The Surge in Computing Power Rental Prices: The “Awakening of Pricing Power” of the Scarcest Asset Amid the AI Boom and the Restructuring of the Industrial Chain

Over the past week, the hottest topic in the computing power sector has undoubtedly been “the surge in computing power rental prices.” On September 29, computing power rental concept stocks surged collectively. Nanwei Software posted four limit-ups in six sessions, Maixinlin hit the 20% limit-up, and Chaoxun Technology, Fengzhushou, Jiahua Technology and others followed the rally. The core catalyst behind this round of market movement came from a pricing announcement across the Pacific—cloud computing service provider Nebius announced that starting October 1, it would fully raise on-demand rental prices for NVIDIA GPUs, with H100 rising from $3.85 per GPU hour to $4.50, and B300 rising from $7.85 to $9.50, an overall increase of about 17% to 21%. This is not an ordinary price adjustment, but a key signal that the supply-demand contradiction in computing power is moving from “narrative” to “financial statements,” and a landmark event in computing power’s leap from a “cost item” to a “core asset.”

I. The Full Picture of Price Increases: Overseas Leading, Domestic Following, and a “Triple Resonance”

Nebius’s price increase covers current mainstream high-end GPU models: H100 rose about 17%, H200 rose 20%, B200 rose nearly 19%, and B300 had the largest absolute increase at 21%. Zheshang Securities pointed out that the on-demand price is the most direct pricing anchor for new signings and small and medium-sized customers; its across-the-board increase means the marginal retail price system is shifting upward in sync, with extremely strong indicative significance for the industry’s price center. In other words, the on-demand price is not the discounted price in large customers’ long-term contracts, but the “listed price” of the computing power spot market. When the listed price rises overall, procurement costs for all new contracts, renewals, and small and medium-sized customers will rise accordingly.

Overseas price increases quickly transmitted to the domestic market. Domestic computing rental price increases have entered an “announcement verification stage”: Tianyang Technology’s 303 million yuan computing cloud service agreement was raised to 402 million yuan less than a month later; Xingyun Technology’s three contract orders increased by 201%, 79%, and 13% respectively; Saiyi Information’s two high-performance computing service contracts rose from 6.45 billion yuan to 6.72 billion yuan. Overseas on-demand prices, overseas long-term contract prices, and domestic contract prices rising simultaneously formed a rare “triple resonance” in the computing power rental market.

This round of price increases differs from past short-term speculation in that it is supported by real contracts and revenue recognition. In the past, computing power rental remained more at the level of “planned capacity” and “strategic cooperation,” making it hard for the market to verify actual utilization and price levels. Now, many listed companies have disclosed upward revisions in contract amounts through announcements, meaning price increases have entered a commercial closed loop: computing power operators can not only pass cost increases downstream, but also obtain excess pricing gains amid tight supply and demand. The price transmission chain has also become clear: GPU procurement costs, HBM and advanced packaging costs, power and cooling costs, rack and land costs, and O&M and depreciation costs ultimately all reflect in the per-hour unit price of computing power. When demand growth outpaces the pace of supply release, rental providers’ gross margins enter an upward channel.

II. Capital Frenzy: From “Land Grabbing” to “Refined Operations”

The capital market reacted sharply. In the computing power rental sector, Nanwei Software hit consecutive limit-ups, and Xingyun Technology’s stock price rose by as much as 586% during the year, earning it the name “Computing Power Order King.” Maixinlin, Chuangye Heima, Jiahua Technology and other targets also followed the rally. What capital is chasing is not a simple concept, but order visibility, rack-up rate, and price elasticity. Whoever has stable racks, power facilities, and orders in hand can seize the initiative in this round of price increases.

The deeper logic lies in a qualitative change in the industrial landscape. Data from the China Academy of Information and Communications Technology shows that China’s intelligent computing power rental market is expected to exceed 260 billion yuan in 2026. The computing power rental industry is shifting from extensive expansion of “land grabbing” to a structural leap toward “high-end premium and refined operations,” and the industrial landscape is accelerating from “a hundred flowers blooming” to a Matthew effect stage of concentration among leaders. Operating companies with stable racks, power facilities, and orders in hand are seeing their bargaining power substantially strengthen.

