. This shortens the power path by roughly 90 percent, reducing both resistive losses and the transient response challenges that emerge when thousands of amps must be delivered to a processor that can swing its current draw by hundreds of amps in microseconds.

The interconnect wall: copper at 224G and beyond

The third constraint gets less attention but is equally binding at scale. Electrical signaling over copper traces at 224 Gbps per lane is already operating at the ragged edge of signal integrity. Channel loss, insertion loss, and crosstalk at these frequencies demand extraordinary equalization and error correction. Beyond 224G, the physics of copper become unforgiving.

The transition to optics is underway. Co-packaged optics (CPO) - placing photonic engines directly on or adjacent to switch silicon - collapses the distance penalty that limits electrical reach and unlocks bandwidth density that copper backplanes cannot match. For AI training clusters running thousands of accelerators in tightly coupled scale-up domains, interconnect bandwidth directly determines how large a distributed job can scale and how efficiently inference workloads distribute.

xFusion has commercially deployed 224G high-density orthogonal backplanes and is conducting pre-research on 448G. Its Super Node architecture - an open-architecture liquid-cooled system - delivers up to 240 kW per rack, supports 128 XPU per cabinet and 256 XPU across dual cabinets, and is designed for multi-vendor processor compatibility from the ground up. That openness is not philosophical - it is a hedge against the accelerator market's rapid fragmentation, where locking into a single vendor's proprietary interconnect fabric trades a bandwidth constraint for a supply chain dependency.

The compound problem

None of these three walls - cooling, electrical, optical - respects the others' timeline. They are all arriving in the same product generation. And they interact in ways that make partial solutions dangerous.

World-class liquid cooling paired with legacy 54V power delivery? The electrical bottleneck throttles the rack before the coolant warms up. Pristine 800V DC power feeding a cluster with inadequate optical fabric? Stranded GPUs waiting for data. High-bandwidth networking behind a thermally throttled compute tier? Wasted tokens. Wasted capital.

This is why the "computing-power synergy" framing - treating cooling, power, and interconnect as a single coupled optimization problem from grid to chip - is more than a marketing theme. It is an engineering discipline. And it is why xFusion has invested in the full vertical stack: from chip-level thermal interface materials and cold plates, through board-level and rack-level power conversion, to cabinet-scale liquid cooling and backplane interconnect.

The company's participation in more than 60 industry organizations, contribution to more than 60 open standards, and ecosystem of over 30,000 partners worldwide reflects a bet that no single vendor can solve the converged problem alone - but someone has to integrate the pieces coherently.

As Wang put it at OCP: "In the Agentic era, competitiveness is not about how many chips you have. It is about how much real business value you can extract from every watt and every joule of heat."

That statement maps to a value chain xFusion presented on stage -WATT → FLOPS → TOKEN → AGENT → VALUE - where each stage introduces conversion losses and the facility's true performance is measured not by watts consumed but by the useful business outcomes produced. It is, in essence, an argument that the AI datacenter is not a building that houses computers. It is an energy conversion system. And three physical walls are about to determine who converts efficiently and who does not.

Explore more real cases here, or request a consultation here.

Supplied by xFusion