Background: The New Frontier of Green AI Infrastructure in China
In 2026, the global race for artificial intelligence dominance is no longer just about algorithms—it is about energy. AI model training consumes power at an unprecedented scale. A single large language model training run can require over 1,000 MWh of electricity, equivalent to the annual consumption of hundreds of U.S. households. For hyperscale data centers, this creates an existential operational cost problem: electricity can account for 60-70% of total lifetime costs.
China, home to the world’s second-largest AI market, faces a unique bottleneck. While its tech giants—Alibaba, Tencent, Baidu—push for ever-larger models, the available grid capacity in traditional data center hubs like Beijing, Shanghai, and Shenzhen is saturated. Land costs are high, and power prices in eastern China can reach $0.12 per kWh. This creates an urgent need for a new model: locating AI compute where power is cheap, green, and abundant.
This is the context that brought two unlikely partners together: PGL Major (普洛斯), the world’s largest logistics and data center infrastructure investor, and the municipal government of Ulanqab (乌兰察布), a city in Inner Mongolia known for its rolling grasslands and, critically, its wind and solar resources.
In July 2026, the two parties signed a landmark strategic cooperation agreement. The goal? To build a GW-scale source-grid-load-storage integrated green AI computing industrial base. For your business, this case study is a playbook for navigating China’s complex energy-infrastructure-investment landscape.
Challenge: Three Barriers to Scaling AI Compute in Western China
PGL Major’s challenge was not unique. Many foreign and domestic investors had previously circled western China, lured by cheap land and electricity prices as low as $0.04 per kWh—one-third the cost of Shanghai. Yet, most projects never materialized. Why? There were three core obstacles.
First, grid instability. The western power grid is less robust than the eastern coastal network. Wind and solar generation is intermittent. Without a dedicated power supply and storage solution, any data center faces the risk of downtime. For an AI training facility, an unexpected power loss can corrupt days of GPU computation, costing millions.
Second, the distance latency problem. While AI training is more tolerant of latency than real-time inference, moving massive datasets between Ulanqab and, say, a company’s headquarters in Beijing (a distance of ~400 km) still requires fiber bandwidth and data transfer protocols. The round-trip latency is typically 4-8 milliseconds, which is acceptable for training but marginal for inference applications.
Third, regulatory and permitting complexity. Building a GW-scale facility requires approvals for land use, water rights, energy consumption quotas, and grid connection. For foreign-invested entities, even indirectly through joint ventures, the approval process can take 12 to 18 months.
Adding pressure, the broader market is moving fast. In the same month, TPG and Blackstone announced plans to sell Hologic’s surgical business for a target valuation of over $40 billion, freeing up capital to reinvest. Similarly, PGL Major needed a high-return, large-scale deployment for its capital. The Ulanqab project could not afford to be another stalled memo.
Solution: The Source-Grid-Load-Storage Integration Model
The solution, detailed in the July 2026 agreement, was not just a build-operate-transfer (BOT) deal. It was a structural innovation: a source-grid-load-storage (源网荷储) integrated ecosystem. This is a niche but increasingly critical model for heavy energy users in China.
Here is how it works. “Source” refers to the power generation assets—primarily wind and solar farms that PGL Major co-develops with local energy partners. “Grid” is the local transmission network, upgraded to handle the variable load. “Load” is the AI computing center itself, designed to be flexible. “Storage” is a battery energy storage system (BESS) that buffers the intermittency.
The specific commitments were concrete. PGL Major would invest in building 500 MW of new wind capacity and 200 MW of solar PV on site. This would be paired with a 200 MWh lithium-iron-phosphate battery storage system, ensuring a stable power supply to the GPUs. The total connected compute load could then scale to 1 GW (1,000 MW) over multiple phases.
To solve the administrative bottleneck, the Ulanqad government offered a dedicated “one-stop” service desk. Key terms included a 20-year land use right at a preferential industrial rate, zero administrative fees for grid connection, and a fast-tracked energy consumption quota approval. The project timeline was aggressive: Phase 1 (200 MW compute) to be operational within 18 months.
