Hoarding GPUs for Three Years: More Cards, More Debt, But the Money Is Gone

[[**The computing power market is not comprehensively oversupplied; rather, there is a severe mismatch between new and old cards and between eastern and western supply and demand: high-end training cards remain in short supply, while old cards and western data centers sit largely idle, rents have plunged, and cross-industry card hoarders are mired in losses and insolvency.**

**Key points**

**1. Oversupply is an illusion; mismatch is the main cause**

The national average utilization rate of intelligent computing is about 30%, and that of domestically produced clusters in third- and fourth-tier cities is only 10%–22%; H100 rents have fallen from more than $7 per hour to about $3.4.

**2. Cross-industry card hoarders are mired in losses and insolvency**

Hongbo's computing power subsidiary had net assets of -5.9 million yuan in the first half of 2026; Qunxing Toys' intelligent computing gross profit was negative 15 million yuan; Gaole Co. signed a 6.752 billion yuan long-term order while loss-making and conducted sale-and-leaseback of about 4.887 billion yuan.

**3. Large customers building their own clusters and low-priced APIs are drawing away tenants**

ByteDance, Alibaba, Tencent and others have built their own 10,000-card clusters, and shipments of non-GPU accelerators surpassed GPUs for the first time; Token prices have fallen 84% from the 2023 benchmark, making API calls an order of magnitude cheaper than renting cards.

**4. Old cards are squeezed by iteration and the inference price war**

High-end new cards are still rationed and rising in price, while old cards flood into the inference price war; cards iterate in about a year, while intelligent computing centers have a payback period of 5–8 years, and idle capacity still incurs depreciation and electricity costs.

**5. Western idle capacity and blind construction are accelerating the shakeout**

More than 70% of high-computing-power demand is in the east, while western data centers are constrained by bandwidth and latency; governments and urban investment platforms account for about 35% of investors, regulatory inquiries are increasing, and sublessors without scenarios are the first to exit. 2026-10-27 10:22 Three years of hoarding cards: more cards, more debt, but the money is gone | Pianke Sheng quick overview

This article is from the WeChat public account: Pianke Sheng, author: Li Xiaonian

"

Toy sellers, dye makers, film and TV people... all came to grab cards. Three years later, some are insolvent, and some are still borrowing while carrying tens of billions in long-term orders. Where exactly did this mismatch go wrong?

In 2024, renting a decent H100 required connections.

The delivery cycle for a single card was as long as 52 weeks. Placing an order a year in advance and actually getting the card in hand was proof you really had an in.

The median rent for an H100 surged above $7 per hour in early 2024, and some people scrambled to pay domestic quotes of more than 100,000 yuan a month for an eight-card server.

As long as a batch of cards was sitting in the data center, there was no worry about customers—large model companies lined up at the door, middlemen resold scheduling slots, and whoever could get the goods simply waited to collect money.

In less than three years, the same batch of cards lies in the data centers, and the market has taken on a completely different face.

Estimates from multiple industry research institutions show that the average computing power utilization rate of intelligent computing centers nationwide is only about 30%; newly built domestically produced computing clusters in third- and fourth-tier cities have an average utilization rate of only 10% to 22%, and the idle rate of some domestically produced chip clusters is as high as 70% to 80%.

Prices are even more telling. According to media reports, the monthly rent for low-end computing nodes from some IDC service providers crashed from 30,000 yuan to less than 6,000 yuan, a drop of more than 80%; some intelligent computing parks are losing tens of millions of yuan a year and are still worrying about tens of thousands of yuan in daily electricity bills.

The data centers that spent heavily to hoard cards back then now have computing power spinning idle, fans still turning, and bills still arriving.

From one card being hard to find to not even earning back the electricity bill, only three years have passed. Many people's first reaction is: computing power is oversupplied, the bubble has burst.

What actually happened is not an oversupply of computing power but a mismatch of computing power—the newest cards are fought over, while the previous generation of cards cannot be rented out; demand in the east is queuing up, while data centers in the west are idling.

In 2023, ChatGPT ignited the large model arms race, and every company feared being the first to fall behind in the computing power competition.

And there was only one type of top-tier training card available on the market, supply was tightened layer by layer, and at the most extreme the delivery cycle approached a year.

During those two years, stories of training computing power being in short supply emerged endlessly, and the judgment repeatedly cited by the industry from Nvidia's earnings call was: cloud vendors' computing power has completely sold out.

The panic spread to the rental market, and that meant prices.

In early 2024, the median rent for an H100 surged above $7 per hour. Middlemen who could not get cards began reselling scheduling slots and subletting at marked-up prices, exactly the same logic as concert scalpers.

