중국 AI 개발사 딥시크는 기가와트 규모 컴퓨팅 클러스터 계획이 현지 전력 인프라 준비 속도를 앞지르며 심각한 계통 연계 병목 현상에 직면했다.

Sixteen thousand... wait, 160,000 Huawei Ascend chips. A gigawatt-scale data center. Full-load consumption equivalent to 750,000 households. This is not a municipal power plan; it is a project currently underway in Ulanqab, Inner Mongolia, spearheaded by DeepSeek.

What does 750,000 households represent? At an average of three people per household, that translates to the electricity consumption of over two million residents. A single data center is effectively consuming the power of a mid-sized city.

While the AI community debates the compute capabilities of the Ascend 950, the inference speed of DeepSeek-V4, and potential API price reductions, the power sector is asking different questions: Where will the electricity come from? Who will facilitate the connection? And once connected, can the grid withstand the load?

The rationale behind choosing Ulanqab lies in its wind and solar resources. Inner Mongolia leads China in installed new energy capacity, and Ulanqab stands out as one of the province’s most concentrated renewable energy hubs. The region offers abundant wind and solar generation, low-cost power, and green energy quotas.

For AI companies, electricity constitutes the largest portion of computing costs beyond hardware. Chips can be purchased, and technology can be caught up with, but electricity is a 24/7 operational expense. By rough industry estimates, a gigawatt-scale data center could consume over eight billion kilowatt-hours annually. A mere one-cent difference in electricity pricing translates to tens or even hundreds of millions of yuan in annual cost variations.

Consequently, DeepSeek’s relocation to Ulanqab is fundamentally a power strategy rather than a purely technological one. The company is not merely seeking Huawei or land; it is pursuing affordable, green, and reliably supplied electricity. This siting logic mirrors the historical migration of cryptocurrency mining operations from northwest China to Kazakhstan and Texas: where power is cheapest, computing thrives.

However, a critical distinction must be made. AI computing and cryptocurrency mining differ fundamentally in social impact and policy positioning. Cryptocurrency mining is purely consumptive, generating no tangible output, which led to its regulatory closure in China. In contrast, AI computing holds substantial industrial and strategic value, explicitly endorsed by the state as a driver of “new quality productive forces.” Their only shared characteristic lies in the physical principle that “computing follows electricity prices.” Beyond that, their natures diverge completely and should not be conflated.

It is precisely this policy backing that grants DeepSeek the leverage to negotiate such a massive project in Ulanqab. Conversely, the sheer scale of the undertaking compels the power sector to take it seriously.

A gigawatt-level load cannot simply be plugged into the grid. Within the power industry, such a data center represents an exceptionally unique consumer. Unlike factories that operate during daytime shifts or commercial malls with distinct peak and off-peak hours, a data center’s load profile resembles a flat line. It operates at full capacity around the clock, without interruption or negotiation.

Integrating this type of load into the grid extends far beyond laying a single cable. First is connection capacity. A gigawatt-scale demand requires dedicated substations, specialized transmission lines, and tailored interconnection schemes. Determining whether the local grid can currently bear the load, whether new infrastructure is needed, and how long construction will take falls outside the purview of AI companies.

Second is green power supply. DeepSeek requires renewable energy, but wind and solar are inherently intermittent. Ensuring a 24/7 fully loaded data center runs on green power necessitates integrated energy storage, regulating power sources, and sophisticated procurement strategies. This supporting infrastructure is often more complex than the data center itself.

Third is electricity pricing. Following marketization, renewable energy prices in Inner Mongolia fluctuate significantly. Midday solar surges can drive prices to the floor or even into negative territory, while evening peaks see prices spike. Data centers seeking cheap power must accept this volatility. Yet AI training and inference workloads cannot be paused or shifted at will. Resolving this contradiction remains an unsolved challenge.

Therefore, DeepSeek’s gigawatt project functions less as a simple “high-consumption user” initiative and more as a stress test. It probes whether Inner Mongolia’s power grid, its green energy absorption framework, and its electricity market mechanisms can sustain a novel, continuously operating megaproject.

