Era of Compute Shortage: Why Enterprises Are Rushing to Rent NVIDIA Blackwell Compute Power

The global artificial intelligence industry is currently experiencing an unprecedented era of compute shortage. The exponential growth of large language model parameters, the large‑scale deployment of AI agents, and the explosion of multimodal applications have jointly driven a wave of demand far beyond expectations. Against this backdrop, compute leasing has become the fastest and most economical way for enterprises to obtain high‑performance GPU compute power, evolving from a “nice‑to‑have” into a “must‑have.” Whether you are an AI startup, a traditional enterprise undergoing digital transformation, or a large internet platform, you are actively seeking stable, flexible, and cost‑effective compute leasing solutions.

I. The Reality of Compute Shortage: NVIDIA GB300/B300 “Impossible to Find”

In June 2026, the alarm for compute shortage has sounded again. According to the latest report from Wedbush Securities, demand for NVIDIA’s Grace Blackwell systems (including GB300 and B300) continues to soar, delivery lead times are extending, and availability is declining. Analysts Matt Bryson and Antoine Legault noted: “We don’t remember a supply tightness this severe in the late stages of an NVIDIA accelerator cycle since Ampere/Hopper, suggesting demand is growing faster than previously expected.” It is reported that some customers’ Blackwell orders have been pushed back to the first quarter of 2027, with significant premiums over the official guide prices.

Multiple factors have contributed to the supply tightness. On one hand, global AI infrastructure construction is growing much faster than expected – hyperscale cloud providers such as Microsoft, Google, Amazon, and Meta increased their combined capital expenditures by more than 35% year‑over‑year in the first quarter of 2026, with most of the spending going into AI servers and data centers. On the other hand, high‑bandwidth memory (HBM) and advanced packaging capacity, which are critical for the Blackwell systems, remain key supply chain bottlenecks. HBM3E and HBM4 production yields are still ramping up, and TSMC’s CoWoS advanced packaging capacity, although expanding, remains in short supply. Nevertheless, NVIDIA, with its strategic move to lock in DRAM and HBM supply early, is “in the best position across the entire technology ecosystem.”

From an institutional forecast perspective, NVIDIA’s market dominance continues to solidify. TrendForce predicts that in 2026, Blackwell solutions will account for 71% of NVIDIA’s high‑end GPU shipments. Meanwhile, the next‑generation Rubin architecture faces challenges such as HBM4 certification and power management, and is expected to start shipping only in the second half of 2026. This means that for at least the next 6‑9 months, the Blackwell family will continue to dominate the high‑end AI compute market, and the supply‑demand imbalance will not ease quickly.

At the same time, the U.S. Department of Commerce upgraded its AI chip export controls on June 1, 2026, closing the previous “loophole” that allowed Chinese companies to purchase through overseas subsidiaries. The new guidelines explicitly state that even if an entity is located outside China, as long as its headquarters is in China, it must still comply with export licensing requirements for advanced AI chips, including NVIDIA Rubin, Blackwell, and AMD MI350X. This policy has further increased the difficulty for domestic Chinese enterprises to obtain high‑end overseas compute power, sharply boosting the demand for domestic AI chip alternatives and local compute leasing services.