Then the wealth effect spread to the secondary market.

Starting in 2023, the earliest companies to cross into computing power tasted the sweetness: when the words "GPU" and "intelligent computing center" appeared in announcements, the stock price rose.

More and more followers joined in—toy sellers, dye makers, integrated stove manufacturers, film and TV companies—successively announcing their entry into computing power.

Hongbo Co. was one of the earliest companies to cross into computing power and was once the banner of the entire sector.

Its background had nothing to do with computing power—a Fujian company founded in 1999 and listed in 2008, its main business was lottery printing, a leading domestic supplier of sports lottery printing, a sign it had carried for more than 20 years.

In 2022, the company set up a wholly owned subsidiary, Yingbo Digital Technology, and, leveraging its cooperative relationship with Nvidia, jointly built an AI innovation empowerment center in Beijing to do computing power business.

As soon as the news came out, the A-share market boiled over. In 2023, Hongbo Co. rose nearly 600% in a year, with its stock price surging from around 6 yuan to more than 40 yuan, and its market value once approaching 21 billion yuan, hailed as the "No. 1 leader in computing power."

The climax of the story came in 2025. That year the company's computing power and related services revenue was 614 million yuan, which looked respectable—but the bulk came from one-time acceptance confirmation of a large project, and the core computing power subsidiary Yingbo Digital Technology's full-year net profit was only 347,000 yuan.

By the first half of 2026, computing power revenue had plunged 91.3% year on year to only 51.73 million yuan, Yingbo Digital Technology had a net loss of 15.37 million yuan, and net assets had fallen to -5.9 million yuan—insolvent.

Qunxing Toys, a toy seller, offered another version of the answer.

In 2025 the company made a cross-industry move into intelligent computing, generating full-year intelligent computing revenue of 53.36 million yuan; after deducting costs, gross profit was negative 15 million yuan.

Selling cards is not as good as selling toys.

Gaole Co., Ltd., also originally a toy manufacturer, was bolder.

Against the backdrop of continued losses in 2025 and the first half of 2026, it signed two long-term computing power service contracts totaling about 6.752 billion yuan in succession from June to August 2026. Yet as of the half-year report, the company's consolidated contract liabilities were only 2.92 million yuan, and the computing power subsidiary's revenue was zero.

Where did the money come from? From July to August it carried out two sale-and-leaseback financings totaling about 4.887 billion yuan, using subsidiary equity and accounts receivable as pledged collateral.

A small frame carrying a story worth tens of billions.

Similar cases include Lianhua Holdings (terminated contracts of about 575 million yuan) and Yitian Smart (terminated about 113 million yuan); earlier, Feilixin terminated a contract of nearly 1.1 billion yuan because GPU servers were not delivered on schedule.

There is another way to survive called being a "computing power sublandlord."

Jiuzhang Cloud, which is sprinting toward a Hong Kong listing, does not spend heavily on buying cards itself; it leases computing power long-term and then subleases it.

Its prospectus shows: lease liabilities snowballed from 11.7 million yuan at the end of 2023 to 3.087 billion yuan by the end of July 2026, with interest expense alone reaching 32.46 million yuan in the first half of 2026.

The industry's books are even more striking.

According to industry statistics, as of June 2026, China had more than 500 intelligent computing centers in use, with nearly 400 under construction or planned.

According to estimates by several brokerages, the break-even point for the computing power leasing business generally falls at a rack utilization rate of 60% to 70%, yet the industry's real cost structure and depreciation rhythm are far more complex than this number.

By the time those who placed their bets remembered to do the math, the market had already changed.

The first to leave were the big customers. ByteDance, Alibaba, and Tencent buy cards to use them, not to speculate on them.

As AI business became their main business, building their own ten-thousand-card clusters became an inevitable choice; for cards they could not obtain, they made their own. Shipments of custom ASIC chips from Huawei Ascend, T-Head, and Kunlunxin have surged.

By the first half of 2026, shipments of non-GPU accelerators in China's data centers had grown nearly 80% year on year, surpassing GPUs for the first time.

The biggest tenants have successively become landlords themselves. And those who hoarded cards soon discovered that retail customers could not be relied on either.

Big tech drove model API prices down to a few cents per million tokens, and the frontier model price index shows that by October 2026, Token prices had fallen 84% from the March 2023 baseline.

For the vast majority of enterprises, calling APIs directly is an order of magnitude cheaper than renting cards and building their own. After all, no company would build a power plant on its own roof just for lighting.

