TL;DR · 30-second read

The Short Version

Artificial intelligence computers draw electricity in sudden bursts, which can jolt the power grid the way a heavy factory switching on and off does. At a conference in Shanghai, Huawei showed gear meant to let large AI computing sites soak up those swings with batteries, and even help steady the grid instead of straining it.

It also pushed a different scorecard. Instead of measuring how much power a building wastes, measure how much useful artificial intelligence work each unit of electricity buys. The Chinese firm SenseTime says it has improved that figure by 80 percent.

Huawei used the HUAWEI AIDC Facility Summit, held alongside HUAWEI CONNECT 2026 in Shanghai on September 20, 2026, to launch what it calls a grid-interactive AIDC solution — a package of power and cooling equipment for AI data centers built around what the company labels “3+1” innovations. The announced version 1.0 combines a grid-friendly uninterruptible power supply, intelligent lithium batteries and grid-forming energy storage, alongside an AI-managed liquid cooling system that predicts the condition of its working fluids to schedule maintenance before failures. In the power domain, Huawei also outlined a longer-horizon “MIMO” power supply architecture.

Hou Jinlong, Director of the Board of Huawei and President of Huawei Digital Power, opened the summit with four design arguments: pairing new power architecture with liquid cooling, using multi-layer hybrid storage to absorb local power fluctuations, coupling electricity and compute so data centers stop behaving as rigid loads, and shifting to modular, prefabricated delivery to cut construction lead times. Speakers from VNET Group and SenseTime shared the stage; SenseTime’s Lin Hai said end-to-end optimization had improved SenseCore’s tokens per watt by 80%.

Executive Summary

The substance of the announcement is architectural rather than financial. Huawei is proposing that an AI data center should be designed as a participant in the power system — able to hold its own voltage and frequency behavior steady through grid-forming storage, and to buffer the second-to-second swings that large training and inference clusters impose — rather than as a fixed block of demand that a utility must simply serve.

That framing matters because grid connection, not construction, is increasingly what sets the calendar for AI capacity. A site that arrives at the utility as a controllable, power-quality-friendly asset is a different negotiation from one that arrives as raw megawatts of volatile load. Huawei’s pitch, delivered by Digital Power Vice President Bob He, is that as renewables raise the share of power-electronic sources on the grid, compute and energy infrastructure can no longer be planned independently.

The second half of the pitch is a measurement change. Huawei and SenseTime both argued for scoring AI facilities on tokens per watt — useful model output per unit of electricity — instead of power usage effectiveness, the long-standing ratio of total facility power to IT power. That is a bigger change than it sounds, because it moves the boundary of the efficiency calculation outward to the grid and inward to the model at the same time.

Why a Data Center Would Want to Look Like a Grid Asset

A conventional data center is, from a utility’s point of view, a well-behaved but inflexible customer: it wants a large, constant, highly reliable supply and offers nothing back. AI clusters break the “well-behaved” half. Synchronized training jobs can ramp thousands of accelerators together, producing what the release describes as rapid and severe power fluctuations, and Huawei’s speakers named the consequences utilities worry about — disconnection events and unstable wideband oscillation, meaning sustained electrical ringing across a range of frequencies that neither the site nor the grid damps out.

The engineering answer Huawei describes is layered storage sitting between the workload and the grid. Batteries and grid-forming inverters absorb the swing locally, so the fluctuation never reaches the interconnection point; grid-forming units go further by actively setting voltage and frequency rather than following the grid, which is what allows a site to support a network with a high share of inverter-based renewable generation instead of leaning on it. VNET Group Founder and Chairman Josh Chen extended the idea to campus scale, describing direct green power connections, active microgrids and direct-current distribution inside buildings as the way to serve gigawatt-class AI campuses.

The commercial logic follows from scarcity. When interconnection capacity is the binding constraint on AI buildout, the cheapest new megawatt is often the one a developer does not have to request. Equipment that lets a site ride through disturbances, shave its own peaks and behave predictably at the meter is, in effect, a substitute for grid capacity that has not been built yet — and a bargaining chip in queue discussions. Huawei has not published evidence on how much connection time or capacity this saves in practice, which is the number that would turn the argument into a business case.

Tokens Per Watt Redraws the Boundary of the Efficiency Question

Power usage effectiveness has governed data center efficiency for two decades. It is a facility-boundary ratio: total power into the building divided by power reaching the IT equipment. It says nothing about whether the servers do anything useful. Once a site is spending most of its energy on inference, that blind spot becomes the main event — a facility can post an excellent PUE while running inefficient hardware on badly scheduled work.