II. Compute Leasing Prices Continue to Rise: A Seller’s Market Has Arrived

The compute shortage is directly reflected in market prices. A research report from Guojin Securities shows that the on‑demand leasing price of NVIDIA B300 has surged by 105% compared to November 2025, while the leasing price of H100 has also increased by about 25%. Specifically, the hourly leasing price for a single B300 GPU has exceeded US$\mathrm{<span\; style="color:\#FFFFFF;">4.5</span>}<span\; style="color:\#FFFFFF;">,</span>\mathrm{<span\; style="color:\#FFFFFF;">w</span>}\mathrm{<span\; style="color:\#FFFFFF;">h</span>}\mathrm{<span\; style="color:\#FFFFFF;">i</span>}\mathrm{<span\; style="color:\#FFFFFF;">l</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">t</span>}\mathrm{<span\; style="color:\#FFFFFF;">h</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">u</span>}\mathrm{<span\; style="color:\#FFFFFF;">n</span>}\mathrm{<span\; style="color:\#FFFFFF;">i</span>}\mathrm{<span\; style="color:\#FFFFFF;">t</span>}\mathrm{<span\; style="color:\#FFFFFF;">p</span>}\mathrm{<span\; style="color:\#FFFFFF;">r</span>}\mathrm{<span\; style="color:\#FFFFFF;">i</span>}\mathrm{<span\; style="color:\#FFFFFF;">c</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">f</span>}\mathrm{<span\; style="color:\#FFFFFF;">o</span>}\mathrm{<span\; style="color:\#FFFFFF;">r</span>}\mathrm{<span\; style="color:\#FFFFFF;">a</span>}\mathrm{<span\; style="color:\#FFFFFF;">o</span>}\mathrm{<span\; style="color:\#FFFFFF;">n</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\text{<span style="color:\#FFFFFF;">‑</span>}\mathrm{<span\; style="color:\#FFFFFF;">y</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">a</span>}\mathrm{<span\; style="color:\#FFFFFF;">r</span>}\mathrm{<span\; style="color:\#FFFFFF;">l</span>}\mathrm{<span\; style="color:\#FFFFFF;">o</span>}\mathrm{<span\; style="color:\#FFFFFF;">n</span>}\mathrm{<span\; style="color:\#FFFFFF;">g</span>}\text{<span style="color:\#FFFFFF;">‑</span>}\mathrm{<span\; style="color:\#FFFFFF;">t</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">r</span>}\mathrm{<span\; style="color:\#FFFFFF;">m</span>}\mathrm{<span\; style="color:\#FFFFFF;">l</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">a</span>}\mathrm{<span\; style="color:\#FFFFFF;">s</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">h</span>}\mathrm{<span\; style="color:\#FFFFFF;">a</span>}\mathrm{<span\; style="color:\#FFFFFF;">s</span>}\mathrm{<span\; style="color:\#FFFFFF;">r</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">a</span>}\mathrm{<span\; style="color:\#FFFFFF;">c</span>}\mathrm{<span\; style="color:\#FFFFFF;">h</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">d</span>}\mathrm{<span\; style="color:\#FFFFFF;">a</span>}\mathrm{<span\; style="color:\#FFFFFF;">b</span>}\mathrm{<span\; style="color:\#FFFFFF;">o</span>}\mathrm{<span\; style="color:\#FFFFFF;">v</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">U</span>}\mathrm{<span\; style="color:\#FFFFFF;">S</span>}$3.2 per hour. Data released by semiconductor research firm SemiAnalysis confirms the same trend – the one‑year GPU lease contract price for H100 has soared from a low of about US$\mathrm{<span\; style="color:\#FFFFFF;">1.70</span>}\mathrm{<span\; style="color:\#FFFFFF;">p</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">r</span>}\mathrm{<span\; style="color:\#FFFFFF;">G</span>}\mathrm{<span\; style="color:\#FFFFFF;">P</span>}\mathrm{<span\; style="color:\#FFFFFF;">U</span>}\mathrm{<span\; style="color:\#FFFFFF;">p</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">r</span>}\mathrm{<span\; style="color:\#FFFFFF;">h</span>}\mathrm{<span\; style="color:\#FFFFFF;">o</span>}\mathrm{<span\; style="color:\#FFFFFF;">u</span>}\mathrm{<span\; style="color:\#FFFFFF;">r</span>}\mathrm{<span\; style="color:\#FFFFFF;">i</span>}\mathrm{<span\; style="color:\#FFFFFF;">n</span>}\mathrm{<span\; style="color:\#FFFFFF;">O</span>}\mathrm{<span\; style="color:\#FFFFFF;">c</span>}\mathrm{<span\; style="color:\#FFFFFF;">t</span>}\mathrm{<span\; style="color:\#FFFFFF;">o</span>}\mathrm{<span\; style="color:\#FFFFFF;">b</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">r</span>}\mathrm{<span\; style="color:\#FFFFFF;">2025</span>}\mathrm{<span\; style="color:\#FFFFFF;">t</span>}\mathrm{<span\; style="color:\#FFFFFF;">o</span>}\mathrm{<span\; style="color:\#FFFFFF;">U</span>}\mathrm{<span\; style="color:\#FFFFFF;">S</span>}$2.35 per GPU per hour in March 2026, an increase of nearly 40%. By the end of May 2026, spot leasing prices for H100 on some compute platforms had even touched US$2.80.

