The GPU Rental Boom Amid a Computing Crunch: A 68‑Billion‑Yuan "Hunger Game"

On July 31, 2026, A‑share stocks related to computing power leasing surged across the board. According to data from Tonghuashun, 174 out of 184 stocks in the sector closed higher, with more than a dozen names—including Hongjing Technology, Xinneng Energy, and Litong Electronics—hitting the daily limit. Behind this rally lies a striking statistic: data from the China Academy of Information and Communications Technology shows that the domestic computing power leasing market reached RMB 68 billion in the first quarter of 2026, up 62% year‑on‑year, and is expected to exceed RMB 260 billion for the full year. Globally, industry analysts project that worldwide revenue from computing‑power‑related leasing services will surpass $120 billion in 2026, maintaining an annual growth rate above 50%.

On one side, capital markets are racing ahead; on the other, the thirst for computing power remains unquenched. In the first quarter, domestic AI computing demand surged 417% year‑on‑year, while supply grew only 128%—a nearly four‑fold gap that is pushing computing power leasing to centre stage in the AI boom. From large‑model training to inference applications, from research institutes to traditional manufacturers, almost every sector is scrambling to "find cards." Yet interestingly, the market is not truly "card‑less"; rather, cards are sitting in the wrong places or in the wrong forms. This distorted supply‑demand dynamic has fuelled an unprecedented leasing frenzy.

I. The Computing Crunch: A "Structural" Shortage

If computing power is the water and electricity of the AI era, the current situation is that the taps are fully open, but the pipes are not wide enough.

Guo Tao, deputy director of the China E‑commerce Expert Service Centre, points out that the accelerated rollout of AI agents and other applications is pushing token consumption to exponential heights, while the near‑term surge in enterprise demand for high‑end computing is colliding with the slow ramp‑up of GPU production capacity. Together, these forces have created a severe supply‑demand imbalance. For example, the spot delivery lead time for H100‑equivalent computing from leading domestic cloud providers has stretched to more than 12 weeks, and some customised clusters require up to six months of waiting.

However, this "shortage" is not monolithic. In early July 2026, Meta announced that it was formally pushing forward with its internal "Meta Compute" plan to sell or lease idle AI computing power to external customers. The news sent the Philadelphia Semiconductor Index tumbling more than 6% at one point, sparking fears of a "compute glut." Soon after, Microsoft was also reported to have some data centres unable to activate already‑delivered GPUs due to power‑infrastructure delays, leaving billions of dollars of assets "plugged in but offline."

On one hand there is a vast overall computing gap; on the other, idle capacity sits with the tech giants. This seemingly contradictory picture reveals the "structural stratification" the computing market is undergoing. The massive investments of top AI firms have created temporary surpluses in certain specific GPU specifications, but the broader market's hunger for general‑purpose computing is far from satisfied. As one industry insider put it, roughly 15‑20% of high‑performance GPUs worldwide are in a "boxed‑but‑unpowered" state—not because there are no cards, but because supporting infrastructure such as data‑centre space, power, and cooling cannot keep up. In China, in particular, energy‑consumption permit approvals for data centres in first‑tier cities have tightened, so many procured computing devices must wait for government clearance, further exacerbating the "cards‑without‑power" dilemma.

This structural contradiction provides the most fertile ground for the explosion of computing leasing. Leasing models do not require every user to build their own data centre or bear the risks of energy approvals; they simply call capacity on demand, effectively hedging against various "soft costs" of hardware deployment. Meanwhile, the national "East‑to‑West Computing" project has entered its second phase, with bandwidth among the eight national computing hub nodes upgraded to 400G, greatly improving cross‑regional scheduling. This makes low‑cost western computing power more efficiently leasable to eastern demand centres, further expanding the reach of the leasing market.

II. Leasing Models: From "Resource Rental" to "Value‑Sharing"

Computing leasing is not new, but the leasing market in 2026 is undergoing a qualitative transformation.

