DeepSeek-R2 Model Accelerates Deployment and the Rise of Computing Power Leasing

How Low-Cost AI Reshapes Market Dynamics Amid Stable GPU Supply

In March 2025, Chinese AI company DeepSeek’s next-generation R2 model launched ahead of schedule, sending shockwaves through the industry. Simultaneously, the computing power leasing sector has experienced explosive growth due to cooling GPU market demand and shifting industry needs, becoming a pivotal variable in the AI supply chain. Falling prices for NVIDIA’s H200 chips, declining resale values of used H100 GPUs, and delays in GB200 mass production have driven enterprises to transition from stockpiling GPUs to flexible leasing models, forming a dual "model optimization + computing power sharing" techno-economic paradigm.

I. Technological Breakthrough: Synergy Between Low-Cost Models and GPU Leasing

The core competitiveness of DeepSeek-R2 lies in its “low training costs + high-performance inference”, enabled by two key trends:

1. Stable GPU Supply and Leasing Adoption: NVIDIA’s H200 chips are now priced at 2.4 million yuan through distributors, with delivery times shortened to 3-4 weeks. As used H100 prices continue to drop, SMEs and startups increasingly opt for GPU leasing to reduce upfront costs.

2. Hardware-Adaptive Model Architecture: R2 reduces reliance on high-end GPUs through parameter optimization and distillation technology, supporting localized deployment. For instance, its lightweight models run on edge devices, complementing RTX 3090/4090 GPU leasing services on platforms like Xianyu, which cater to AI image generation and lightweight inference needs.

This approach not only alleviates dependence on expensive GPUs but also fuels a sharing economy for computing resources. Xianyu reports a 300% surge in searches for high-end GPU rentals over three months, with daily rates as low as 30 yuan, underscoring demand for flexible computing solutions.

II. Market Polarization: Capital Bets on Leasing and Vertical Applications

The capital market reflects a split focus:

• Boom in Computing Power Leasing: China’s GPU leasing market is projected to reach 260 billion yuan, growing over 20% annually. Listed firms like Lotus Holdings and Maxscend are building AI data centers or adopting financial leasing models, securing contracts exceeding 500 million yuan. Globally, AI cloud giant CoreWeave (CRWV.US) targets a $35 billion IPO valuation, with its 250,000-GPU data center network setting industry benchmarks.

• Cooling GPU Infrastructure Investments: Wall Street is reducing exposure to chip giants like NVIDIA, pivoting to vertical AI applications. The 2% rise in the SSE STAR Market AI ETF (588930) signals market confidence in the “leasing + application” model over pure hardware investments.

This shift highlights an industry consensus: AI computing demand is migrating from centralized training to distributed inference. Omdia predicts global server capital expenditure will hit $280 billion in 2025, with inference workloads driving demand for ASIC chips and elastic leasing services.

III. Challenges: Financial Risks and Ethical Dilemmas in GPU Leasing

Despite lowering entry barriers, the leasing model faces unresolved issues:

1. Funding Pressures and Hardware Risks: Companies like Lotus Purple Star face a 780-million-yuan funding gap for GPU procurement, relying on credit lines and financial leasing. Rapid GPU obsolescence and maintenance costs threaten profitability.

2. Regulatory Fragmentation: Leasing providers must navigate policies like the EU’s AI Act and China’s proposed AI governance rules, requiring strict data compliance and privacy safeguards.

Additionally, open-source tensions persist. While DeepSeek’s open-source strategy is praised, misuse risks—such as AI-generated disinformation—could escalate if models spread via leasing platforms.

IV. Future Trends: GPU Leasing Ecosystems and Domestic Substitution

Three key developments will shape 2025’s AI computing landscape:

1. Hybrid Computing Models: Enterprises will blend in-house clusters with leased resources. Honor’s $10 billion AI terminal network and Xiaomi’s automotive AI expansion rely on dynamic GPU allocation.

2. Domestic GPU Adoption: Local chips like Zhaoxin and Biren gain traction. Despite NVIDIA’s dominance in high-end leasing, policy-backed domestic alternatives will thrive in low-latency sectors like government and education.

3. Efficiency Innovations: Copper cabling, silicon photonics, and glass substrates optimize computing efficiency. Copper connections, boosted by GB200 production, are projected to grow at 25% CAGR, cutting data center costs.

Commentary: Leasing Reshapes AI Value Chains, But Beware of Bubbles

The synergy between DeepSeek-R2 and GPU leasing marks a shift from “heavy-asset races” to “asset-light services”. Yet, risks loom:

• Speculative Hype: Some firms exploit leasing as a buzzword, with idle data centers exposing supply-demand mismatches.

• Overreliance on NVIDIA: Supply chain vulnerabilities necessitate stronger domestic chip ecosystems and open-source collaboration.

Ultimately, balancing innovation, capital, and ethics will define success. As CoreWeave’s IPO demonstrates, only players integrating “computing power, models, and scenarios” will thrive in this transformative era.

--The above article is purely AI-generated. If there is any infringement, please contact us for deletion!