On the night of May 29, 2026, Shao Shuwei, an executive at Litong Electronics, likely experienced a restless sleep.

That day, Litong Electronics' stock price soared to an all-time high of 240.88 yuan. However, in the evening, the company issued a cautionary announcement highlighting high customer concentration in its computing power business, with Tencent accounting for over 70% of its revenue. It also warned of intensifying industry competition leading to potential price declines and a significant gap between its current valuation and actual performance.

Over the following two days, Litong Electronics' stock price plummeted by the daily limit, wiping out nearly a quarter of its market capitalization.

Shao Shuwei was not the only one affected. Three other individuals—Ouyang Hua in Guangzhou, Zhou Chaonan in Langfang, and Geng Kangming in Shenzhen—found themselves in similar predicaments. By June 8, Hongjing Technology's stock had tumbled from a May peak of 345 yuan to 159 yuan, nearly halving in just twenty days. Runze Technology and Xiechuang Data also saw declines of 5.19% and 6.64%, respectively.

This was no mere coincidence. The new business track they had ventured into was undergoing a market revaluation.

| A Bitter Triumph |

Shao Shuwei, Zhou Chaonan, Geng Kangming, and Ouyang Hua were all engaged in the same business model. They purchased chips from NVIDIA and leased them to internet giants, profiting from the price differential. While the concept seemed straightforward, the execution was fraught with challenges. NVIDIA's price hikes drove up their costs, while pressure from internet giants squeezed their revenues. Caught in the middle, their profit margins were razor-thin.

Zhou Chaonan's Runze Technology relied heavily on its top five clients, with ByteDance alone contributing over 60% of its revenue. Litong Electronics' situation was even more extreme, with Tencent accounting for over 70%. These major clients held significant sway over their fortunes. Litong Electronics' true gross margin of 4.51% reflected the intense pressure from both ends.

Now, clients like ByteDance, Tencent, and Alibaba are ramping up the construction of their own data centers. These four individuals find themselves at the mercy of the industrial chain, their fates closely tied to Jensen Huang's decisions at NVIDIA.

From the Hopper to the Blackwell and now the Rubin architectures, each new generation of NVIDIA chips has brought significant performance leaps. When NVIDIA released a new chip, rental prices for the previous generation would plummet. An H100 chip, purchased for 2.5-2.8 million yuan, would be worth only 1.7-1.8 million yuan three years later, nearly halving in value. Their equipment was not an asset but a rapidly depreciating consumable.

NVIDIA was not just a supplier but also a rule-setter. Its NCP Preferred certification was essentially a license to operate in this market, with Litong Electronics being the only mainland recipient. However, these licenses were not permanent. NVIDIA had the power to change rules, issue additional licenses, or revoke existing ones at will. Each new chip release triggered a sharp decline in rental prices for older models. Monthly rents for A100s fell from 28,000 to 16,000 yuan. This rhythm was dictated not by market supply and demand but by NVIDIA's product cycles.

NVIDIA bore no depreciation risks. Its profits were locked in with each sale. When chips depreciated, buyers absorbed the losses. When rental prices fell, lessors took the hit. NVIDIA, sitting at the top of the food chain, was the true landlord of this business.

At GTC 2026, Jensen Huang introduced a new metric: Tokens per Watt. Computing power became synonymous with revenue, with each Token representing potential income. This signaled a shift in competition within the computing power industry—from who had the most chips to who could generate more AI services per chip. Technical capabilities were displacing resource advantages, and technology was precisely what these four individuals lacked.

On May 31, 2026, the U.S. Department of Commerce's Bureau of Industry and Security (BIS) issued new guidelines. Exporting high-end AI chips to Chinese companies would now be based on the headquarters location of the recipient, not the shipping address. Even purchases through overseas subsidiaries in Singapore or Malaysia required U.S. licenses, with approvals presumed to be denied. For Runze Technology and Xiechuang Data, which relied on NCP certifications, this rule effectively choked off their supply chain.