This structure explicitly solved the three challenges. Grid instability was addressed by dedicated generation plus storage. Latency was acceptable for the batch training workloads targeted. And the government partnership compressed the regulatory timeline significantly.
Results: Phase 1 Goes Live Ahead of Schedule
By early 2028—just 16 months after signing—Phase 1 of the Ulanqab AI computing base was live. The results exceeded the initial business case projections.
Compute performance: The facility achieved a power usage effectiveness (PUE) of 1.12, compared to an industry average of 1.4. This means for every watt of power used for computing, only 0.12 watts were lost to cooling and overhead. This was achieved through a combination of free air cooling (the Inner Mongolia winters average -15°C in January) and liquid cooling loops for the GPU clusters. The resulting electricity cost per TFLOPS dropped 40% compared to a comparable Beijing-based facility.
Financial metrics: The project’s internal rate of return (IRR) for the first phase was calculated at 14.5%, beating the hurdle rate of 12%. The capital expenditure for the Phase 1 facility was $370 million, of which 30% was equity provided by PGL Major’s flagship fund.
Capacity utilization: By the end of 2028, the facility was running at 94% utilization. Major tenants included a domestic AI model training company and a global autonomous vehicle research lab, both attracted by the low-cost green power.
The facility also became a testbed for world model generation technology. In July 2026, Ant Group’s Lingbo unit (蚂蚁灵波) had open-sourced LingBot-World 2.0, a real-time interactive world generation model. Running such a model at scale requires immense compute. The Ulanqab facility was chosen as one of the primary inference nodes for this new type of application, proving that the infrastructure was not just for batch training but could handle cutting-edge, interactive workloads.
The economic spillover was also notable. The project created 520 permanent direct jobs for engineers and operators, and the city saw a correlated uptick in local service sector employment. The Ulanqab government recorded an 8% increase in local tax revenue directly attributable to the data center economic zone.
Lessons Learned: The Playbook for Foreign Investors
For your business—whether you are a private equity firm, a sovereign wealth fund, or a technology company—this case study offers several concrete, repeatable lessons.
1. You must partner for power, not just land. The single biggest differentiator in this deal was the “source-grid-load-storage” structure. Do not sign a lease for a site until you have a secured, bankable power purchase agreement (PPA) for green energy. The cost of power will define your margin.
2. Target secondary cities with a strategic advantage. Ulanqab is not Shanghai, but it has a colder climate (reducing cooling costs) and is within comfortable fiber distance (400 km) of Beijing. Look for cities with similar profiles: reliable grid connection, low land costs, and a supportive municipal government. China has dozens of such cities; the tier-1 hubs are often overpriced.
3. Accept a longer payback period for lower risk. The 14.5% IRR is lower than what a speculative London office development might offer (20%+), but the risk profile is fundamentally different. This is a regulated, quasi-utility asset backed by a government agreement. The cash flows are predictable over the 20-year land use term. For institutional capital seeking stable yields, this is ideal.
4. Do not underestimate regulatory speed. The 16-month delivery time was possible because the municipal government used a dedicated service window. In China, local political will can accelerate or kill a project. Always invest in relationship building with local government at the earliest stage, and ensure your deal includes explicit timelines for permitting approvals.
5. Prepare for technological evolution. The facility was originally designed for training workloads, but quickly pivoted to support inference for LingBot-World 2.0. Design your data center with modular, flexible power distribution and liquid cooling compatibility to handle next-generation chips (e.g., GPUs with 1,200W TDP). Obsolescence risk is one of the biggest hidden costs.
This case study is not an anomaly. Similar projects are being structured across Xinjiang, Gansu, and Sichuan. The companies that move fast, partner with municipalities, and commit to green power will capture the next wave of China’s AI infrastructure buildout. The window is open, but it will not remain so forever. By 2030, the best sites in western China may already be claimed.
Source: China Gateway 360 analysis based on public filings, official press releases from Ulanqad Municipal Government (July 2026), agreements between PGL Major and Ulanqad City, and industry data from the China Data Center Alliance. | July 2026
