梁文锋的500亿 投向一座北方小城

Main point 1: Core message: DeepSeek's first financing round of RMB 50 billion is going to Ulanqab, where it plans to build a ~1 GW ultra-large AI data center, deploy at least 160,000 high-end AI chips, and bring some computing capacity online from late 2027 to early 2028.

1. The funds are mainly for facility construction, server procurement, and deployment of domestic AI computing clusters; at full load, a 1 GW data center would consume about 8.76 billion kWh a year, close to the annual electricity use of a medium-sized city's residents.

2. Because of the huge investment, DeepSeek immediately started a second financing round, with recent funding reaching nearly RMB 100 billion.

Main point 2: Three criteria explain why Ulanqab was chosen.

1. Electricity price: the delivered price is about RMB 0.358/kWh, less than half of eastern industrial electricity prices, and total electricity costs over the full life cycle are 30%–40% lower than on the eastern coast.

2. Climate: the annual average temperature is 4.3°C, and natural cooling can be used for nearly 10 months a year, reducing the second-largest cost.

3. Network: the dedicated fiber link to Beijing has a one-way latency of only 2.1 ms, suitable for tech giants needing rapid delivery.

Main point 3: Ordinary people can also see changes in jobs and the city.

1. Local data center senior delivery managers and senior operations engineers earn monthly salaries of RMB 15,000–30,000 with 14 months' pay, and such positions are increasing.

2. Signed computing center projects have grown from 36 to 89, and Ulanqab is shifting from the Potato Capital to the Token Capital.

Main point 1: Brand positioning and city branding.

1. Ulanqab has upgraded its identity along a three-stage path — Grassland Cloud Valley, Smart Computing City, Token Capital — making it, according to Goldman Sachs, one of the largest and fastest-growing AI computing clusters in Asia-Pacific.

2. Slogans and landmarks reinforce memory: "South has Guiyang, North has Ulanqab" and "One step west from Beijing is Ulanqab", with streets named Huawei Avenue, Apple Avenue, and Alibaba Avenue.

Main point 2: Cost-based pricing and competitive strategy.

1. Low electricity prices are the core weapon: the delivered price is about RMB 0.358/kWh, and total electricity costs over the life cycle for comparable computing capacity are 30%–40% lower than on the eastern coast.

2. This is similar to a brand using supply-chain advantages for price competition, packaging the cool climate, green power, and land resources into product selling points.

Main point 3: How user scale changes the delivery model.

1. DeepSeek has 129 million monthly active users, is moving toward trillion-parameter models, and inference demand is growing exponentially, so it is shifting from renting computing capacity to building its own.

2. Brands can observe: after the user base crosses a threshold, service delivery is forced to become asset-heavy.

Main point 1: A growing market and demand signals.

1. Ulanqab's computing center projects have grown from 36 to 89, of which 86 are intelligent computing centers; Goldman Sachs estimates the capacity in operation and planned is about 12.5 GW, more than 10 times the scale in operation in 2025.

2. With DeepSeek at 129 million MAU and inference requests growing exponentially, large-model services are moving from R&D to application, and computing demand is in an explosive phase.

Main point 2: Cooperation models and policy opportunities.

1. The Token Capital Three-Year Action Plan targets computing capacity above 200,000 P and revenue of RMB 8 billion in 2026, and explores a Token trading and settlement platform, opening a policy window for equipment, operations, and trading services.

2. DeepSeek is running both self-built and externally leased capacity, showing that asset-light cooperation and asset-heavy construction can proceed in parallel; 21Vianet, GDS, ByteDance, and Envision are potential major customers or partners.

Main point 3: Risks and responses.

1. The total investment for a 1 GW data center is about RMB 160–170 billion; local officials admit the computing industry is still losing money.

2. With computing capacity concentrated online from late 2027 to early 2028, sellers should guard against overcapacity and price competition and leave sufficient safety margin in advance financing and payment terms.

Main point 1: Manufacturing orders and design needs.

1. DeepSeek plans to deploy at least 160,000 high-end AI chips, 21Vianet will provide 210,000 standard racks, and Envision's Xinghe base covers 120,000 square meters of building area, generating manufacturing demand for data center shells, cabinets, power supply, and cooling.

2. Job postings include senior data center delivery managers, IDC design and planning engineers, and senior operations engineers, showing that delivery management and planning/design are the core gaps.