In the past, computing leasing was essentially a "per‑card/per‑hour" resource‑rental business. Customers rented a certain number of GPUs and paid by the hour or month; the lessor took no responsibility for how many useful tokens were generated or how fast the model converged. That crude model worked when computing was extremely scarce, but as supply gradually grows, customers are demanding more sophisticated pricing.

With the proliferation of token‑based billing, the rules are shifting from "selling hardware" to "selling outcomes." Customers no longer pay simply for hardware specs; their core metrics become cost per delivered token, effective output efficiency, and service stability. A growing number of leasing platforms now offer "pay‑per‑effective‑compute‑output" pricing, including value‑added services such as model fine‑tuning and distributed‑training‑framework adaptation. For example, some third‑party computing service providers have launched "training‑as‑a‑service" packages: customers only upload their data and model code, and the platform automatically handles cluster scheduling, fault recovery, and checkpoint resume, billing ultimately based on the number of tokens successfully trained.

This shift directly unlocks the differentiated advantages of domestic Chinese chips. Without needing to compete on peak single‑card performance, domestic chips can leverage their cost advantages and, combined with full‑stack performance‑optimisation technologies, raise overall cluster throughput to deliver better cost‑performance than foreign alternatives. Many domestic leasing providers have launched dedicated leasing pools based on home‑grown chips such as Huawei's Ascend and Hygon's DCU, with prices 30‑50% lower than Nvidia A100/H100 series. After framework‑level optimisation, the effective training compute utilisation can reach over 75%, approaching 90% of the level of mainstream foreign products. By relying on existing installed capacity rather than waiting for new production, leasing providers have become a "critical force" in bridging the current computing gap.

From a policy perspective, the picture is even more supportive. On July 31, the National Development and Reform Commission disclosed at a press conference that, according to relevant estimates, the construction of the national computing network during the "15th Five‑Year Plan" period will generate RMB 4 trillion in direct investment. Given that computing construction is primarily enterprise‑financed, this creates enormous room for private investment. The same day, the Political Bureau of the CPC Central Committee stressed the solid promotion of the "six networks" planning and construction, with the computing network explicitly included. Galaxy Securities noted that the computing network, which enables cross‑regional scheduling to support large‑model training, has now entered its implementation phase. The RMB 4 trillion top‑level design is laying the institutional runway for the long‑term growth of computing leasing. Concurrently, many local governments have introduced "computing vouchers" and subsidies, directly lowering the threshold for smaller enterprises to lease computing power and further stimulating demand.

III. Earnings Delivery: Who Is Quietly Making a Fortune?

The frenzy in concept stocks is not unfounded—earnings are the ultimate litmus test.

Litong Electronics' semi‑annual performance pre‑announcement for 2026 shows expected net profit attributable to shareholders of RMB 650‑750 million for the first half, an astonishing year‑on‑year increase of 1172‑1368%, driven by continued growth in various computing‑business revenues. Notably, Litong's computing‑leasing business, only two years old, now contributes over 70% of the company's profit, vividly reflecting the sector's explosive momentum.

This is not an isolated case. From the semi‑annual reports released so far, a dense flow of AI infrastructure orders is leaving clear marks on hardware companies' income statements. Foxconn Industrial Internet expects first‑half net profit of RMB 23.4‑24.4 billion, up 93‑101% year‑on‑year, with cloud‑service‑provider AI server revenue growing more than 230%. Inspur Information expects net profit of RMB 2.6‑3.1 billion, a rise of 226‑288%. Dawning Information expects a 150‑180% increase, with shipments of liquid‑cooled computing clusters doubling. Unisplendour's subsidiary, H3C, also reported full AI server orders and a revenue increase of over 200% in the first half.

On the demand side, large computing orders are also being placed at a rapid pace. Jiechuang Intelligence announced on July 20 that it plans to purchase up to RMB 10 billion in IT equipment. Sieyuan Information announced on July 14 that it intends to buy high‑performance computing servers, with a total contract value not exceeding RMB 5.079 billion. Even Yitian Intelligence, whose main business is integrated stoves, disclosed that its subsidiary would purchase no more than RMB 550 million in server equipment. More notably, many listed companies are adopting a hybrid "lease‑first‑then‑buy" model, testing through leasing before deciding on purchases, which further supports sustained revenue for leasing service providers.

