Huawei unveiled a network-interactive AI data center solution at HUAWEI CONNECT 2026 in Shanghai on September 21, introducing grid-integrated infrastructure designed to maximize tokens per watt as the core metric of computing factory competitiveness, according to a PR Newswire release from the company.
Key points
- Huawei’s AIDC 1.0 integrates grid-friendly UPS, intelligent lithium batteries, grid-forming energy storage and AI-driven liquid cooling to adapt to and support the power grid while smoothing AI workload fluctuations
- Hou Jinlong, Director of the Board and President of Huawei Digital Power, identified maximizing tokens per watt and minimizing cost per token as the basis of computing factory competitiveness
- SenseTime’s SenseCore improved tokens per watt by 80% through end-to-end optimization from power plants through grid, IDC rack services, cloud and model services
- VNET Group’s Josh Chen described redefining energy sources, campuses and buildings for mutual support to enable GW-scale AIDC reliable power
- The solution marks Huawei’s expansion beyond Ascend AI chips into full-stack AI infrastructure including power, cooling and grid integration
Grid-interactive architecture targets power-quality challenges
Bob He, Vice President of Huawei Digital Power, stated that as the share of renewable energy grows, the electricity grid increasingly exhibits power-electronics characteristics. Combined with rapid and strong fluctuations in AI workloads, compute-power and energy infrastructure are no longer separate but tightly integrated. Huawei’s “3+1” innovation approach for the grid-interactive AIDC 1.0 solution uses a future-oriented MIMO power architecture. On the power side, grid-friendly UPS systems, intelligent lithium batteries and grid-forming energy storage systems enable AIDCs to adapt to and support the grid while flattening AI workload fluctuations. On the thermal side, an AI-driven liquid cooling system predicts coolant conditions and enables predictive maintenance.
SenseTime reports 80% TPW gain via end-to-end chain
Lin Hai, General Manager of SenseCore Business Group at SenseTime, detailed the evolution of intelligent computing infrastructure efficiency from Power Usage Effectiveness (PUE) to tokens per watt (TPW). The TPW metric links power consumption directly to token output and covers the full chain from power plants and grids through IDC rack services, cloud computing services and model services. Through end-to-end optimization, SenseCore improved TPW by 80%, according to the company’s presentation at the summit.
VNET outlines GW-scale campus redesign for green energy direct access
Josh Chen, Founder and Chairman of VNET Group, said energy sources, campuses and buildings can be redefined to enable mutual support between energy sources, campus planning and DC power supply for buildings. This creates a new energy system oriented toward direct connection to green energy and active microgrids that guarantees reliable and stable power supply for GW-scale AIDCs. Traditional data centers have limited interaction with the grid, but AI workloads’ sensitivity to frequent changes can lead to grid disconnection and unstable broadband oscillations.
Huawei positions power as infrastructure foundation for integrated compute-network-storage
Xia Qin, Chief Expert in Huawei Computing Strategy Planning, noted AIDC is evolving toward an integrated, coordinated architecture for compute, network and storage with energy as the infrastructure foundation. A highly reliable, energy-efficient and rapidly deployable power system enables hyperscale AI compute clusters to operate reliably and lays a solid foundation for large-scale training and inference in the era of agentic AI. The announcement comes as Chinese technology firms build sovereign AI data center capacity amid U.S. export controls on advanced chips.
Huawei will present its one-stop AI inference solution alongside the network-interactive AIDC at HUAWEI CONNECT 2026, aiming to create a highly reliable, energy-efficient and rapidly deployable compute foundation for AI cluster infrastructure.