More notably, the tightness in the compute market is not a short‑term phenomenon. As delivery lead times for new Blackwell, H100, and H200 clusters lengthen (pushed into 2027), delivery bottlenecks will continue to push prices upward. From a total cost of ownership (TCO) perspective, given the current price expectations where a single GPU is hard to find, compute leasing solves the entire compute demand – from R&D testing to large‑scale inference – at a fraction of the upfront cost of a one‑time purchase, and its economic advantages are becoming increasingly evident. For example, the upfront purchase cost of an eight‑GPU H100 server is approximately US$\mathrm{<span\; style="color:\#FFFFFF;">300</span>}<span\; style="color:\#FFFFFF;">,</span>\mathrm{<span\; style="color:\#FFFFFF;">000</span>}\text{<span style="color:\#FFFFFF;">‑</span>}\mathrm{<span\; style="color:\#FFFFFF;">350</span>}<span\; style="color:\#FFFFFF;">,</span>\mathrm{<span\; style="color:\#FFFFFF;">000</span>}<span\; style="color:\#FFFFFF;">,</span>\mathrm{<span\; style="color:\#FFFFFF;">w</span>}\mathrm{<span\; style="color:\#FFFFFF;">h</span>}\mathrm{<span\; style="color:\#FFFFFF;">i</span>}\mathrm{<span\; style="color:\#FFFFFF;">l</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">t</span>}\mathrm{<span\; style="color:\#FFFFFF;">h</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">t</span>}\mathrm{<span\; style="color:\#FFFFFF;">o</span>}\mathrm{<span\; style="color:\#FFFFFF;">t</span>}\mathrm{<span\; style="color:\#FFFFFF;">a</span>}\mathrm{<span\; style="color:\#FFFFFF;">l</span>}\mathrm{<span\; style="color:\#FFFFFF;">c</span>}\mathrm{<span\; style="color:\#FFFFFF;">o</span>}\mathrm{<span\; style="color:\#FFFFFF;">s</span>}\mathrm{<span\; style="color:\#FFFFFF;">t</span>}\mathrm{<span\; style="color:\#FFFFFF;">o</span>}\mathrm{<span\; style="color:\#FFFFFF;">f</span>}\mathrm{<span\; style="color:\#FFFFFF;">a</span>}\mathrm{<span\; style="color:\#FFFFFF;">o</span>}\mathrm{<span\; style="color:\#FFFFFF;">n</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\text{<span style="color:\#FFFFFF;">‑</span>}\mathrm{<span\; style="color:\#FFFFFF;">y</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">a</span>}\mathrm{<span\; style="color:\#FFFFFF;">r</span>}\mathrm{<span\; style="color:\#FFFFFF;">l</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">a</span>}\mathrm{<span\; style="color:\#FFFFFF;">s</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">f</span>}\mathrm{<span\; style="color:\#FFFFFF;">o</span>}\mathrm{<span\; style="color:\#FFFFFF;">r</span>}\mathrm{<span\; style="color:\#FFFFFF;">t</span>}\mathrm{<span\; style="color:\#FFFFFF;">h</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">s</span>}\mathrm{<span\; style="color:\#FFFFFF;">a</span>}\mathrm{<span\; style="color:\#FFFFFF;">m</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">c</span>}\mathrm{<span\; style="color:\#FFFFFF;">o</span>}\mathrm{<span\; style="color:\#FFFFFF;">n</span>}\mathrm{<span\; style="color:\#FFFFFF;">f</span>}\mathrm{<span\; style="color:\#FFFFFF;">i</span>}\mathrm{<span\; style="color:\#FFFFFF;">g</span>}\mathrm{<span\; style="color:\#FFFFFF;">u</span>}\mathrm{<span\; style="color:\#FFFFFF;">r</span>}\mathrm{<span\; style="color:\#FFFFFF;">a</span>}\mathrm{<span\; style="color:\#FFFFFF;">t</span>}\mathrm{<span\; style="color:\#FFFFFF;">i</span>}\mathrm{<span\; style="color:\#FFFFFF;">o</span>}\mathrm{<span\; style="color:\#FFFFFF;">n</span>}\mathrm{<span\; style="color:\#FFFFFF;">i</span>}\mathrm{<span\; style="color:\#FFFFFF;">s</span>}\mathrm{<span\; style="color:\#FFFFFF;">a</span>}\mathrm{<span\; style="color:\#FFFFFF;">b</span>}\mathrm{<span\; style="color:\#FFFFFF;">o</span>}\mathrm{<span\; style="color:\#FFFFFF;">u</span>}\mathrm{<span\; style="color:\#FFFFFF;">t</span>}\mathrm{<span\; style="color:\#FFFFFF;">U</span>}\mathrm{<span\; style="color:\#FFFFFF;">S</span>}$160,000‑200,000, saving nearly 40% of the initial investment, without the need to bear additional expenses such as equipment depreciation, electricity, cooling, and maintenance.