Main point 2: Industrial foundation and business opportunities.

1. Ulanqab originally used cheap electricity to produce ferroalloys; 2025 output was 11.32 million tons, accounting for 72% of Inner Mongolia and exceeding RMB 100 billion in output value. This resource-conversion capability can extend to computing hardware, energy storage, and temperature control.

2. The local area calls potatoes, ferroalloys, and tokens the "3T", showing that traditional factories can use new industry labels to regain investment attraction and policy resources.

Main point 3: Digital implications.

1. As Ulanqab moves from Potato Capital to Token Capital, factories should watch how local procurement catalogs change with industrial planning shifts.

2. A 1 GW data center uses about 8.76 billion kWh per year, making power infrastructure and green energy equipment a long-term source of orders.

Main point 1: Industry trends.

1. AI computing is entering a stage where competition centers on electricity, land, and cooling; both Goldman Sachs and Morgan Stanley name Ulanqab as a key beneficiary cluster.

2. Signed computing projects have grown from 36 to 89, and Goldman Sachs estimates planned capacity at about 12.5 GW; Envision's Xinghe base claims to be the world's largest single AI computing facility, and the industry is moving toward hyperscale.

Main point 2: Client pain points and solutions.

1. Clients' pain points are electricity bills, cooling, and network latency; Ulanqab responds with RMB 0.358/kWh power, nearly 10 months of natural cooling per year, and 2.1 ms direct connection to Beijing.

2. This solution can be refined into a standardized site-selection model — cool climate + low electricity prices + proximity to core markets — and exported as consulting services.

Main point 3: Entry points for service providers.

1. DeepSeek's self-build plus external leasing model creates new business in IDC design, operations, and server delivery.

2. The local exploration of a Token trading and settlement platform opens a new track for data services, trading systems, and settlement service providers.

Main point 1: Platform expansion moves.

1. Huawei led the way in 2013, launching the Grassland Cloud Valley; Apple, Alibaba, UCloud, and Kuaishou followed. Kuaishou's RMB 10-billion data center is its first self-built hyperscale data center.

2. ByteDance plans to add 5–6 GW of computing capacity, to be delivered in early 2028; DeepSeek is building a 1 GW data center in Ulanqab through a combination of self-building and leasing.

Main point 2: Investment attraction and cooperation.

1. The government issued the Token Capital Three-Year Action Plan, targeting 200,000 P of computing capacity and RMB 8 billion in revenue by 2026, and exploring a Token trading and settlement platform; platform companies can seek policy support or participate in co-construction.

2. GDS has invested more than RMB 30 billion over five years, and 21Vianet has invested RMB 10.6 billion in total and built 210,000 racks, showing that third-party data center platforms are also seizing positions at scale.

Main point 3: Operations management and risk avoidance.

1. Local officials admit the computing industry is still losing money; a 1 GW project costs about RMB 160–170 billion, so platform companies should control the pace of self-building.

2. After concentrated delivery between late 2027 and 2028, idle computing capacity and price wars may occur; projects closer to Beijing with lower latency and access to low electricity prices are more resilient.

Main point 1: New industry developments.

1. The focus of AI competition is shifting from model algorithms to electricity, land, and cooling; Goldman Sachs calls Ulanqab one of Asia-Pacific's largest and fastest-growing AI computing clusters.

2. Computing center projects have grown from 36 to 89, and Goldman Sachs estimates capacity in operation and planned at about 12.5 GW, reflecting an investment boom in intelligent computing infrastructure.

Main point 2: Business model evolution.

1. After raising RMB 50 billion in its first round, DeepSeek built a 1 GW data center and then launched a new round of about RMB 100 billion, representing big-model companies moving from renting computing capacity to asset-heavy Token factories.

2. Envision's Xinghe base has reached one million P with a total plan exceeding 2 GW; officially described as the world's largest single AI computing facility, it represents a new form of computing organization.

Main point 3: Policy implications and issues.

1. Ulanqab puts potatoes, ferroalloys, and tokens together as the "3T" and uses the Token Capital Three-Year Action Plan to turn low electricity prices, cold resources, and proximity to Beijing into digital-industry rent.

2. But the region has net population outflows and its GDP ranks in the lower-middle range nationally; officials admit it is still losing money, underscoring that resource-based cities developing computing face fiscal pressure during the investment period, capacity absorption, and cross-regional allocation issues.