Cold Plains Light Engines Accelerated Compute Buildout In Northern Provinces

Orbital views gave me my first shock. More than one thousand facilities stand counted across the country, installed power has crossed 30GW, and another 20GW waits in preparation. With the United States heading toward 60GW by year end, the picture marks the second largest artificial intelligence estate in the world. The contrast sharpens once Europe with the Middle East and Africa is placed near 15GW in total. I read it this way: training an artificial intelligence model is no longer a single building job, it resembles wiring a city grid. I leave the figures to the table below and I continue the comment from there.

The pace impresses as much as the scope. A 100MW class campus can finish in under a year. I found the secret in kitchen cabinet logic: power and cooling assemblies leave the factory as finished blocks, arrive ready, and join on site. Ground preparation, money flow, and grid hookup advance at the same time. Local authorities line up plots, funding, and grid access in advance. A contest between provinces adds fuel: once one region announces an artificial intelligence campus, the neighbor offers cheaper land and power to pull the capital. Against projects that wait long for permits elsewhere, this tempo stands apart.

The answer to placement lies in splitting east from west. Crowded hubs such as Beijing, Shanghai, and Shenzhen generate data yet land stays tight and power stays dear. So heavy model training travels west while quick response work remains near users. Eight national compute hubs and ten clusters form the skeleton of the plan. I compare it with a home: noisy bulky jobs live in the cool basement while serving happens in the front kitchen. Fast links join both sides. The split eases land pressure and power bills at once.

The city of Ulanqab shows the plan in action. The city alone signed 19 projects totaling 12.5GW, more than ten times the operating load of the prior year. According to SCMP , ByteDance negotiates 5 to 6GW of fresh power there, while Huawei and Alibaba expand in the same basin. More than 30 facilities stand under construction inside a three kilometer radius near Hohhot. The Wired file on Ulanqab confirms the wave: nearly one hundred facilities opened or broke ground since 2016. The reason is plain: average air near 4 degrees cheapens cooling, wind and solar run rich, and distance to Beijing keeps network delay low. Energy bills in Ulanqab run near half the Beijing level, a vast gap for artificial intelligence models that run thousands of parts around the clock.

The wave reaches beyond one city. New groups rise near Zhongwei in Ningxia by the desert edge and around Yiwu in Xinjiang further west. The user table surprised me most: five firms take about sixty percent of fresh capacity. ByteDance , Alibaba, Tencent, Baidu, and Huawei absorb the bulk. The firm behind TikTok and CapCut beside the firm behind WeChat makes the pattern clear: recommendation systems, generated video, agent services, and cloud workloads hunger for compute without pause. The lesson I draw: the user base that feeds models each day drives this build as much as model count.

Now we step inside the rack. The building block of the Huawei line is Ascend, a part that carries model training, token creation, and answer generation. Eight parts meet in one module, eight modules enter one rack, and each rack holds 64 parts. Sixteen racks join into an Atlas 950 group of 1024 parts, while larger layouts walk toward 8000 parts. The Huawei team reports groups of 256000 parts in deployment with a road map toward one million parts. I tell it as an apartment story: the module is a flat, the rack is a floor, the group is a block, and the super group is a district. Fine flats help little when stairs and lifts fall short, and here links between parts are those stairs.

The light link section excited me most. Moving data over copper grows hard and power hungry as rates climb. Huawei places the Hi-ONE optical engine beside parts: data walks a short electric path, turns into light, and travels over fiber. A single engine carries 7.2 Tbit/s. The shift saves over 550kW per group, more than half a megawatt returned to model work. Beside it stands UnifiedBus, a shared tongue for parts, memory, and storage, with 4096 parts under unified memory address in one group. The logic is simple: thousands of fast parts idling for data make a slow system. It resembles luxury cars stuck at a crossing with broken signals.

Software matters as much as hardware. Nvidia holds a strong position through CUDA, a developer platform with ready function sets, tools, and years of accumulated practice. Building a rival part is only the first move; the real test is whether existing writing runs fast there. TileLang from DeepSeek with Huawei answers that test: one writing aims to run on both Nvidia and Huawei lines, shared as open source over GitHub . DeepSeek frames the path as a domestic option against lock in. CUDA still counts, yet it no longer stands as the sole trail. My shorthand is a charger story: once a single brand cable ruled, now one cable fits two sockets.

Shells stand with buyers waiting, so where are the parts to fill them. That question leads to the fab. SMIC lines reported loading near 95 percent in the second quarter of 2026, with expansions in Beijing, Shanghai, Shenzhen, and Tianjin. Reuters confirms the loading and the hunt for fresh tools. With leading ASML EUV tools restricted, output advances on DUV gear with multi pass patterning; picture a fine drawing traced in several aligned passes, where each added pass adds time, cost, and fault risk. Memory follows the same arc: CXMT of Hefei grew in phone and server memory and now chases stacked memory to feed parts. HuaweiCentral records show Huawei announcing HiBL 1.0 and HiZQ 2.0 designs, with 128GB and 1.6TB/s discussed on the Ascend line. Unless part output and memory output grow together, one bottleneck merely hands over to the next.

Let me close. Round table estimates place global fill near 2029 and domestic fill near 2030; I read both as estimates, not pledges, since demand could grow faster. The near term outcome looks like cheaper artificial intelligence tools with wider model and agent choice. The reverse side is attention and data: recommendation systems will race for focus while custody of data gains weight. My plain lesson: we type a short order while factories, grids, and thousands of parts labor behind it. Once concrete meets silicon in one sentence, the artificial intelligence debate changes shape. Track the figures in the table above and the records at the end.