In the past, the competitive focus of computing power rental providers was “can we get the cards” and “can we build the data centers.” Today, the focus has shifted to “can we provide stable power,” “can we improve GPU utilization,” and “can we lock in long-term customers.” Computing power rental is no longer simple hardware subleasing, but a comprehensive service integrating power engineering, cooling and heat dissipation, network scheduling, cluster O&M, and financial solutions. The rack-up rate, failure rate, network latency, and scheduling efficiency of high-end GPU clusters directly determine rental providers’ profitability. The industry is moving from “selling computing power hours” to “selling stable services,” and from a “resource-based business” to an “operations-based business.”

At the same time, the valuation logic of the capital market is changing. In the early days, the market focused more on computing power rental companies’ planned capacity and GPU procurement volumes. Now it pays more attention to contract liabilities, advance receipts, orders in hand, and operating cash flow. Because only real revenue can prove that computing power is being consumed, and only stable collections can support subsequent capital expenditures. The rise of the computing power rental sector is essentially the market repricing “computing power pricing power.”

III. Supply-Demand Gap: Why Computing Power Has Become the “Scarcest Core Asset”

Behind the price increases is the acute contradiction between supply bottlenecks that are hard to resolve in the short term and explosive demand. GPU iteration cycles are long, advanced process capacity is limited, and model token call volumes and ARR (annual recurring revenue) are entering a stage of rapid growth. DeepSeek’s annualized revenue has exceeded $1 billion and it is using Huawei chips to train a 2-trillion-parameter large model; Alibaba’s Qwen 4 is being trained, with subsequent plans to evolve toward 5 trillion to 10 trillion parameters. Each order-of-magnitude increase in parameter scale drives an exponential rise in computing power demand.

From the demand side, the AI industry is experiencing a triple overlay: First, basic model training continues, with leading vendors pursuing larger parameters, longer context, and stronger multimodal capabilities. Second, inference demand is exploding rapidly, with token call volumes in scenarios such as chatbots, coding assistants, search Q&A, and intelligent customer service continuing to climb. Third, AI agents are beginning to execute tasks, moving from “answering questions” to “operating software, calling tools, and completing transactions.” Each task execution requires multiple rounds of inference and external tool calls, consuming far more computing power than traditional dialogue.

The explosion on the application side further amplifies demand. Meta’s personal AI assistant Muse reached 2.8 million downloads in 12 days after launch and is advancing integration with the Shopify payment ecosystem, promoting the commercialization of consumer-grade agents. AI is upgrading from “dialogue” to “task execution,” and demand for inference, tool calls, and cloud resources continues to expand. Goldman Sachs strategists expect the five largest U.S. cloud service providers to increase AI infrastructure spending by more than 50% next year, reaching $1.2 trillion. Under the supply-demand gap, computing power pricing power is turning from “narrative” into real “financial statement numbers.”

From the supply side, delivery cycles for high-end GPUs, HBM memory supply, CoWoS advanced packaging capacity, server assembly capacity, data center racks, and power quotas constitute multiple bottlenecks. Even if chip capacity increases, data center construction, power access, and cooling systems take time. Especially in first-tier cities and core computing hubs, energy consumption quotas and power allocations have become scarce resources. Computing power is not simply “buy cards and plug in power”; it is a systems engineering effort requiring land, power, network, cooling, and O&M. The pace of supply release is far behind the speed of demand explosion.

This supply-demand gap brings a direct result: computing power rental prices have strong upward elasticity. When the spot market is in short supply, on-demand prices rise first; when on-demand prices rise to a certain level, large customers’ long-term contract prices also shift upward; when long-term contract prices rise, computing power operators’ profit expectations improve, further driving capital expenditures and stock prices upward. Once this chain starts, it is difficult to reverse in the short term.

IV. Computing-Power-Electricity Coordination: The Underlying Constraint and New Solution for Computing Power Expansion

Computing power expansion does not lack a ceiling. The stability and cost of power supply are becoming core variables constraining the scale-up of computing power. Just this week, China’s first storage-computing integrated project officially started in Baotou, Inner Mongolia. Through coordinated construction of energy storage stations and distributed intelligent computing centers, it builds a green industrial closed loop of “input electricity, output tokens.” The project uses solid-state transformer technology to achieve efficient power conversion and precise voltage regulation, connecting the entire chain of energy supply, energy storage peak shaving, and computing power scheduling.

The significance of computing-power-electricity coordination is that it attempts to solve the most fundamental contradiction in the computing power industry: computing power demand fluctuates, power supply fluctuates, yet AI training and inference require relatively stable energy and computing power support. Through energy storage peak shaving, electricity can be stored during low-price periods and released during computing power peaks; through computing power scheduling, non-real-time tasks can be migrated to periods of surplus power; through green power direct connection, electricity costs can be reduced and carbon emission requirements met. If storage-computing integration can be replicated at scale, it will significantly improve the cost structure of computing power rental providers and also change the geographic distribution of computing hubs.

