ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
News4 min read·2026-10-21DisconfirmAI Editorial

From PUE to Tokens Per Watt: Why AI Data Center Efficiency Metrics Need to Change

In large AI clusters, many devices simultaneously enter compute, communication, or idle phases, causing power to shift rapidly. Traditional PUE measures facility overhead but cannot reflect effective computational output. SenseTime introduced TPW (Tokens Per Watt) as a full-chain efficiency metric at the 2026 AIDC Infrastructure Summit, while Huawei emphasized tokens-per-watt output and cost optimization from a compute-factory perspective, bringing grid supply, power delivery, cooling, and compute scheduling under a single evaluation target.

ImportanceMaterialEvidenceE3 inspectableWrite-upStandard

Original publication time 2026-09-28 19:18 GMT+8

The data center industry has long used PUE to measure energy efficiency: total facility energy divided by IT equipment energy. This number tells you how much electricity the power and cooling systems consume as overhead, but it cannot answer a more practical question—how much effective computation the electricity entering the servers actually produces.

AI workloads make this question urgent. In large GPU clusters, many devices simultaneously enter compute, communication, or idle phases as tasks progress, causing power to shift rapidly. Two data centers with identical annual energy consumption may have entirely different requirements for power control and energy storage if one runs a steady load while the other experiences frequent synchronous power fluctuations. Under the same PUE, device utilization, task completion volume, and business cost can differ substantially.

Lin Hai, General Manager of SenseTime's Large Device Business Group Intelligent Computing Center, introduced at the AIDC Infrastructure Summit held alongside Huawei Connect 2026 that SenseTime is moving from PUE-centric energy management further toward TPW (Tokens Per Watt) as the core of full-chain efficiency management. This metric directly links electricity efficiency to token output, placing grid power plants, IDC rack services, compute cloud services, and model services under a single evaluation chain. According to his disclosure, SenseTime's Large Device platform has achieved an 80% improvement in token output per unit of electricity cost through end-to-end full-chain optimization; the Lingang Intelligent Computing Center is also the industry's first to receive the China Academy of Information and Communications Technology's 5A certification for compute-power synergy.

Sign in / Sign up
ReadingResearchRadarInvestment framework

Hou Jinlong, Huawei Board Member and President of Huawei Digital Power, at the same summit summarized 'more tokens per watt, lower cost per token' as the core competitiveness of a compute factory.

The introduction of TPW is not merely swapping one number for another—it changes the coordination logic across every layer of infrastructure. When the evaluation target extends to actual computational output, power systems must respond to compute load fluctuations in real time, thermal management systems must share operational information with compute devices, and scheduling strategies must jointly account for electricity prices, energy storage costs, and task latency constraints. The TMU thermal management controller Huawei demonstrated at the summit achieves native adaptation with Ascend servers, with cooling-power coordination responding to compute load—this is one concrete implementation of joint design.

This trend is not unique to Huawei. NVIDIA, Google, and Emerald AI announced the founding of the AI Energy Management Alliance (AEMA) on September 16, promoting dynamic adjustment of data center power consumption based on grid operating conditions. Google disclosed in March that it has incorporated a combined 1 GW of demand response capacity into long-term energy contracts with multiple U.S. utility companies, coordinating grid operations by limiting or shifting portions of machine learning tasks. The International Energy Agency (IEA) projected in its 2026 report that global data center electricity consumption will grow from 485 TWh in 2025 to 950 TWh by 2030, nearly doubling; AI-focused data center consumption is expected to roughly triple over the same period.

Policy is tightening as well. A royal decree draft on data center regulation published by Spain in August 2026 requires that for each operating hour, at least 80% of electricity consumption be covered by new renewable generation within that same hour—annual aggregate compliance is insufficient. Four Chinese national agencies jointly issued an action plan in April 2026 explicitly encouraging grid-forming energy storage for compute facilities and supporting direct green electricity connections. The European Commission adopted a unified data center sustainability rating scheme on September 21, with first labels expected in 2027.

TPW as an evaluation metric still has clear boundaries. Different models, output quality requirements, and latency constraints all affect token production; energy accounting periods and system boundaries must also be consistent, making direct cross-task TPW comparisons invalid. SenseTime's 80% improvement is self-reported data—its specific optimization path, baseline conditions, and verification methodology have not been fully disclosed in public materials. Huawei's AIDC 1.0 specifications, including the Taishan UPS single-rack capacity of 1,280 kVA and power module delivery time reduced from 7 days to 3, are likewise vendor self-described figures.

PUE retains value—it measures the energy overhead of the infrastructure itself. TPW adds another observation dimension: under conditions that meet quality and service requirements, how much effective computation a given AI data center can continuously deliver. The two metrics coexist rather than replace each other, but the full-chain coordinated design that TPW demands is pushing power supply, energy storage, cooling, and compute scheduling from siloed operation toward joint optimization.

Sources

  • 原始出处qbitai.com↗
Continue researching →