The artificial intelligence industry has entered a period of rapid growth.

Attracting computing power investments has become a popular choice for local governments seeking to develop their digital economies.

Utilization rates remain below 30%, while operating costs continue to rise.

Some regions have blindly followed the trend, engaged in subsidy races, and built computing facilities under the guise of other projects, triggering a vicious cycle of competition in computing power investment.

Just last week, Xinhua News Agency specifically warned against “the vicious cycle of computing power investment,” directly pointing to the short-sightedness of local development strategies.

The fact that the agency has personally stepped in to sound the alarm speaks volumes. It is time to hit the brakes on this computing power race.

Intelligent computing centers—building them leads to losses, but not building them leads to even greater losses—reflect the collective dilemma of computing power investment promotion.

01 The Subsidy Race: A Striking Contrast

First, consider a stark contrast.

In southern coastal cities, computing power vouchers can reach up to 10 million yuan annually.

In provincial capitals in the Northeast, the annual subsidy cap for enterprises is only 1 million yuan.

We see that the former is 10 times that of the latter.

It’s not that the industrial gap is this wide; rather, subsidy standards have deviated from industrial logic and turned into a “bidding war.”

Some cities in the Yangtze River Delta are even more aggressive, offering computing power subsidies as high as 50% and allowing policies at the provincial, municipal, and district levels to be stacked, driving the overall subsidy rate higher and higher. Furthermore, beyond computing power vouchers, “model vouchers,” “data vouchers,” and “corpus vouchers” have emerged.

The names may differ, but the essence remains the same: the government foots the bill, with subsidies piling up at every level.

The problem is that not every city dishing out the money has a corresponding industrial need.

There is a city in the Yangtze River Delta known for its manufacturing sector. Despite its small local AI industry and limited computing power needs, it was compelled to introduce a computing power voucher policy under pressure from neighboring cities.

The result? A portion of the computing power vouchers were issued, but the redemption rate was very low; the portion that wasn’t issued simply sat idle in the accounts.

Funds were allocated but went unused—so why follow suit when everyone knows it’s not cost-effective?

Local authorities “fear missing out on industrial opportunities” and “worry about falling behind neighboring cities”—no one wants to stop or be left behind.

The result is that while everyone is making what appears to be a “rational” choice individually, collectively, it amounts to collective irrationality. The competition is becoming increasingly fierce, and every round of escalation makes it harder to back down in the next round.

“Taking a detour”: if they can’t build a computing power center, they’ll just build it under a different name.

The state has set barriers for computing center construction—including approval processes, technical standards, and hub node clustering—aimed at preventing uncontrolled proliferation and resource idling. Yet, driven by the impulse to attract investment, some localities have developed two “workaround” models.

The first: private enterprises take the lead.

They adopt a “self-construction + leasing” approach to build computing centers, while simultaneously requiring these private enterprises to undertake investment promotion tasks. By bundling construction and investment promotion, the project is not formally built by the government, but in essence, it is the government’s will being implemented.

The second model: State-owned enterprises (SOEs) procure equipment.

Instead of directly building complete computing centers, state-owned enterprises are tasked with purchasing core equipment such as servers to provide low-latency local computing services to local small and medium-sized enterprises.

The head of a state-owned enterprise in eastern China put it bluntly:

“This workaround allows us to circumvent regulatory guidance, reduces cash flow pressure on SOEs, and helps local governments attract businesses.”

It seems like everyone benefits, right? But a closer look reveals that all the risk is being shouldered by the SOEs.

Government computing power subsidies are directed solely at the end-users—that is, the AI companies that have been attracted to the area.

As the computing power providers, SOEs receive not a single penny in subsidies. They barely break even on daily operating costs, which is nowhere near enough to cover the high depreciation costs of the servers. While the book depreciation cycle for GPUs is 4 to 5 years, in reality, chip architectures undergo a generational update roughly every two years.

Many localities haven’t seriously calculated this depreciation cost yet.

In traditional investment promotion, even if a factory building stands empty, the steel and concrete don’t vanish, and the land remains. Computing power investment promotion is entirely different—what’s being purchased is “digital perishables” with an expiration date.

