China's plan to make computing power as easy as plugging into the grid

There’s a quiet tension inside the 2026 World AI Conference in Shanghai this week. Walk past the 1,100 exhibitors showing off 3,486 products — 349 of them global debuts, the organizers claim — and you’ll see the usual spectacle: robots serving tea, autonomous driving demos, the whole polished show. But in a side forum on Tuesday, Gao Wen, director of the Peng Cheng Laboratory and an academician at the Chinese Academy of Engineering, laid out something more honest: the long, grinding reality of building national-scale computing infrastructure.

Gao described the China Computing Network in three phases. Phase 1, the aggregation stage, is about 70 percent done. Roughly 70 percent of the country’s computing resources have been folded into a real-time unified database that the National Data Administration now uses to monitor the utilization and status of data centers across the country. As of March 2026, China’s total intelligent computing power reached 1,882 EFlops — exaflops, or a quintillion floating-point operations per second — across 1.445 million standard server racks. More than 70 dedicated computing corridors have been built. In April of this year, the computing network was formally designated one of the “six networks” in China’s strategic infrastructure portfolio, alongside power, transportation, and telecommunications.

Phase 2 is where it gets trickier. Once the aggregation layer is solid, the network will open up collaborative computing to specific user groups — think internal deployments within large companies, or partnerships between friendly vendors. It won’t be open to the general public yet. Gao described this phase as analogous to smart grid dispatching: the system will route compute tasks to the most cost-effective or geographically optimal data center automatically, just like the power grid balances load. He estimated this will take about five more years to mature.

The real prize is Phase 3 — the “computing tasks can be collaborative” stage. The vision is that a user anywhere in China could fire off a computation and the network would transparently pull from whatever resources are available — cross-province, cross-operator, cross-chip-architecture — without the user ever knowing. Gao was blunt about the timeline: this isn’t a three-to-five-year thing. It’s going to take “ten, maybe twenty years.”

The comparison he keeps returning to is electricity. The goal, he said, is to make computing as convenient as plugging an appliance into a wall socket. You don’t think about which power plant generated the electrons, which grid segment routed them, or what voltage step-down happened along the way. The computing network should work the same way. Getting there means nailing not just the technical architecture — interoperability between different chipmakers’ accelerators, consistent billing and metering standards, reliability guarantees — but also the institutional layer: who pays whom, who is liable when a computation fails mid-stream, and how security boundaries work when your data is being processed on hardware you don’t control.

During the conference, Gao’s team also ran a parallel forum on turning data centers from energy hogs into “grid-friendly nodes,” with a focus on real-time coordination between computing dispatch and power grid scheduling. The idea: if you can time-shift compute loads to when renewable energy is abundant, you solve two problems at once.

A separate announcement at the same forum — the “Star Pivot” (星枢) space computing constellation’s maiden launch — suggests the computing network ambition extends beyond terrestrial infrastructure. The project, billed as a Shanghai-anchored space computing industry benchmark, put its first satellite batch into orbit.

For now, though, Phase 1 grinds forward at 70 percent. The database exists. The national policy framework is locked in. But the distance between 70 percent and the seamless utility vision Gao described is measured in decades, not quarters.