China has spent billions of dollars building far too many data centers for AI and compute - could it lead to a huge market crash?

[Barnsley]

80% of the country's new data center capacity is unused, reports say!

• DeepSeek is just one of the many reasons China's AI growth simply didn't materialize.
• 
  
• Up to 80% of new data center capacity hasn't been used according to local sources.
• 
  
• Should this capacity hit the wider market, it could cause a major headache to data center developers?

China’s AI infrastructure boom is faltering, as according to a report in MIT Technology Review, the country built hundreds of data centers to support its AI ambitions, but many are now sitting unused.

Billions were invested by both state and private entities in 2023 and 2024, with the expectation that demand for GPU rentals would keep growing, but uptake has in fact dropped off, and as a result many operators are now struggling to survive.

Much of the early momentum was driven by hype. The government, keen for China to become a global leader in AI, encouraged local officials to fast-track data center construction with the result that more than 500 projects were announced nationwide, and at least 150 were completed by the end of 2024, according to state-affiliated sources. However, MIT Technology Review says local publications are reporting that up to 80% of this new computing capacity remains idle.


Selling off GPUs

Location is also a problem, MIT Technology Review notes. Facilities built in central and western China, where electricity is cheap, now face issues meeting latency requirements. In cities like Zhengzhou, operators are reportedly even giving away free compute vouchers in an attempt to lure users.

In some regions, developers began selling off GPUs after failing to secure long-term clients.

Xiao Li, a data center project manager who spoke with MIT Technology Review, said many WeChat groups that once boasted about Nvidia chip deals have gone quiet. “It seems like everyone is selling, but few are buying,” he noted.

Should this capacity hit the wider market, it could cause a major headache to data center developers, flooding an already soft sector with even more supply and pushing prices down further.

https://www.techradar.com/pro/china...-compute-could-it-lead-to-a-huge-market-crash

 

[KevinK]

Thanks for the pointer. The original MIT tech review article is even better at highlighting the real issues.

China built hundreds of AI data centers to catch the AI boom. Now many stand unused.

The country poured billions into AI infrastructure, but the data center gold rush is unraveling as speculative investments collide with weak demand and DeepSeek shifts AI trends.

www.technologyreview.com

 

[tonyget]

So who says China struggle to get more GPU？Who says GPU is the bottleneck of Chinese AI development? There is no GPU shortage in China，in fact，there is a GPU surplus

 

[Jert]

This is what always happens in authoritarian planned economy, and always just the complete opposite of "Skate to where the puck is going to be, not where it has been". No surprise.

By the way, this is how CCP bureaucrats get promotion !
Same thing is happening for fabs.

 

[KevinK]

So who says China struggle to get more GPU？Who says GPU is the bottleneck of Chinese AI development? There is no GPU shortage in China，in fact，there is a GPU surplus

If you read a little closer, especially the MIT article, you'll get a different picture:
* China has too many data centers / GPUs configured for TRAINING.
* DeepSeek has greatly reduced the demand for TRAINING, but inference is a different story
* The data centers are poorly designed / implemented in general, and especially for inference.
* Plus many are geographically isolated from where there is demand for inference, rendering them mostly valuless due to latency.

 

[Daniel Nenni]

Techradar did not do the MIT article justice.

Training versus inference, we may see this same issue here since the majority of datacenters are still in the US. AI model training datacenters would be overpowered for inference but they can still do inference, correct? Just overkill, so outdated AI datacenters can be refit for inference?

 

[count]

Techradar did not do the MIT article justice.

Training versus inference, we may see this same issue here since the majority of datacenters are still in the US. AI model training datacenters would be overpowered for inference but they can still do inference, correct? Just overkill, so outdated AI datacenters can be refit for inference?

Inference needs low latency. Most inference will actually happen on the edge, IMO. Less important for chat bots, but very important for a lot of other applications.

Think about a self driving car or humanoid robot, you don't want your car to have to make a request to the cloud and have to wait 100ms for a response before every turn of the steering wheel. The inference engine has to live on the edge for these kinds of applications.

In the long term, the shift to inference and especially inference at edge may reduce demand growth for compute, but will increase demand for memory.

 

[KevinK]

Most inference will actually happen on the edge, IMO. Less important for chat bots, but very important for a lot of other applications.

Agree with edge as being important for super low latency, personalization and autonomous operation - robots, car and personal assistants. But there are a couple of other use cases for inference that demand big shared-user iron, perhaps connected to edge models.

