From PUE to Tokens per Watt: Huawei AIDC Is Changing the Yardstick

For nearly two decades, the AI data center industry has used one yardstick: PUE. It measures how much of the total electricity consumed actually reaches IT equipment, but it cannot answer a more fundamental question: how much useful compute does that IT power ultimately produce? As AIDCs scale from megawatt campuses to gigawatt campuses, and as the “Token factory” becomes an industry consensus, PUE’s blind spots are being exposed.

Huawei’s response in 2026 is not to squeeze another few tenths of a point out of PUE, but to change the yardstick: from PUE to “Tokens per Watt” (TPW)—the number of Tokens ultimately produced per watt of electricity. SenseTime’s SenseCore has already validated this path in practice: through end-to-end optimization from power plants and the grid to IDC racks, cloud services, and model services, SenseCore improved TPW by 80%.

This new yardstick forces power and cooling to move from “supporting systems” to “components of output.”

Power: From “Stable Supply” to “Grid Support,” and Then to “Absorbing Fluctuations” Measured by TPW, the task of the power chain is no longer just “keeping the lights on.” AI load fluctuations span multiple timescales, from microseconds to hours. Pushing all of them onto the grid is neither fair nor realistic. Huawei’s answer is multi-layer hybrid energy storage coordination: rack-side storage handles microsecond/second-level fluctuations, facility-side storage handles millisecond/minute-level fluctuations, and campus-side grid-forming storage handles hour-level fluctuations—absorbing as much of the AI load’s “temperament” inside the AIDC as possible.

At the product level, two key nodes deserve attention. Taishan UPS achieves 1,280 kVA per cabinet and 98% double-conversion efficiency, making it the industry’s first single-cabinet MW-class UPS, capable of smoothing about 70% of AI load fluctuations. Hengshan DC UPS is the first software-defined DC UPS, with a single hardware platform supporting 270 V, 400 V, and 800 V simultaneously, so the power architecture is not locked to the voltage standard of any single chip generation. Together, they point in the same direction: power supply is no longer a “one-time engineering project,” but an elastic foundation that can evolve with compute architectures.

The further anchor is SST (solid-state transformer). Huawei has explicitly stated that it will gradually evolve toward an integrated grid-forming energy router centered on SST, replacing conventional iron-core transformers with power electronics to enable bidirectional power flow and active voltage regulation. Once this step is completed, the AIDC will change from a “power-hungry beast” in the eyes of the grid into an adjustable, supporting node.

Cooling: From “Removing Heat” to “Letting Compute Run at Full Speed” If power answers “how much electricity is available,” cooling answers “how much useful compute that electricity can produce.” The passive logic of a traditional CDU is: supply cooling at a set value, regardless of what workload the servers are running. But AI load power curves fluctuate sharply. Fixed cooling either under-cools and causes throttling, or over-cools and wastes electricity that could otherwise be used for compute.

The breakthrough of Huawei TMU is not heat-exchange capability itself, but that it breaks the data barrier between the cooling system and the compute system. TMU is natively compatible with Ascend servers. Through cooling-power synergy, the cooling system can respond quickly to changes in compute load rather than waiting for temperatures to rise and then adjusting reactively. This means that when load drops, cooling output drops synchronously, and the saved electricity goes directly back into the numerator of “electricity available for Token production.” On the safety baseline, TMU uses a dual AC/DC power architecture to achieve 0 ms primary/backup switching, and its high-efficiency plate heat exchanger achieves an approach temperature as low as 3°C—moving from “usable” to “safe to rely on long-term.”

Huawei Cloud’s Wuhu Data Center offers a reference point: annual average liquid-cooling PUE as low as 1.1, saving nearly 1 billion kWh per year per million servers. In the TPW context, every kilowatt-hour saved is not the endpoint, but the starting point for producing more Tokens.

The Real Variable: Compute and Power No Longer Belong to Two Separate Departments When power and cooling are placed under the TPW yardstick, a fact previously hidden by PUE becomes visible: the efficiency bottleneck of AIDCs often lies not in the parameters of individual devices, but in the depth of synergy between “power” and “compute.”

Multi-layer energy storage absorbs fluctuations so the power side does not push the burden onto the grid; TMU works with Ascend servers so the cooling side does not push the burden onto over-cooling; SST and grid-forming storage move toward grid interaction so the AIDC can change from a passive load into an active node. These three things share the same logic: along the path of every watt from entering the campus to becoming a Token, the “inaction” or “over-action” of every link will ultimately show up in the numerator or denominator of TPW.

While the industry is still debating how low PUE can go, Huawei and its customers are already using TPW to answer a more essential question: the value of an AIDC is not how much electricity it consumes, but how close it brings every watt of electricity to the Token.