Tokens per watt closes that gap by measuring output rather than overhead, and this is where the 80% figure does its work. SenseTime’s Lin Hai described SenseCore’s improvement as end-to-end, spanning power plants and grids, rack services in the data center, compute cloud services and model services. That is the whole chain, not the mechanical plant. A number of that size is therefore not evidence that any single piece of equipment got 80% better; it is evidence that when you draw the measurement boundary that wide, most of the available gain sits in layers the facility team has never been scored on — and, at the far end, in the grid supplying the site. Huawei’s Hou Jinlong put the same point commercially: maximizing tokens per watt and minimizing cost per token will be the core competitiveness of what he called compute factories.

Who this affects is concrete. Colocation operators sell space, power and a PUE commitment; they do not control the model, the scheduler or the chip, so they cannot be held to a tokens-per-watt number without a very different contract. Chip vendors and inference-software teams gain, because their contribution finally appears in the facility scorecard. Utilities gain a vocabulary for load that varies in usefulness as well as size. The catch is standardization: tokens differ by tokenizer, model and precision, so an 80% improvement is a credible internal engineering result and not yet a figure that can be compared across operators. No published methodology accompanied the claim.

The Delivery Clock and the Question of Who Can Buy It

Hou’s fourth argument — modular, prefabricated, product-based delivery to shorten lead time — is the least glamorous and arguably the most operationally consequential. In a market where the constraint is time to first megawatt, converting site-built electrical and cooling systems into factory-built blocks shifts the schedule risk from a construction crew to a production line. It also suits Huawei’s structural position: the company is selling the power chain and the cooling chain and the compute stack, and Bob He explicitly named end-to-end energy plus full-stack AI as the differentiator.

That integration is genuinely hard for point-product vendors to match, and it is also the strongest argument against buying it. A grid-interactive architecture in which the uninterruptible power supply, the batteries, the grid-forming storage, the liquid cooling controller and the workload scheduler are co-designed is, by construction, difficult to unbundle later. Buyers evaluating it should ask what the interfaces are and whether any layer can be swapped — questions the launch materials do not address.

Market context also shapes who this is for. The named practitioners on stage — VNET and SenseTime — are China-based, and the summit sat inside a conference where Huawei also presented an upgraded Stellar AI Network solution and released its Intelligent World 2035 report series setting out ten key directions. The ideas here travel further than the product does: measuring output per watt, buffering workload volatility on site and offering grid services are design responses any operator facing a long interconnection queue will recognize, whatever equipment ends up in the building.

Background

Huawei Digital Power is the Huawei division that sells power conversion, energy storage, solar inverter and data center power and cooling systems — a business adjacent to, but distinct from, the company’s networking and computing lines. Its data center pitch has increasingly merged with Huawei’s AI compute strategy, on the argument that the electricity system and the compute system are now a single engineering problem. HUAWEI CONNECT is the company’s annual flagship conference; the 2026 edition in Shanghai also produced an upgraded Stellar AI Network solution and a set of Intelligent World 2035 research reports laying out ten key directions.

The wider context is a buildout constrained by electricity rather than silicon supply alone. Racks designed for AI accelerators draw many times what general-purpose servers did, pushing operators to liquid cooling and to far larger grid connections, while utilities in many markets face multi-year queues to study and approve new large loads. That has pushed the industry toward on-site generation, storage and demand flexibility, and toward efficiency metrics that account for useful work rather than facility overhead. VNET Group is a Chinese carrier-neutral data center operator; SenseTime is a Chinese AI company whose SenseCore platform runs its own intelligent computing centers.

Sources

Source: Huawei Grid-Interactive AIDC Solution: Shaping the New Paradigm of AI Infrastructure — Huawei’s September 20, 2026 announcement from the HUAWEI AIDC Facility Summit in Shanghai, detailing the grid-interactive AIDC 1.0 solution and remarks from Huawei, VNET Group and SenseTime executives.

Primary sources: Huawei представляет инновационное решение в области иммерсивного телеприсутствия, которое предлагает эффект погружения и интеллектуальные услуги (Huawei’s September 20, 2026 immersive telepresence announcement from the same conference); Huawei представляет обновленную сеть Stellar AI для обеспечения безопасного и интеллектуального подключения к агентскому миру (the upgraded Stellar AI Network solution presented at HUAWEI CONNECT 2026 on September 18, 2026); В последнем отчете Huawei предлагаются 10 ключевых направлений Интеллектуального мира-2035 (Huawei’s September 18, 2026 launch of its Intelligent World 2035 report series).