Nevertheless, the project brings mutual benefits to Ulanqab. Inner Mongolia experiences some of the nation’s fastest new energy capacity growth alongside significant wind and solar curtailment pressures. Ulanqab has long sought ways to absorb surplus renewable generation. A stable gigawatt load provides an ideal mechanism to locally consume vast amounts of wind and solar output, alleviate transmission bottlenecks, and improve renewable utilization rates. Both local government and grid operators possess strong incentives to advance the project.

The reality is a “mutual necessity”: DeepSeek requires affordable green power, while Ulanqab needs a stable baseload. Both parties are willing, but technical and regulatory barriers regarding grid connection, energy storage, and pricing stand in the way. Overcoming these hurdles will determine the project’s viability.

This development unfolds against a broader backdrop: computing power is rapidly emerging as the fastest-growing new variable within the power sector. Public data indicates that national computing facility electricity consumption reached approximately 170 billion kilowatt-hours in 2025, representing a year-on-year increase of roughly 30 percent—five to six times the overall societal electricity growth rate. This figure is projected to climb further in 2026.

Computing’s electricity demand is transitioning from marginal experimentation to major integration. Cities like Ulanqab, Hohhot, Qingyang, Zhongwei, and Gui’an are evolving into new data center hubs for a singular reason: affordable, abundant green power.

Yet the power sector is clearly unprepared. For decades, grid planning revolved around a “factories plus cities” model, with load centers concentrated along the eastern coast and power generation bases located in the west, linked by ultra-high-voltage transmission lines. The AI era is disrupting this paradigm. Computing loads are no longer anchored in eastern metropolises but are relocating to western energy bases, consuming power locally without requiring cross-provincial transmission. This shift presents both advantages and complications.

On the positive side, it enables local absorption of renewable energy, reducing transmission burdens. On the challenging side, its load characteristics starkly contrast with traditional consumers. Grid planning, interconnection standards, pricing mechanisms, and safety parameters all require fundamental recalibration.

Compounding the issue is the finite nature of “affordable green power.” As computing projects multiply, a single gigawatt data center can consume a substantial portion of an entire renewable base’s output. When AI begins demanding electricity in increments equivalent to “750,000 households,” grid managers must confront a previously uncalculated dilemma: how should limited cheap green power be allocated?

Prioritizing computing can stimulate regional economies and accelerate industrial upgrading, yet the long-term sustainability of computing projects remains uncertain. Will data center utilization rates hold steady? Could intensifying price wars in the AI sector leave facilities idling as empty power drains?

Allocating power to traditional industry and residential users follows established logic, but growth is comparatively modest, offering less immediate economic stimulus than computing. This allocation dilemma did not exist previously. Now it has arrived, and the power sector lacks a mature evaluation framework to address it.

The true test lies in perspective. While the AI community views this project through the lens of chips, compute capacity, and domestic substitution, the power industry sees a fundamental examination: Can a gigawatt-scale, 24/7 fully loaded, green-power-requiring, cost-sensitive giant secure a long-term, stable, and economically viable position within Inner Mongolia’s existing power system? There is no ready-made answer.

Historically, the power sector operated on the principle of “where the user is, the power is delivered.” Computing flips this logic: “where the power is, the computing is built.” This emergence of “load following power sources” represents an unprecedented shift for the industry. It forces a fundamental reevaluation of interconnection upgrades, grid planning, tariff structures, green energy allocation, and, most critically, whether gigawatt-scale computing loads constitute a “new opportunity” or a “new risk.”

Looking ahead, three indicators will determine whether this project warrants continued monitoring. First, the approval and construction timeline for supporting substations and interconnection infrastructure. This serves as the primary feasibility threshold; delays here stall everything downstream.

Second, the development of energy storage and regulating power source solutions. Whether relying on direct green power supply or market-based procurement, volatility must be smoothed by balancing resources. The maturity of this plan dictates whether the data center consumes “genuine green power” or merely “nominal” certificates.

Third, the execution of long-term power purchase agreements. Data centers are long-cycle assets requiring predictable long-term pricing rather than exposure to spot market volatility. Securing such contracts would validate the business model’s closure; failure to do so leaves the project perpetually suspended.

Ultimately, DeepSeek’s gigawatt ambitions will not hinge solely on Huawei’s chip production capacity, corporate financing, or order pipelines. They must successfully navigate the power sector’s rigorous requirements.