Data centers used to exist solely for enterprise applications and the cloud. Today, they are increasingly used to produce tokens, the unit in which AI output is generated and billed. And the numbers suggest a “token factory” era is already here. According to Huawei’s He Bo, Vice President of Huawei Digital Power, daily token usage of ByteDance’s Doubao model alone grew roughly 1,500-fold in the year to June.
This has consequences, because token factories are a very different kind of load. The power draw of an AI cluster can fluctuate by up to 200% amid training and inference runs, says He, unlike a typical data center that draws a steady amount of power around the clock.
Stable power is crucial
And such swings are happening more often, as token processing volumes surge. In China alone, volume grew roughly 2,500-fold in the last two years, according to Huawei’s figures. Overall electricity consumption is climbing steeply too, with the International Energy Agency (IEA) expecting global data center electricity consumption to double from 2024 levels by 2030. Yet securing enough power is only half the problem.
The other half is whether the grid can tolerate how that power is used. The fluctuations happen on several fronts simultaneously, says He. At the chip level, power draw changes within milliseconds, while swings in the data hall happen over minutes. Zoom out further, and consumption can rise and fall over hours as new jobs get scheduled. The problem? No single piece of equipment is designed to handle all of that.
The power grids themselves are also changing. As renewable energy displaces conventional power generation across Asia Pacific, grids are becoming electrically weaker, with lower short-circuit ratios that make them more prone to voltage swings and oscillation, says He. And as facilities scale into the gigawatts, grids are now far less able to absorb disturbances than they once were.
Getting enough power is just the start. Power densities inside the building keep rising with each iteration of AI hardware, and the heat has to be removed just as quickly. Crucially, the data center can no longer be a passive load. It is no longer enough to simply ride through voltage and frequency events. In some cases, the facility will be expected to play a part to help the grid recover from them.
From the grid to the chip
The fix, according to He, lies in looking at the whole facility rather than its parts. A more efficient UPS or a better CDU helps, but the bigger gains come from rethinking the entire electricity pathway through the facility. His framework for this is what Huawei calls[z1.1] its “3+1” Strategy, for watt innovation, heat innovation, and bit innovation, as well as construction[z2.1] innovation using prefabrication.
Take power first. The grid to chip chain must get shorter and lose fewer watts along the way. Yet that chain is getting more complicated. Utility power, renewables, batteries and backup generators all feed in, while AC and several flavors of DC will coexist on the server side for years. Huawei’s answer is an architecture it calls MIMO, built around a grid-forming energy router that takes in multiple energy inputs and turns them into stable power. This allows a single power foundation to serve every output type.
The same thinking applies to energy storage, which should be layered according to how fast the swings are. Capacitors on the board absorb the fastest ones, supercapacitors and short-duration lithium batteries handle the data hall, and grid-forming battery systems smooth things out at the campus level, so each layer hands off to the next.
Cooling must keep up with density, too. Liquid cooling has gone from an option to a necessity. But at above 100kW per rack, thermal inertia is so low that a cooling failure may leave operators about a minute to respond, says He. The points of failure multiply as well, since the coolant now runs through the server. His fix is AI-driven thermal management that flags minor leaks and declining coolant quality before they become faults, so that operations shift from reacting to problems to predicting them.
Grid strength will decide who builds first
All of which raises the question of whether the current wave of data center construction across Asia Pacific is sustainable, or a temporary bubble. He is firmly in the first camp. Token demand is growing exponentially, and gigawatt-scale AI data centers are moving from a small handful of projects to routine deployments. In his view, this is a long investment cycle driven by a structural shift in computing.
But it will not unfold evenly. Access to power, grid connection capacity, renewable energy availability and construction speed differ widely across markets. And those factors will ultimately decide how fast each market moves. In other words, grid conditions do not just determine how stable a facility is once built. They determine when it gets built at all; some markets will see protracted development over years while others move far quicker.
Huawei’s pitch is breadth. On the energy side, it builds everything from generation and grid equipment to storage, power supply and thermal management. It also makes the chips, the models and the applications that run on them. That lets the company design and simulate the whole system together, He argues, instead of assembling parts from different vendors.
“We are not simply combining UPS systems, lithium batteries, and liquid cooling equipment,” he summed up. In the token factory, power is where the design starts, because the ultimate limit of compute is electricity.
Visit Huawei’s website to learn more about its solutions here.