Moreover, price differentiation in the leasing market is intensifying. On‑demand leasing prices are significantly higher than long‑term reserved instances, but many enterprises are still willing to pay a premium for flexibility because they cannot predict their compute demand for the coming months. Some compute leasing platforms have also introduced “hybrid leasing” models that combine on‑demand and reserved instances to help enterprises balance cost and flexibility.

III. Structural Growth in Compute Demand: A Trillion‑RMB Market on the Verge of Explosion

What is driving such strong demand for compute? The answer lies in the explosive growth of token usage. Public data shows that China’s average daily token calls have exceeded 140 trillion, a more than 40% increase from the end of 2025. From 100 billion in early 2024, to 100 trillion by the end of 2025, and now to the 140 trillion level, the steepness of the curve clearly indicates that AI applications are moving from demos and trials to large‑scale commercial operations. Take leading models such as ByteDance’s Doubao, Baidu’s Wenxin Yiyan, and Alibaba’s Tongyi Qianwen as examples – their API call volumes have grown several times in the past six months, directly driving an exponential increase in back‑end compute demand.

A research report from Dongwu Securities points out that in Q1 2026, top cloud vendors and major model developers have been bidding up prices to acquire high‑end compute power held by compute leasing companies. Only the leading cloud vendors (such as Alibaba Cloud, Huawei Cloud, and Tencent Cloud) are able to secure relatively adequate high‑end compute; second‑tier cloud vendors and model developers face far from satisfied demand, with a huge gap remaining. It is estimated that the total domestic shortfall in high‑end AI compute capacity exceeds 50 EFLOPS (exaflops), equivalent to the compute power of about 300,000 H100 GPUs.

Behind this supply‑demand imbalance are several technical realities: Chinese LLMs are transitioning from hundreds of billions of parameters to trillion‑parameter MoE architectures; multimodal AI consumes 7‑15 times more compute than traditional unimodal models; and the global AI compute infrastructure market is rapidly breaking through the trillion‑dollar mark. Specifically, OpenAI’s GPT‑5 is rumored to adopt a MoE architecture, where the activated parameters are much smaller than the total parameters, but training it still requires an extremely large‑scale compute cluster. Meanwhile, video generation models such as Sora and Kling consume thousands of times more compute for a single inference than text‑based models. Once these applications are deployed at scale, they will place enormous pressure on compute supply.

