AINEWS
SenseTime's Lingang, Shanghai, site took the top mark from CAICT (China's telecom and internet standards body) on September 12 in China's first grid coordination test for AI sites — an early marker that the industry's efficiency story is starting
China's first national test of whether AI data centers can flex with the power grid awarded its top grade on September 12 to a single site: SenseTime's Lingang facility in Shanghai. The test, run by the China Academy of Information and Communications Technology (CAICT) at the 2026 China Computing Power Conference in Langfang, is one early marker of where the industry's efficiency story is heading: away from facility overhead and toward intelligence per joule.
For more than a decade, the dominant efficiency score for data centers has been PUE, or power usage effectiveness, the ratio of total facility power to the power that reaches the servers. PUE measures cooling and power-conversion waste. It never said anything about the useful work the electricity actually bought.
That blind spot is starting to matter. AI training and inference have turned electricity from a corporate cost line into a binding public-interest constraint, and CAICT's new "compute-electricity coordination" test tries to grade something different from facility overhead: how well AI workloads track the grid, store energy, and shift in response to price and demand signals. The scoreboard some operators are now adopting, tokens per watt (or useful AI output per unit of energy), reframes efficiency as productivity rather than thrift.
Combined with the 5A "compute efficiency" grade Lingang earned in 2024, also a CAICT first per Science Net, the Shanghai site is now the country's only dual-5A AI data center. The new compute-electricity coordination test itself grades a site across three axes (economics, reliability, and operations) and 11 measurements that include storage coordination, electricity-cost optimization, load shifting, demand response, safety, and platform operations, according to the QbitAI-syndicated SenseTime release.
The storage backbone was disclosed in late 2025: a 17.888 MW / 35.776 MWh battery system built with CATL, with millisecond response, per Sina Finance and Zhidx. SenseTime's load-forecasting model runs at 96% accuracy by the company's own count, with electricity-trading bid deviations inside ±2%. The company says the combined platform, an in-house eight-layer data and five-decision system linking energy forecasting, battery management, distribution monitoring, HVAC, and power-trading interfaces, gives Lingang 25 MW of demand-response capacity and the top two demand-response scores among Shanghai data centers during the 2026 summer peak.
The economics are the more interesting claims, and they are SenseTime's claims, not audited figures. The company projects a 2026 PUE of 1.24 against China's 1.25 threshold for new large data centers, an 80% lift in tokens produced per unit of electricity, an average power cost roughly 10% below comparable Shanghai sites, and about 24,000 tons of annual carbon abatement per 10,000 P of compute. The QbitAI piece is a sponsored SenseTime release republished by the outlet; CAICT certifies that the compute and power systems are coordinated enough to earn 5A, not the unit economics behind the 80% number.
The structural shift, by contrast, is independently verifiable. CAICT's 2025 Compute and Power Coordinated Development white paper and the Open Data Center Committee's coverage of the new standards track a national policy shift: the question is no longer how lean a data center is, but how productive per joule and how grid-friendly. Whether the 80% productivity lift survives outside SenseTime's telemetry is the question that will determine whether the 5A test becomes a national benchmark or stays a single-company showcase.
Watch item: the second site to clear 5A under CAICT's compute-electricity coordination test, expected at the next China Computing Power Conference, will be the first data point on whether the framework is scaling.