NCP, or NVIDIA Cloud Partner, was NVIDIA's highest certification for cloud service providers. Globally, there were dozens of NCPs, but only five in China: Alibaba Cloud, Tencent Cloud, Runze Technology, Aojia Software (a subsidiary of Xiechuang Data), and GMI Cloud (Taiwan). Litong Electronics' subsidiary, Century Litong, held a Preferred-level dealership for DGX servers under the NPN system, a general channel certification shared by over a hundred domestic firms.

On May 31, 2026, the U.S. Department of Commerce implemented new, penetrative regulatory rules. Orders signed before this date remained unaffected, but new orders for top-tier chips would undergo thorough scrutiny of Chinese actual controllers, with licenses presumed to be rejected. This effectively cut off the latter half of their supply pipeline while the first half still flowed.

An even greater risk loomed: the closing window of scarcity.

On the supply side, NVIDIA ramped up production, flooding the market with Blackwell and Rubin chips. Domestic players like Cambricon and Huawei's Ascend accelerated their efforts, with China's AI chip shipments surging 2.7 times year-on-year in Q1 2026. On the demand side, internet giants accelerated their data center builds. ByteDance raised its 2026 capital expenditures from 160 billion to over 200 billion yuan, with half earmarked for chips and self-built computing infrastructure. Alibaba CEO Wu Yongming stated that AI investments could "far exceed 380 billion yuan," with self-developed GPUs already in mass production. Tencent and DeepSeek also ramped up their efforts. Their strategy had shifted from "addressing shortcomings" to "seizing positions."

The computing power leasing market was stratifying. Low-end supply overflowed, with A100 monthly rents plummeting from 28,000 to 16,000 yuan (a 43% drop). High-end demand outstripped supply, with H100 rents surging nearly 40% against the trend. Internet giants poured money into proprietary high-end clusters, offloading low-end inventory to the leasing market. The market was transitioning from uniform gains/losses to a tiered structure of high-end scarcity and low-end collapse.

The current climate was unforgiving for Shao Shuwei, Zhou Chaonan, Geng Kangming, and Ouyang Hua.

| The Price of Defying Fate |

On paper, Litong Electronics, Hongjing Technology, Runze Technology, and Xiechuang Data appeared profitable.

Shao Shuwei, fifty-three, had spent thirty years manufacturing TV casings. While the LCD TV market declined, he never truly stood tall, nor did he fall.

Zhou Chaonan's journey had been arduous. At forty-nine, she had planned a 1.34 million-square-meter data center on saline-alkali soil in Langfang, with her debt-to-asset ratio soaring to 99%. She waited nine years for profitability. During those nine years, China's real estate market tripled, and Langfang land prices skyrocketed. Exiting at any point to buy property would have been wiser than persisting, but she dared not retreat.

Geng Kangming's path was narrower. After six years on a Foxconn assembly line, he ventured into storage contract manufacturing, where razor-thin margins left him at the mercy of his clients. Lacking pricing power or a competitive moat, he shared Shao Shuwei's plight.

Ouyang Hua had walked this road the longest. After fifteen years in the Shantou Municipal Party Committee office, he resigned in 1995 to start a smart city company, which remained stagnant for over two decades. At sixty-three, he was still helping companies stem losses.

For these four, there was no light ahead, no path behind.

In the first half of 2023, ChatGPT launched, igniting the AI boom and a global surge in computing power demand. NVIDIA's H100 chips were in short supply, with deliveries pushed to the following year. Internet companies, flush with cash but initially conservative, avoided heavy assets. Chip firms had technology but not infrastructure. The gap—buying equipment, building data centers, and renting out usable computing power—remained unfilled. A lifeline appeared, and four stranded individuals seized it.

Zhou Chaonan, after nine years, secured ByteDance as a client. In 2025, Runze Technology's net profit soared 182% to 5.05 billion yuan. She transformed her losses into nine computing clusters, issued China's first data center REITs, and recouped 4.5 billion yuan in one fell swoop. Shao Shuwei invested 450 million yuan in Tencent's computing network, with net profit surging 1,088% to 293 million yuan in 2025, exceeding the previous seven years' earnings. Geng Kangming applied his two-decade strategy of "expanding production with every order" to GPUs. Xiechuang Data's smart computing revenue hit 2.761 billion yuan, up 1,727%.