With state-owned enterprises acting as a safety net and the government reaping the benefits, this model may make the data look good in the short term, but in the long run, it could drag state-owned enterprises into a sustainability crisis, undermining the preservation and appreciation of state assets.

02 Corporate Arbitrage, Industry Fails to Take Root

What has been attracted by pouring subsidies into the market?

Some cities have attracted not high-quality enterprises capable of taking root, but a batch of “arbitrage players.”

Lured primarily by hefty subsidies, they set up shell subsidiaries locally, while their core technology, talent, and operations remain entirely off-site.

Amid local governments’ blind pursuit of investment, the threshold for verifying a company’s actual operational status has been lowered—whether intentionally or not.

Without thoroughly assessing core competitiveness or verifying the company’s ability to drive industrial growth, the money is disbursed first—questions to follow later. When these companies eventually fail or even abscond, there are few assets left to liquidate, leaving local governments with the immense pressure of cleaning up the mess.

What is truly disheartening is the flip side: AI companies with core technologies and growth potential have failed to receive targeted support due to the homogenization and oversimplification of subsidy policies.

This is the deepest-seated danger of the “computing power investment race”: it is not merely a waste of money, but a distortion of the industrial ecosystem.

In any industry, when subsidies become a company’s primary consideration for location—rather than industrial chain support, application scenarios, or the talent environment—the entire logic of investment promotion goes off track.

It is also worth noting that the AI industry is characterized by extremely rapid technological iteration.

In some regions, a lack of scientific assessment of their own industrial foundations has led to blind, trend-following development, resulting in an imbalance characterized by “heavy investment, light returns” and “heavy construction, light demand.”

Once the technological trajectory shifts, the capital and resources poured in during the early stages face the risk of becoming obsolete. Amid homogenized competition, how can cities without a differentiated positioning retain enterprises?

03 Expanding Application Scenarios: A Window of Opportunity Opens

The problems have been laid out clearly enough—where is the way forward?

The first cut must be made at the top.

At the national level, we need to strengthen routine, dynamic monitoring of computing power usage, coordinate the layout of computing nodes nationwide, and concentrate resources at national hub nodes.

Drawing on past experiences in reducing phone and internet fees—through coordinated allocation, we can truly drive down the cost of computing power usage across society.

Establish cross-regional computing power coordination mechanisms to break down local barriers, allowing idle computing power to circulate rather than having every city build its own system only to let it sit idle.

The second cut targets market chaos.

Establish a filing and review system for subsidy policies, issuing early warnings for subsidies with excessively high ratios or excessive scales.

The key is to shift the approach to subsidies: move from “blanket handouts” to “targeted support.” Subsidies should be based on results, not scale, and blindly increasing subsidies beyond fiscal capacity must be strictly prohibited.

The third cut targets the mindset.

This is the most critical cut: shifting from “subsidizing computing power” to “expanding application scenarios.” Computing power itself does not generate tax revenue; it is the application scenarios that utilize that power that do.

Local governments’ investment promotion efforts should shift from “competing to attract GPU companies to set up shop” to “cultivating the local industry’s ability to utilize computing power.”

The key is to promote the deep integration of AI with manufacturing, services, and agriculture, using application needs to drive computing power development—rather than first piling on hardware and then scrambling to find users.

The fourth cut targets performance evaluations.

Performance evaluations should incorporate metrics on computing power utilization efficiency and the actual economic impact on industries, while de-emphasizing short-term scale targets.

We must stop focusing solely on “how much has been built” and instead look at “how much is being used.” At the same time, we must regulate the behavior of state-owned enterprises (SOEs) in computing power development, strictly prohibit SOEs from bearing investment promotion costs in disguised forms, establish a risk assessment mechanism for computing power project investments, and strictly control debt risks.

Final Thoughts

Ultimately, computing power investment promotion is not off-limits. But we cannot approach digital-age initiatives with an “industrial park mindset.”

GPUs are not like steel and concrete; the key is that they have a shelf life. For local investment promotion, the competition lies in whose computing power is actually used, utilized to full capacity, and retained.

When the advantage of subsidies gives way to the advantage of an ecosystem, companies will naturally “vote with their feet.” The window for course correction has opened; it is time to settle the accounts.