Staff Reporter |
Editor | Wen Shuqi
In the past, internet giants mostly chose to build their own data centers to meet their ever-growing computing power needs. But the explosion in AI demand is turning that model on its head.
Recently, our reporters visited Ulanqab and found that this city, which once attracted internet giants to set up data centers thanks to its cool climate and cheap power, is now in the middle of a fresh wave of expansion. Over in the Chahar High-tech Industrial Development Zone, multiple data center projects are under construction. But the builders aren’t Alibaba or Tencent anymore – they’re third-party operators like CICC Data, 21Vianet, and Envision Energy.
These new data centers aren’t being built for the operators’ own use. They’re being leased out to the very giants that used to build their own – Alibaba, ByteDance, Tencent, and the like. One CICC Data facility we checked out was snapped up by Alibaba Cloud the moment construction kicked off, and it’s already been delivered. Another data center built by Envision has already inked lease deals with several top-tier internet companies, though it hasn’t been handed over just yet.
An Alibaba Cloud technician told us that the company used to rely primarily on self-built data centers. But that strategy has shifted. Going forward, new data centers will be mostly leased, not built.
This isn’t just a domestic trend. U.S. tech heavyweights like Meta, Microsoft, Amazon, and Oracle are also locking in future data center capacity through long-term leases. Meta’s Q1 disclosures this year showed its lease obligations for data centers and related infrastructure had ballooned to roughly $183 billion.
What’s driving this shift is the explosive growth in AI-fueled compute demand. McKinsey projected in a June report that global data center demand will nearly double from about 82GW in 2025 to around 220GW by 2030. AI-related demand alone is expected to surge from about 44GW to 155GW – a 3.5-fold increase.
So why are tech and internet giants pivoting from self-building to leasing in the AI era? It all comes down to the economics.
Big Tech Starts Running the Numbers
AI isn’t just boosting demand for compute; it’s dramatically scaling up the size of individual data center projects.
An insider at a top cloud computing firm told us that the industry has gone through four generations of data centers, from enterprise self-built EDCs to shared colocation IDCs, then to hyperscale cloud CDCs, and now to AI-specific AIDCs. The scale of an AIDC is anywhere from ten to dozens of times larger than a traditional data center.
According to our sources, just a few years ago, the data centers that internet giants were building were mostly in the hundreds of megawatts. But once AI demand took off, they started needing GW-scale facilities.
Investment research firm Bernstein estimates that building and equipping a single 1GW AI computing center requires total capital expenditure (Capex) of around $35 billion.
And based on actual GW-scale data center construction in China over the past two years, the upfront one-time investment typically exceeds 20 billion yuan, according to the cloud computing insider.
The upfront costs that big tech has to shoulder have skyrocketed, forcing a rethink of the economics of self-building. Li Jun, deputy general manager of Envision Energy’s AI Computing Center division, told us that GW-level AI computing centers are now heavy-asset investments. There’s no reason for big tech to tie up all its capital in infrastructure. “By signing long-term leases, these giants can channel more capital into chips, models, and their AI businesses, while letting specialized operators take on and manage the infrastructure assets,” he said.
Time is another critical factor in the build-versus-lease equation. The cloud insider revealed that traditional data center construction takes around 12 months. But for large facilities that require new substations and other power infrastructure, the energy build-out alone can take three to five years. Some projects might need several years just from acquiring land to getting the power infrastructure in place. “In a compute crunch, the traditional self-build model simply can’t keep up with how fast demand changes,” he added.
Multiple industry insiders have pointed out a clear mismatch between the speed of AI compute expansion and the cycle of traditional power grid construction. Facing rapidly growing AI demand, top-tier giants want data centers delivered in about six months from project approval. The main bottlenecks are substations, switch stations, and other grid connection points.
“Big tech is increasingly drawn to parks that already have ‘source-grid-load-storage’ capabilities – it solves the green power question on one hand, and the overall power supply reliability on the other,” one industry insider told us. A giant building alone rarely has the ability to secure power resources of the same scale and at the same low cost. But third-party AIDC operators have been planting wind and solar farms in Inner Mongolia, Ningxia, and other regions for years. Wherever there’s favorable policy and contiguous land, they’ve already staked their claim.
Sitting in central Inner Mongolia, Ulanqab has become one of the key battlegrounds for third-party AIDC operators. With an altitude of 1,500 meters, an annual average temperature of 4.3°C, and electricity prices between 0.32 and 0.35 yuan per kWh, running the same computing system in Ulanqab can cut annual electricity bills by nearly half compared to the east.
We’ve learned that the total investment in signed computing power projects in Ulanqab has already surpassed 250 billion yuan, with third-party AIDC operators like CICC and 21Vianet each pouring billions into multiple projects.
It’s safe to say the data center leasing business will keep growing steadily over the next few years.
Compute Crunch Forces Faster Data Center Delivery
During our visits, we noticed that the AIDC leased by Alibaba Cloud was delivered by operator CICC Data in just 100 days – a 50% to 75% reduction compared to the previous average timeline.
Behind this leap in delivery efficiency is a complete overhaul of how data centers are built. Wang Chaoyang, general manager of Alibaba Cloud’s global data centers, told us that modularization has supercharged construction. If the old way of building a data center was like a traditional construction project, then after modularization, it’s more like assembling a product.
Wang compares the shift to Apple’s OEM manufacturing model. Power supply, cooling, security, smart systems, and fire protection are all standardized and versioned into modules, with modularization rates exceeding 90%. Instead of assembling everything on site, these modules are pre-assembled in factories. The construction site no longer needs thousands of workers – just a crane to lift the containerized modules into place and some wiring and debugging before going live.
This modular delivery approach also signals that Alibaba Cloud is gaining more leverage in data center leasing deals. According to our sources, in AIDCs leased by big tech, the vast majority of architectural design is still controlled by the third-party operators – meaning tenants had to accept whatever the operator built. That’s changing now.
Wang revealed that after this modular overhaul, delivery times, costs, and quality all improved simultaneously, with overall construction costs dropping by more than 10%. One big reason: module installation is so much simpler now that regular workers can assemble them after minimal training.
Modularization isn’t just about cutting costs and boosting efficiency – it also makes data centers more resilient in the face of relentless AI compute growth. With GPU iteration speeding up, a traditionally built data center might become obsolete and need retrofitting for next-gen chips almost as soon as it’s finished. But with modular design, data centers can scale up faster and adapt more easily to new chips and compute demands.
Simply put, AI is redefining the data center business. In the past, it was all about resources and scale. In the AI era, the game is about speed, efficiency, and how quickly you can respond to next-generation compute demands.
While third-party AIDC operators haven’t fully embraced modular delivery as an industry standard just yet, there’s no denying that once one data center’s delivery cycle gets dramatically shortened, the pressure is on for everyone else to step up their game.