Ouyang Hua took the greatest risk. In Q1 2026, Hongjing Technology's revenue jumped 958%, nearly tenfold in a year, with its stock price soaring from 13 to 111 yuan. At sixty-three, he finally turned a profit.

However, defying fate came at a cost.

Shao Shuwei's transformation was the swiftest. Litong Electronics lacked its own data centers, renting equipment upstream and subleasing to Tencent. Using the net method, it only recognized margin income. Restated as an operator, its gross margin was just 4.51%. Litong Electronics resembled a sublessor—the chips weren't its, the data centers weren't its, and without Tencent's contracts, it would be an empty shell. After decades in TV casings, he had simply changed shells.

Ouyang Hua's gamble was the riskiest. In Q1 2026, interest expenses hit 35 million yuan, exceeding net profit of 30.26 million yuan. His debt-to-asset ratio stood at 89.82%. With 14.6 billion yuan in comprehensive credit lines, he personally guaranteed them all. At sixty-four, his guarantees far outweighed his net worth. If computing power went unleased, even his pension was at risk.

Geng Kangming's Xiechuang Data had procured 32.2 billion yuan in AI servers, over six times its net assets. An H100, purchased for 2.5 million yuan, would fetch only 1.7 million yuan three years later, nearly halving in value. AI chips had economic lives of two to three years, compared to twenty for traditional equipment. Every day, his 32.2 billion yuan in equipment depreciated by millions. Depreciation and interest consumed 80% of profits. The assembly line boss from Foxconn now owed daily interest and faced equipment devaluation the moment he turned off the lights.

Zhou Chaonan appeared the steadiest, but ByteDance accounted for 60% of her revenue. Last year, ByteDance's capital expenditures surged 60%, much of it directed toward self-built computing power. The richer ByteDance became, the less it needed her. After nine years waiting for ByteDance, ByteDance would not wait nine years for her.

Behind these numbers lay GPU-backed leverage. While leverage could be wings when sustained, it became chains when it failed. Hu Rong, a Wuhu computing power industry operator, noted that the greatest泡沫 (bubble) in the current market was the financialization of heavy assets. If next-gen chip deliveries were delayed or AI applications failed to scale profitably, high-priced hoarded older GPUs would face dual blows of plummeting rents and illiquidity.

But with leverage maxed out, they could only push forward.

| A Difficult Road Ahead |

Shao Shuwei, Zhou Chaonan, Geng Kangming, and Ouyang Hua succeeded through heavy assets. Shao invested in stamping machines, Zhou in data centers, Geng in contract manufacturing lines, and Ouyang in government project receivables. Managing equipment, assets, and cycles were precisely what AI infrastructure lacked.

Yet their winning formula contained the seeds of defeat. In nascent industries, resources—land, capital, loss tolerance—determined success. As industries matured, profits shifted from resources to technology. Outsiders' advantages dwindled daily.

Three paths lay ahead:

* **Professionalization**: Upgrade from leasing computing power to selling AI services. Xiechuang Data spent 510 million yuan to acquire Guangwei Technology, pursuing full-stack integration. This was the hardest path but the only one to transcend industrial cycles. * **Computing-Power-Electricity Synergy**: Partner with power companies for cost advantages. This path was stable but had a low ceiling. * **Going Global**: Southeast Asia and the Middle East faced larger computing power deficits than China. However, BIS export controls loomed, threatening to strike at any moment.

These paths were arduous. Standing still meant being sandwiched between Jensen Huang's NVIDIA, which controlled chip iterations, and internet giants building computing centers. Tencent slashed low-end computing leases, while DeepSeek cut API prices by 75%. Small and medium-sized enterprises abandoned monthly rentals for pay-per-use models. Even sublessors' retail clients were exiting.

The sleepless night of May 29 might persist. But Shao Shuwei was not alone.

Image sourced from the internet. Rights reserved for the original creator.