* "Big Brain" massive reasoning models with a huge numbers of parameters (trillions), perhaps with MoE, also requiring super-fast throughput. Reasoning latency depends on throughput (tokes / second), since it uses coupled sequences of inference. The economics of these huge models still favor shared models instead of single user models as long as the utilization is high and the resources are used efficiently.

* Intra-enterprise GenAI and RAG that uses as company's proprietary data for anything from external customer support to assessing employee efficiency to analyzing sales and costs. This kind of stuff can't be easily moved to the edge because it requires realtime access to corporate data.

 

[Paul2]

80% of the country's new data center capacity is unused, reports say!

• DeepSeek is just one of the many reasons China's AI growth simply didn't materialize.
• 
  
• Up to 80% of new data center capacity hasn't been used according to local sources.
• 
  
• Should this capacity hit the wider market, it could cause a major headache to data center developers?

China’s AI infrastructure boom is faltering, as according to a report in MIT Technology Review, the country built hundreds of data centers to support its AI ambitions, but many are now sitting unused.

Billions were invested by both state and private entities in 2023 and 2024, with the expectation that demand for GPU rentals would keep growing, but uptake has in fact dropped off, and as a result many operators are now struggling to survive.

Much of the early momentum was driven by hype. The government, keen for China to become a global leader in AI, encouraged local officials to fast-track data center construction with the result that more than 500 projects were announced nationwide, and at least 150 were completed by the end of 2024, according to state-affiliated sources. However, MIT Technology Review says local publications are reporting that up to 80% of this new computing capacity remains idle.


Selling off GPUs

Location is also a problem, MIT Technology Review notes. Facilities built in central and western China, where electricity is cheap, now face issues meeting latency requirements. In cities like Zhengzhou, operators are reportedly even giving away free compute vouchers in an attempt to lure users.

In some regions, developers began selling off GPUs after failing to secure long-term clients.

Xiao Li, a data center project manager who spoke with MIT Technology Review, said many WeChat groups that once boasted about Nvidia chip deals have gone quiet. “It seems like everyone is selling, but few are buying,” he noted.

Should this capacity hit the wider market, it could cause a major headache to data center developers, flooding an already soft sector with even more supply and pushing prices down further.

https://www.techradar.com/pro/china...-compute-could-it-lead-to-a-huge-market-crash

AI is another O2O

 

[count]

Agree with edge as being important for super low latency, personalization and autonomous operation - robots, car and personal assistants. But there are a couple of other use cases for inference that demand big shared-user iron, perhaps connected to edge models.

* "Big Brain" massive reasoning models with a huge numbers of parameters (trillions), perhaps with MoE, also requiring super-fast throughput. Reasoning latency depends on throughput (tokes / second), since it uses coupled sequences of inference. The economics of these huge models still favor shared models instead of single user models as long as the utilization is high and the resources are used efficiently.

* Intra-enterprise GenAI and RAG that uses as company's proprietary data for anything from external customer support to assessing employee efficiency to analyzing sales and costs. This kind of stuff can't be easily moved to the edge because it requires realtime access to corporate data.

I am not a big believer in big brain do everything models. I think the only purpose of very large models will be for distillation of smaller models that do the thing you actually want them to do at the edge.

 

[DanX]

Are there too many EV companies in China?
Yes. But in the end, only these survivors will remain globally.

If you cannot compete with them in cost, you will also fail to compete with them in the global market.

 

[KevinK]

I am not a big believer in big brain do everything models. I think the only purpose of very large models will be for distillation of smaller models that do the thing you actually want them to do at the edge.

Pretty sure enterprises and governments aren't willing to push their protected and real-time updated data to the edge for use, though DOGE might do something ridiculous and stupid with social security and other government data. That data is needed for most "big brain" and "enterprise" use models.

 

[KevinK]

Are there too many EV companies in China?
Yes. But in the end, only these survivors will remain globally.

If you cannot compete with them in cost, you will also fail to compete with them in the global market.

Agree - but there are two issues with that for China:
* How many failed investments / enterprises and how much job loss can the economy sustain ? LGFV (local government financing vehicles) in many areas in China are already deeply in the red from non-performing housing and infrastructure investments. Pile on shuttered EV factories and inoperable data centers.
* Both EV sales and cloud inference delivery are going to require local "manufacturing" for global markets outside of China, between tariffs and latency.

 

[DanX]

Agree - but there are two issues with that for China:
* How many failed investments / enterprises and how much job loss can the economy sustain ? LGFV (local government financing vehicles) in many areas in China are already deeply in the red from non-performing housing and infrastructure investments. Pile on shuttered EV factories and inoperable data centers.
* Both EV sales and cloud inference delivery are going to require local "manufacturing" for global markets outside of China, between tariffs and latency.