According to the latest forecast from IDC, the global AI compute market will reach US$\mathrm{<span\; style="color:\#FFFFFF;">120</span>}\mathrm{<span\; style="color:\#FFFFFF;">b</span>}\mathrm{<span\; style="color:\#FFFFFF;">i</span>}\mathrm{<span\; style="color:\#FFFFFF;">l</span>}\mathrm{<span\; style="color:\#FFFFFF;">l</span>}\mathrm{<span\; style="color:\#FFFFFF;">i</span>}\mathrm{<span\; style="color:\#FFFFFF;">o</span>}\mathrm{<span\; style="color:\#FFFFFF;">n</span>}\mathrm{<span\; style="color:\#FFFFFF;">i</span>}\mathrm{<span\; style="color:\#FFFFFF;">n</span>}\mathrm{<span\; style="color:\#FFFFFF;">2026</span>}<span\; style="color:\#FFFFFF;">,</span>\mathrm{<span\; style="color:\#FFFFFF;">o</span>}\mathrm{<span\; style="color:\#FFFFFF;">f</span>}\mathrm{<span\; style="color:\#FFFFFF;">w</span>}\mathrm{<span\; style="color:\#FFFFFF;">h</span>}\mathrm{<span\; style="color:\#FFFFFF;">i</span>}\mathrm{<span\; style="color:\#FFFFFF;">c</span>}\mathrm{<span\; style="color:\#FFFFFF;">h</span>}\mathrm{<span\; style="color:\#FFFFFF;">t</span>}\mathrm{<span\; style="color:\#FFFFFF;">h</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">c</span>}\mathrm{<span\; style="color:\#FFFFFF;">o</span>}\mathrm{<span\; style="color:\#FFFFFF;">m</span>}\mathrm{<span\; style="color:\#FFFFFF;">p</span>}\mathrm{<span\; style="color:\#FFFFFF;">u</span>}\mathrm{<span\; style="color:\#FFFFFF;">t</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">l</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">a</span>}\mathrm{<span\; style="color:\#FFFFFF;">s</span>}\mathrm{<span\; style="color:\#FFFFFF;">i</span>}\mathrm{<span\; style="color:\#FFFFFF;">n</span>}\mathrm{<span\; style="color:\#FFFFFF;">g</span>}<span\; style="color:\#FFFFFF;">(</span>\mathrm{<span\; style="color:\#FFFFFF;">G</span>}\mathrm{<span\; style="color:\#FFFFFF;">P</span>}\mathrm{<span\; style="color:\#FFFFFF;">U</span>}\mathrm{<span\; style="color:\#FFFFFF;">a</span>}\mathrm{<span\; style="color:\#FFFFFF;">s</span>}\mathrm{<span\; style="color:\#FFFFFF;">a</span>}\mathrm{<span\; style="color:\#FFFFFF;">S</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">r</span>}\mathrm{<span\; style="color:\#FFFFFF;">v</span>}\mathrm{<span\; style="color:\#FFFFFF;">i</span>}\mathrm{<span\; style="color:\#FFFFFF;">c</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}<span\; style="color:\#FFFFFF;">)</span>\mathrm{<span\; style="color:\#FFFFFF;">s</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">g</span>}\mathrm{<span\; style="color:\#FFFFFF;">m</span>}\mathrm{<span\; style="color:\#FFFFFF;">e</span>}\mathrm{<span\; style="color:\#FFFFFF;">n</span>}\mathrm{<span\; style="color:\#FFFFFF;">t</span>}\mathrm{<span\; style="color:\#FFFFFF;">w</span>}\mathrm{<span\; style="color:\#FFFFFF;">i</span>}\mathrm{<span\; style="color:\#FFFFFF;">l</span>}\mathrm{<span\; style="color:\#FFFFFF;">l</span>}\mathrm{<span\; style="color:\#FFFFFF;">a</span>}\mathrm{<span\; style="color:\#FFFFFF;">c</span>}\mathrm{<span\; style="color:\#FFFFFF;">c</span>}\mathrm{<span\; style="color:\#FFFFFF;">o</span>}\mathrm{<span\; style="color:\#FFFFFF;">u</span>}\mathrm{<span\; style="color:\#FFFFFF;">n</span>}\mathrm{<span\; style="color:\#FFFFFF;">t</span>}\mathrm{<span\; style="color:\#FFFFFF;">f</span>}\mathrm{<span\; style="color:\#FFFFFF;">o</span>}\mathrm{<span\; style="color:\#FFFFFF;">r</span>}\mathrm{<span\; style="color:\#FFFFFF;">a</span>}\mathrm{<span\; style="color:\#FFFFFF;">b</span>}\mathrm{<span\; style="color:\#FFFFFF;">o</span>}\mathrm{<span\; style="color:\#FFFFFF;">u</span>}\mathrm{<span\; style="color:\#FFFFFF;">t</span>}\mathrm{<span\; style="color:\#FFFFFF;">35</span>}$42 billion. In the Chinese market, the compute leasing scale is expected to exceed RMB 80 billion, with an annual growth rate of over 60%.

