Product1 publisher3 min readPublished
Huawei sells AI data centre power gear on a promise to steady the grid
The grid-interactive stack unveiled in Shenzhen pairs a 1,280 kVA UPS cabinet and grid-forming storage with a thermal unit built for Ascend servers. Huawei wants buyers judging it on tokens per watt in place of PUE.
The Product Desk · Product desk
What happened
- Zhou Jianjun unveiled Huawei's grid-interactive AI data centre solution in Shenzhen, and the same kit went on stage days later at a Huawei facility summit in Shanghai held alongside Connect 2026.
- The launch included the Taishan UPS, which Huawei calls the industry's first megawatt-class single-cabinet uninterruptible power supply, rated at 1,280 kVA with up to 98% dual-conversion efficiency.
- VNET Group founder and chairman Josh Chen said at the summit that high-frequency changes in AI workloads can cause grid disconnection and unstable wideband oscillation.
- Huawei says both of its main AIDC power supply series passed a grid wideband impedance field test, and claims to be the first vendor through wideband oscillation damping testing at the Shanghai Electrical Apparatus Research Institute.
- SenseTime's Lin Hai said the company's Lingang facility was the first to receive a compute-power synergy certification from China's state telecoms research academy.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- decision Because the thermal unit is built to fit Ascend servers, the cooling and power decision now gets made in the same room as the accelerator decision, and a buyer choosing other silicon gives up the integration Huawei uses to justify the whole stack.
- constraint Swapping PUE for tokens per watt moves the efficiency target onto model output, so the facilities team gets measured on a number the model and cloud teams largely control.
- exposure An operator waiting on a gigawatt connection now has to argue about how its campus behaves electrically as well as how many megawatts it wants, and in Texas it makes that argument behind 474GW of unverifiable requests.
- precedent If tokens per watt becomes the figure customers publish, comparisons get harder to check, because the number spans power plants through models and a vendor can claim the gain without saying which layer produced it.
Huawei says its new thermal management unit is natively compatible with its Ascend servers [9], and the company was selling chips, pods and interconnect at the same conference [10]. The unit does more than move heat. It uses AI to detect micro leaks and predict the health of the working fluid, and it coordinates cooling with electrical load so the system reacts as compute demand changes [11]. A buyer whose accelerators are not Ascend loses that native fit [9].
Underneath the boxes is what Huawei calls a MIMO power supply architecture, which takes multiple energy sources in and serves multiple compute types out, adjusting for AI load swings across timescales from microseconds to hours [12]. The LUNA2000 grid-forming system handles weak-grid support at campus level [13]. SmartLi 5.0 raises storage density and is built for a move to 800V HVDC [14]. The Hengshan DC UPS, which Huawei describes as the first software-defined DC UPS, supports 270V, 400V and 800V on one hardware platform, because the voltage for the next generation of racks is still unsettled [15].
Then there is the metric. PUE compares total facility power with the power that reaches the IT kit, a ratio that leaves out whether that kit produced anything useful [16]. Lin Hai, who runs SenseTime's intelligent computing centre business, said the company has improved its tokens-per-watt figure by 80% [17]. At 1.8 times the tokens for the same power, the same token volume needs about 56% of the electricity [1]. Lin credited end-to-end optimisation across power plants, grids, racks, cloud services and models [18]. Most of that list sits outside the power room. Huawei did not publish prices for the launched hardware, and it gave no baseline or measurement method for the figure [24].
The sales argument lands on the interconnection conversation. Grid connection queues are now the binding constraint on data centre buildout in Europe, and Texas alone is sitting on 474GW of unverifiable requests [19][20]. "Today, in the AI era, the startup, shutdown, and load fluctuations of a large AIDC campus can all impact grid operations," said Zhou Jianjun, who runs global marketing, sales and services at Huawei Digital Power [2][1]. Zhou also said that as the renewables share rises, local grid strength falls and volatility climbs [6].
Whether the premium is worth paying turns on the grid operator and on the silicon. Does the local operator credit grid-forming behaviour with anything a finance team can book: a tariff, queue priority, a faster study? Where the answer is no, the hybrid energy storage that Hou Jinlong, president of Huawei Digital Power, described as absorbing local fluctuation [23] is insurance the buyer funds alone. Then the silicon: is the compute Ascend? The integration is what separates this stack from an assembly of parts [9].
One saving holds either way. PowerPOD 5.0 cuts footprint by 40% and shortens lead time from seven days to three [21], four days off each pod deployed, or about 57% [2].
What to watch
- Whether any grid operator outside China offers a tariff or queue priority for campuses running grid-forming storage.
- Whether the compute-power synergy certification publishes a measurement method for tokens per watt, and who audits the claim.
- Whether next-generation racks settle on 800V and leave the Hengshan DC UPS's 270V and 400V support unused.