Computing power is turning from "an asset you need to use" into "a utility you can draw on at will," and utilities are not for sublandlords to sell.

In 2026, prices collapsed.

Take the leasing price index tracked by AIMultiple as an example: the median H100 rental price fell from more than $7 per hour in early 2024 to about $3.4 by mid-2026, more than halving in two and a half years. On peer-to-peer leasing platforms, listings below $2 are now everywhere.

Industry surveys show that in the first half of the year, rental prices for some small and medium-sized computing power cards fell by more than 40%.

The story of monthly rent collapsing from 30,000 yuan to 6,000 yuan is not an isolated case. Quotes for newly built intelligent computing centers in second-tier cities are only 60% of those of leading cloud vendors.

Large models have shifted from money-burning training to money-making inference; inference relies on new-architecture cards and domestic specialized chips, and the tasks old cards can handle—inference and batch processing—are precisely the bloodiest red ocean of price competition.

Cards hoarded in anticipation of the old era have met a market that no longer needs them.

Hence that bizarre picture: one end is hot.

The latest Blackwell architecture is still rationed, with on-demand B200 prices rising to $8+/hour, higher than a year ago.

Inside Alibaba Cloud there is almost not a single idle card; Ascend capacity has been swept clean of inventory by ByteDance and Alibaba, while Tencent and Baidu are still queuing.

The "sold out" described by Nvidia executives still holds true in 2026.

The other end is cold—the older generation of H100, A100, and A800.

What is crueler is that things that cannot be sold should be able to wait, but computing power cannot wait.

For a machine idle for a day, depreciation is still charged and electricity is still paid; that day is a day completely wasted. Wafers can sit in warehouses waiting for prices to rise; 24 hours in a server room cannot.

And the iteration cycle for high-end training cards has been compressed to about one year, while the investment payback period for an intelligent computing center is 5 to 8 years. By the time the server room is built and the planned costs should be recovered, the cards in hand are already two generations behind.

Listed companies have already paid a book cost for this mismatch—Zhonganke made a 15.25 million yuan impairment provision for a project due to excess computing power supply and declining leasing quotes, and this is only the beginning of the impairment wave.

Then there is the matter of location.

More than 70% of high computing power demand is concentrated in the Beijing-Tianjin-Hebei region, the Yangtze River Delta, and the Guangdong-Hong Kong-Macao Greater Bay Area, while a large number of server rooms were built in the west because of preferential electricity prices.

Limits on network bandwidth and latency are there; no matter how cheap mid- and low-end server rooms in the west are, they cannot absorb demand from the east.

The little money saved on electricity cannot make up for customers not coming.

Regulatory trends also confirm this is not alarmism.

According to media statistics, in the first nine months of 2026, the Shanghai and Shenzhen stock exchanges issued more than 42 inquiry letters and regulatory work letters concerning computing power-related matters, and Hainan Huatie was placed under investigation for information disclosure violations involving a 3.69 billion yuan computing power mega-order.

Starting in April 2025, development and reform commissions in multiple places carried out "computing power surveys" to prevent blind and duplicate construction.

Looking back at these three years, the script is not unfamiliar.

Local governments regarded intelligent computing centers as an industrial option to "take over the baton from real estate," with fiscal subsidies, special bonds, and urban investment platforms advancing funds together—one analysis estimates that governments and urban investment platforms accounted for about 35% of investment entities in this round.

Server rooms were built, cards were installed, and then they found that customers were not coming.

How similar this is to solar and flat panels back then: the direction was right, the story was real, but the right direction could not stop excessive investment.

The shakeout has already begun.

What dies are the "sublandlords" with no scenarios, no customers, and no cards but leverage, waiting for the wind to come; what survives are those holding new-architecture computing power, tied to real demand, and able to secure power quotas.

The industry is moving from "staking out territory" to "improving quality and efficiency," and this process will eliminate most of the followers who entered in 2024.

Token call volume is still growing exponentially, and the growth rate of inference computing power is already more than twice that of training computing power.

Leading clusters are roaring at full capacity, and customers who cannot get cards are still queuing. As for those server rooms built first and rented later, waiting for customers to come to them, every turn of the fans is electricity money.

This article was originally published by the WeChat public account "Pianke Sheng." The data and cases in the article come from IDC reports, public reports such as Daily Economic News, and announcements and prospectuses of relevant listed companies. Channel: Finance and Economics. This content is published with the author's authorization; the views represent only the author and do not represent Huxiu's position. If you have any objections or complaints about this manuscript, please contact tougao@huxiu.com. Everyone who is changing and wants to change the world is on the Huxiu APP [ ]