IV. How Does Compute Leasing Solve Enterprise Pain Points? Six Core Advantages Explained in Detail

Against the backdrop of extreme supply tightness, compute leasing has become the preferred choice for more and more enterprises, thanks to its unique value proposition. The following six core advantages deserve attention:

1. Solving supply‑demand mismatches and gaining access to scarce compute quickly
High‑end GPUs are extremely hard to find, and procurement lead times can stretch to months or even a year. Compute leasing platforms, however, typically have pre‑locked inventory and supply chain resources, enabling delivery within days or even on the same day, allowing enterprises to start their AI projects immediately. For teams with urgent training tasks or product launch schedules, this is a lifeline.

2. Higher capital efficiency – shifting from CAPEX to OPEX
The upfront purchase cost of a fully loaded H100 server exceeds US$300,000, a heavy capital burden for small and medium‑sized enterprises and startups. Compute leasing breaks this large one‑time expense into manageable monthly or hourly payments, allowing enterprises to retain capital for core R&D and market expansion. From a financial perspective, the tax treatment of operating expenses is often more favorable than asset depreciation.

3. Elastic scaling to cope with business fluctuations
The compute demand of AI projects often shows a peak‑and‑valley pattern – model training requires a large amount of intensive compute, while inference is relatively stable but continuous. Compute leasing allows enterprises to scale up or down on demand: spinning up a thousand‑GPU cluster for training and releasing the resources after training is complete, paying only for what they actually use. This elasticity is difficult to achieve with an on‑premises data center.

4. Avoiding technology obsolescence and always using the latest GPU architectures
GPU iteration is extremely fast, with new generations emerging every 1.5‑2 years on average. Hardware purchased in one go may become obsolete or fall short of mainstream requirements within 2‑3 years, with second‑hand residual values dropping sharply. With compute leasing, enterprises can seamlessly switch to the latest Blackwell, Rubin, or even future architectures, maintaining technological competitiveness at all times.

5. Lowering operational barriers and focusing on AI applications
Building an on‑premises GPU cluster requires dedicated teams for hardware maintenance, network configuration, cooling management, power protection, and more – a heavy burden for companies that are not infrastructure‑focused. Compute leasing platforms provide a full suite of managed services, including 24/7 monitoring, automatic failure recovery, and data backup. Enterprises only need to submit jobs via an API or console.

6. A new option for domestic compute leasing – balancing performance and compliance
As U.S. export controls continue to tighten, domestic AI chips – such as Huawei Ascend, Hygon DCU, Cambricon, and Enflame – are rapidly gaining ground. Many domestic compute leasing platforms have started offering compute instances based on Ascend 910B/950PR, whose performance in mainstream LLM training and inference scenarios is approaching or even surpassing (on certain operators) that of NVIDIA A100/H100 to a certain degree. Choosing domestic compute leasing not only avoids supply chain risks but also meets regulatory compliance requirements, making it particularly suitable for customers in critical sectors such as government, finance, and energy.

V. Accelerating Domestic Substitution: A Strategic Opportunity for Domestic Compute Leasing

Against the backdrop of tightening U.S. export controls on AI chips, domestic AI chips are ushering in a historic window of opportunity. Public data shows that in 2025, domestic manufacturers shipped approximately 1.65 million AI accelerator cards in China, capturing over 40% of the market share for the first time. Huawei Ascend shipped about 812,000 units in 2025, and the Ascend 950PR launched in 2026 is benchmarked against and surpasses NVIDIA H200 (for certain operators). Hygon’s Deep Computing二号 (Deep Computing二号) DCU has also made significant progress in commercial applications, with server vendors such as Lenovo and Inspur launching AI servers based on Hygon DCU.

Institutions predict that domestic AI chip shipments could more than double in 2026. At the same time, the domestic compute leasing market is rapidly shifting toward “domestication.” Local governments in core regions such as Beijing, Shanghai, Shenzhen, and Chengdu have rolled out compute vouchers and subsidies to encourage enterprises to use domestic compute. For example, the “AI Compute Voucher” policy introduced by the Beijing Municipal Bureau of Economy and Information Technology provides up to 30% subsidies for enterprises leasing local domestic compute, with an annual cap of RMB 5 million per enterprise.

Several leading compute leasing platforms – including Parallel Technology, QingCloud, UCloud, and Capital Online – have all launched compute leasing products based on Huawei Ascend. Some platforms have even implemented “heterogeneous compute pools” that mix NVIDIA GPUs with domestic GPUs, allowing users to choose the optimal compute resources based on task characteristics. With the continuous improvement of the Ascend ecosystem (CANN software stack, PyTorch adaptation, model migration tools), the usability and ease‑of‑use of domestic compute leasing are rapidly advancing, and it is expected to become a major pillar of the compute leasing market within the next 1‑2 years.