Don't bet on it.
Marxism considers profit (surplus value) to be evil.
The government intends to suppress the profit on capital to near zero.

How strong would the competitiveness of the United States be if the ROI on Wall Street were to become near zero?

 

[KevinK]

Marxism considers profit (surplus value) to be evil.
The government intends to suppress the profit on capital to near zero.

But debt is debt, and if you include the off-balance-sheet LGFVs and other hidden debt, China's debt-to-GDP ratio is higher than that of many other major economies, including the United States and the European Union.

And China's globally focused EV companies are public companies that need profits to maintain an above zero stock price. In the fourth quarter of 2024, BYD's net profit in China jumped by 73.1% to a record 15 billion yuan ($2.1 billion). For the entire year, BYD's profit increased by 34% to 40.3 billion yuan, driven by a 29% rise in revenue. The company's success is attributed to lower prices and strong sales, surpassing rivals including Volkswagen in China.

 

[count]

Don't bet on it.
Marxism considers profit (surplus value) to be evil.
The government intends to suppress the profit on capital to near zero.

How strong would the competitiveness of the United States be if the ROI on Wall Street were to become near zero?

China is not Marxist or communist and hasn't been for a very long time. It's sort of a hybrid system, which is mostly capitalistic with some elements of fascism mixed in. Fascism itself has elements of socialism ingrained with it as well. But modern China is hardly more socialist than the USA.

Under a pure capitalist system with no artificial barriers to entry, margins are also supposed to be close to zero. That's the function of the invisible hand after all.

 

[DanX]

But debt is debt, and if you include the off-balance-sheet LGFVs and other hidden debt, China's debt-to-GDP ratio is higher than that of many other major economies, including the United States and the European Union.

Government subsidies are relatively small compared to LGFV.
LGFV has nothing to do with Chinese government‘s investment in the tech sector , it's from the real estate bubble.

In fact, the Chinese government is against LGFV and is in favor of investment in technology.
And the funds for government investment in technology come from the central government, not from local governments.

The central government has ample room for investment in technology and industry subsidies.

 

[DanX]

Under a pure capitalist system with no artificial barriers to entry, margins are also supposed to be close to zero. That's the function of the invisible hand after all.

This is the difference between theory and actual operation. Look at the operations of sovereign funds in Singapore and China; the advantage of having government power is just so evident.

 

[KevinK]

LGFV has nothing to do with Chinese government‘s investment in the tech sector , it's from the real estate bubble.

LGFVs have been used to develop and finance land for favored projects like car factories, semi factories and data centers
Local governments then plough in plenty more in subsidies to help finance factories, that eventually go under.
Failed factories are then sold back to local governments who use LGFVs to buy back.

We've already seen this cycle for gas powered automobiles.

‘It Is Desolate’: China’s Glut of Unused Car Factories​

Manufacturers like BYD, Tesla and Li Auto are cutting prices to move their electric cars. For gasoline-powered vehicles, the surplus of factories is even worse.

China has more than 100 factories with the capacity to build close to 40 million internal combustion engine cars a year. That is roughly twice as many as people in China want to buy, and sales of these cars are dropping fast as electric vehicles become more popular.
Last month, for the first time (March 2024), sales of battery-electric and plug-in gasoline-electric hybrid cars together surpassed those of gasoline-powered cars in China’s 35 largest cities. Dozens of gasoline-powered vehicle factories are barely running or have already been mothballed.

Sales of gasoline-powered cars plummeted to 17.7 million last year from 28.3 million in 2017, the year that Hyundai opened its Chongqing complex. That drop is equivalent to the entire European Union car market last year, or all of the United States’ annual car and light truck production.

Hyundai’s sales in China have plunged 69 percent since 2017. The company put the factory up for sale last summer, but no other automaker wanted it. Hyundai ended up selling the land, the buildings and much of the equipment back to a municipal development company in Chongqing for just $224 million, or 20 cents on the dollar.

https://www.nytimes.com/2024/04/23/business/china-auto-factories-price-cuts.html?unlocked_article_code=1.-k4.0THE.eBV1-uMwG21w&smid=url-share

 

[DanX]

Local governments then plough in plenty more in subsidies to help finance factories, that eventually go under.
Failed factories are then sold back to local governments who use LGFVs to buy back.

Most local government's subsidies are lands used to build those factories.
The cost is minimal compared to the outcome of job creation.