The cross‑border entry of traditional manufacturers and the frequent appearance of billion‑yuan procurement deals all point to the same conclusion: computing power is no longer the exclusive domain of tech giants; it is becoming infrastructure that all industries are racing to deploy. According to industry statistics, more than 40 A‑share companies from non‑TMT sectors—covering building materials, chemicals, textiles, and other traditional fields—have announced computing‑leasing or procurement plans in the first half of 2026, reflecting the deep penetration of industrial digitalisation.

IV. Undercurrents: The Other Side of the Boom

Yet behind the feast, there are hidden concerns.

The computing‑leasing market's white‑hot fervour rests on a deep‑seated contradiction: the chronic shortfall in high‑end GPU supply. Global AI computing demand is growing at 45‑60% annually, while high‑end GPU supply is expanding at only 15‑20%. HBM capacity is short by 50‑60%. This imbalance cannot be fundamentally reversed in the near term, meaning leasing prices will remain elevated, squeezing profit margins for downstream SMEs and potentially dampening application innovation over the long run.

Meanwhile, the tech giants continue to pour money in. Meta's 2026 capital expenditure guidance stands at $125‑145 billion; Microsoft's quarterly capex hit $41 billion, up 70% year‑on‑year, all for GPUs to expand Azure computing clusters. Some forecasts put global annual AI infrastructure investment at $4 trillion by 2030. But once the large cloud providers have built their own clusters, any surplus capacity could spill over into the leasing market at any time, directly competing with smaller lessors and eroding their market share.

Another non‑negligible risk is technological depreciation. Nvidia has already unveiled its next‑generation Rubin GPU architecture, scheduled for mass production in 2027, which is expected to be several times more powerful than the current Blackwell generation. Once new chips hit the market in volume, rental prices for current mainstream models (such as A100 and H100) could drop precipitously, and lessors holding large inventories would face significant asset‑write‑down pressure. China faces a similar challenge from rapid domestic chip iteration: the Ascend 910B has not yet been fully deployed before the Ascend 950 has already been announced, shortening the investment‑recovery cycle for leasing providers.

Wall Street has already begun to worry about a "cautionary tale"—during the Internet bubble, massive capital poured into fibre‑optic construction, which delivered long‑term value but caused serious short‑term problems. When computing power transitions from a "scarce resource" to a "commodity," can those leasing players that entered at high valuations and lack sophisticated operational capabilities survive a cyclical downturn? In addition, energy consumption and carbon emissions represent potential regulatory risks. Computing clusters are notoriously energy‑intensive. Under China's "dual carbon" goals, energy‑consumption assessments in some regions may tighten, further raising operating costs for leasing providers and reducing their economic viability.

Closing

The computing‑leasing market in July 2026 resembles nothing so much as a "Hunger Game"—demand is voraciously fighting for scraps, supply is stretched thin, and the leasing intermediaries have become the biggest winners.

RMB 68 billion in first‑quarter scale, 417% demand growth, 1172% profit growth—the numbers are seductive, and the story is compelling. But what will truly determine how far this sector can go is not the short‑term supply‑demand mismatch, but whether computing power can genuinely evolve from an "expensive rarity" into "ubiquitous infrastructure." The RMB 4 trillion computing‑network investment blueprint has already been laid out, and the upgrade from "resource rental" to "value‑sharing" is well underway. With continuous improvements in computing‑scheduling technologies, domestic‑chip ecosystems, and token‑economy models, the leasing market will inevitably move from wild growth to refined operations.

It is foreseeable that the computing‑leasing industry will undergo a major shakeout over the next two years: large platforms with technical optimisation capabilities, deep ties to domestic chips, and full‑stack service offerings will dominate, while those "middleman" players that merely flip hardware will gradually be weeded out. The golden age of computing leasing may have just begun, but only those who can ride out the cycles will have the last laugh.