VI. Real‑World Case Studies: How Compute Leasing Helps Different Types of Enterprises Succeed

To illustrate the practical value of compute leasing more concretely, here are three typical case studies from different types of enterprises:

Case 1: LLM startup – quickly validating a hundred‑billion‑parameter model
A Beijing‑based AI startup planned to train a hundred‑billion‑parameter foundation model from scratch within three months. Unable to procure enough H800 GPUs through conventional channels, the company leased 64 eight‑GPU H800 servers (512 H800 GPUs in total) for a period of two months through a compute leasing platform. The platform completed environment setup and network debugging within five working days. In the end, the training task was completed one week ahead of schedule, and the total leasing cost was about one‑fifth of the cost of purchasing equivalent hardware. After the project ended, the compute resources were released with no idle capacity concerns.

Case 2: Traditional manufacturing enterprise – deploying an AI visual inspection system
An auto parts manufacturer wanted to deploy an AI visual inspection system based on a vision foundation model, but building its own GPU servers would have been expensive and time‑consuming. Instead, it chose to rent inference‑optimized GPU instances based on T4 and A10, paying by the hour. Initially, it rented just two GPUs for model fine‑tuning; after going live, it elastically scaled up to 20 GPUs based on the number of production lines. Total upfront investment was less than RMB 50,000, and the average daily inference cost was kept below RMB 2,000. Compared to a purchase‑based approach, this saved 80% of the initial investment.

Case 3: University research team – using domestic compute for academic research
A natural language processing lab at a top‑tier university needed to train a Chinese multimodal model based on the Ascend platform, but the lab lacked physical Ascend hardware. Through a compute leasing platform, the lab rented four Ascend 910B servers for three months and received a 30% cost reduction via a local compute voucher. The team successfully completed the experiment, published a paper at a top international conference, and gained valuable experience in tuning the Ascend platform.

VII. Future Outlook: Compute Leasing as the Core Form of AI Infrastructure

Looking at the booming development of leading models in both China and the U.S., the simultaneous rise of domestic AI chips, the global shortage of high‑end NVIDIA computing capacity, and the intensive signing of cross‑industry compute contracts in China – no matter how the external environment changes, the foundational value of compute as the “water, electricity, and coal” of the AI industry is being continuously verified and amplified. In the coming years, compute leasing is expected to exhibit the following trends:

• Consolidation and platformization: Large compute leasing platforms will integrate scattered small and medium‑sized compute resources, providing unified APIs, unified scheduling, and unified operations to lower the barrier to entry for enterprises.

• Hybrid compute becomes the norm: NVIDIA GPUs will be used for training and part of inference, while domestic GPUs will be used for inference and compliance‑sensitive scenarios. Hybrid leasing will become a standard strategy for large enterprises.

• Compute securitization and futures models: Drawing on the reserved instance model of cloud computing, compute leasing platforms will introduce more flexible “compute futures,” allowing enterprises to lock in prices for the next six months or a year in advance, hedging against market volatility.

• Green compute gains favor: Driven by the “dual carbon” policy, compute leasing products that use liquid‑cooled data centers or clean energy power will command a premium and benefit from subsidies.

Conclusion: Act Fast to Seize the Scarcity Compute Advantage

Compute has become the core resource that determines AI competitiveness. In the current environment of supply tightness, rising prices, and accelerating domestic substitution, enterprises need to formulate their compute acquisition strategies as soon as possible. Whether you are facing development bottlenecks in training large AI models or planning to deploy AI inference applications at scale, starting with a single cost‑effective GPU server and scaling up to thousand‑GPU or ten‑thousand‑GPU compute clusters, compute leasing provides a one‑stop compute solution.

High‑performance GPU compute is ready to use, with professional teams providing full technical support, flexible billing models (hourly, daily, monthly, yearly), and a wide selection of instance types (H100, A100, Ascend 910B, B300, etc.) to help accelerate your AI project. Contact us now to get a customized compute leasing plan and take the lead in the next wave of AI innovation.