At the same time, on September 28, the State-owned Assets Supervision and Administration Commission of the State Council explicitly stated that during the “15th Five-Year Plan” period, central enterprises should continue to increase investment in the computing power network, deepen computing-power-electricity coordination and computing-network integration, and accelerate the implementation of more green power direct connection projects. Continuous policy support provides institutional guarantees for the long-term expansion of computing power infrastructure. The computing power network is no longer just corporate competition; it has been incorporated into the national new infrastructure system. Computing-power-electricity coordination, computing-network integration, and green power direct connection are becoming key words for the next stage of the computing power industry.

V. Risks and Concerns: Is the Price Increase Sustainable?

Although the computing power rental price increase trend is strong, the market still needs to be vigilant about several risks.

First, AI application revenue may fall short of expectations. If downstream AI companies cannot cover computing power costs through subscriptions, advertising, API calls, etc., computing power rental demand may slow temporarily. Especially for small and medium-sized AI startups, rising computing power costs will directly compress gross margins, forcing them to optimize models, reduce calls, or shift to cheaper chips.

Second, GPU technology iteration and depreciation risk. New-generation GPUs with improved performance and energy efficiency may pressure rents for older GPUs. If computing power rental providers purchase old cards at high prices, they may face accelerated depreciation and declining residual value. Lease terms, price lock mechanisms, and chip upgrade clauses will determine whether operators can weather the cycle.

Third, supply release and intensifying competition. As more capital enters the computing power rental market, rack and GPU supply may be released in a concentrated manner at some point in the future. If demand growth slows, prices may shift from soaring to falling. Computing power rental is not always scarce; structural oversupply and structural shortage may coexist.

Fourth, domestic substitution and technology route changes. Domestic chips such as Huawei Ascend and Cambricon continue to iterate, and domestic adaptation for inference scenarios is accelerating. If domestic chips achieve breakthroughs in cost-effectiveness and ease of use, some demand may shift from NVIDIA GPUs to domestic solutions, thereby easing pressure on high-end GPU rental prices.

Fifth, policy and regulatory risks. Factors such as dual control of energy consumption, data security, cross-border computing power services, and electricity price fluctuations may all affect the cost and compliance of computing power rental. Computing power operators need to balance expansion speed and compliance risk.

VI. Industry Implications: Who Benefits and How to Position

In this round of awakening of computing power pricing power, the beneficiaries are not only computing power rental providers. The upstream and downstream of the industrial chain will all be driven to varying degrees.

Computing power operators are the most direct beneficiaries. Enterprises with stable racks, power facilities, orders in hand, and customer resources can improve gross margins through price increases and improve cash flow through advance receipts. IDC companies benefit from higher rack-up rates and rents, especially data centers located in computing hubs with green power resources. Power and energy storage companies benefit from computing-power-electricity coordination investment, while liquid cooling, optical modules, servers, and network equipment companies benefit from cluster construction.

For enterprise customers, strategies need to be more refined. Enterprises with large long-term demand can lock in prices through long-term contracts to reduce spot price increase risk; enterprises with highly volatile inference demand can adopt hybrid clouds, placing stable loads on self-built or long-term contract computing power and peak loads on on-demand computing power; at the same time, they should accelerate model distillation, quantization, MoE architecture optimization, and domestic chip adaptation to reduce per-token computing power costs.

For investors, the valuation of the computing power rental sector should not only look at planned capacity, but also at rack-up rates, contract liabilities, customer concentration, power costs, and depreciation policies. Price increases are a revenue-side signal, but the profit side depends on cost control and operational efficiency. Only operators that truly convert GPUs into stable revenue can enjoy the dividend of pricing power.

Conclusion: The “Awakening Moment” of Computing Power Pricing Power

Nebius’s price increase announcement appears to be just a business decision, but it marks that the computing power market has entered a new stage: computing power is no longer just a “cost item,” but a core asset with independent pricing capability. As model parameters leap from trillions to tens of trillions, and as AI agents move from laboratories to tens of millions of users’ phones, the supply-demand gap in computing power will only widen. In this round of “awakening of pricing power,” computing power operating enterprises with stable racks, power facilities, and orders in hand will become the most direct beneficiaries of the AI era. For the entire industry, the continued rise in computing power rental prices may be just the first footnote in this infrastructure competition.