Behind the Spring Festival AI Red Envelope Battle: The Supply-Demand Game of Computing Power with H100 and H200 in Shortage

The price of an H200 server has rebounded from 1.9 million to 2.3 million yuan. DDR5 memory module prices have skyrocketed by over 300%, and the instantaneous traffic surge triggered by the AI red envelope battle on Lunar New Year's Eve might be nearly 40% higher than in previous years.

During the Spring Festival, Tencent's Yuanbao APP experienced brief service instability due to a sudden traffic surge. Simultaneously, in Shenzhen, computing power leasing professionals were carrying cash from place to place, searching for available H200 units, only to find a scattered handful. They described the current computing power market as operating in "hell mode."

This is the real picture of China's AI computing power market just before the 2026 Spring Festival. On one side, AI companies are vigorously engaged in Spring Festival red envelope marketing campaigns, with investments exceeding 4.5 billion yuan. On the other side, there is a severe supply crunch where high-end computing chips like the H100 and H200 are nearly impossible to procure.

01 Spring Festival Effect: AI Marketing Wars Ignite Computing Power Demand

During the 2026 Year of the Horse Spring Festival, Chinese internet giants engaged in a fierce battle centered on AI red envelopes. Companies like Tencent, Alibaba, Baidu, and ByteDance invested massive funds, competing fiercely for AI user touchpoints and attention through tactics like distributing red envelopes, sponsoring Spring Festival Gala events, and launching new holiday-themed features.

Following the launch of its Spring Festival campaign, Tencent Yuanbao quickly rose to the top of the free app rankings on the Apple App Store. Alibaba's AI platform, Qianwen, announced an investment of 30 billion yuan, partnering with various Alibaba ecosystem services like Taobao Flash Sales, Fliggy, and Freshippo to offer users "free purchase" promotions.

The massive influx of users into these AI applications created unprecedented concurrent access pressure. According to industry insiders, this year's Spring Festival period might see a nearly 40% year-on-year increase in peak instantaneous traffic concurrency.

The Spring Festival is not only a peak period for user traffic but also a concentrated release window for domestic AI models. As the holiday approached, a new wave of domestic model releases began, with several companies planning to launch significant AI model updates.

To cope with the sharply rising traffic demands during the Spring Festival, several cloud service providers and AI companies have leased substantial amounts of temporary bandwidth from data centers.

02 Supply Dilemma: The "Invisible" High-End GPUs and Soaring Prices

At the other end of the surging demand, the supply of high-end computing chips has hit a serious bottleneck. Top-tier GPU chips like NVIDIA's B200 are virtually "invisible" in the domestic market, while prices for mid-to-low-end GPUs have also surged due to rising raw material costs.

Although the H200 has obtained an export license, the actual import approval process remains stringent, leaving the domestic market in a state where the product is "seen but not yet readily accessible." This situation has forced cloud service providers like UCloud to proactively procure chips from 3 to 5 different domestic manufacturers in advance to diversify their supply sources.

Even more startling is the volatility in the storage market. Since September 2025, prices for DDR5 memory modules have surged by over 300%. A single 64GB DDR5 memory module that cost just over 2,000 yuan in July-August last year now costs over 10,000 yuan.

The influx of speculative capital has intensified the market frenzy. Some distributors who previously focused on graphics cards have switched to trading memory, employing a "circular self-reinforcing" tactic—buying low, selling high, then buying back at even higher prices, continuously pushing market prices upward.

03 Chain Reaction: Comprehensive Price Hikes from Hardware to Services

Rising hardware costs have been passed on to the prices of complete server systems and service fees. The price of a complete H200 server system once dropped to a low of around 1.9 million yuan but has now rebounded to over 2.3 million yuan according to latest quotes.

Configuring a single B200 server could incur memory costs alone exceeding 300,000 yuan. Factoring in the CPU, hard drives, chassis, power supply, and other components, the total system price has climbed to between 600,000 and 700,000 yuan.

The consumer market has not been spared. The average price for an RTX 5090 graphics card has increased by over 5,000 yuan, with the specific RTX 5090D model seeing its price jump from 16,899 yuan to 20,399 yuan.

One Tianjin resident who purchased a new computer in September 2025 found that an identical configuration cost over 10,000 yuan more just four months later. Within that increase, the price for two 96GB memory modules rose from just over 2,000 yuan to nearly 8,000 yuan.

Supply chain pressures have also given rise to a grey and black market. "Rebranded" memory modules and outright counterfeit products have flooded the market. These products fail to meet original manufacturer quality standards and cause significant performance issues when installed in servers.

04 Domestic Opportunities: Formation and Challenges of a Multi-Tiered Computing Power Market

The shortage of high-end GPUs is forcing computing power demand to fragment within the market, creating a parallel landscape that is multi-tiered and multi-sourced.

For applications requiring ultimate performance or those with dependencies on international software ecosystems, companies still strive to acquire NVIDIA chips. However, for broader industry applications and in government/enterprise markets with high demands for data security and controllability, domestic computing power has become the mainstream choice.

Domestic chip manufacturers are accelerating their efforts to catch up. At the end of 2025, companies like Moore Threads and MetaX (Muxi) went public on Shanghai's STAR Market, experiencing remarkable first-day stock price surges. These companies also face significant challenges, including high R&D investment and ongoing financial losses.

The trend of software-defined infrastructure is reshaping hardware standards from the top down. The rapid development of domestic large models, like DeepSeek, has created specific requirements for architectures like Mixture of Experts (MoE) and mixed-precision training. This, in turn, pressures domestic hardware manufacturers to overcome the hurdle of software and framework adaptation.

05 Future Trends: The Dual Path of Open Architecture and Demand Diversification

Solving the computing power shortage requires action on both the supply and demand sides. On the supply side, companies like Hygon are promoting an "AI Computing Open Architecture," aiming to bring together players across the AI industry chain to establish unified development standards.

On the demand side, computing power needs are diversifying. Cloud providers like UCloud are developing capabilities for heterogeneous resource management and virtualization, aiming to partition different types of computing power more finely and schedule appropriate workloads to suitable resources.

2026 is widely considered the first year of commercialization for AI applications. As model capabilities mature and inference costs decline, the market focus is shifting from "who has the hardware" to "who has the use cases" and "who has the user traffic."

Although discussions about potential computing power过剩 exist in the market, practitioners on the front lines believe current AI models still have some way to go before becoming genuine, reliable production tools, necessitating continued investment.

If AI agents mature to the point of solving real-world productivity problems effectively, the demand for computing power will continue its sustained growth.

Currently, the price for a complete H200 server system has rebounded to over 2.3 million yuan, while the proportion of data centers equipped with liquid cooling technology is approaching 80%.

The market is undergoing a transition from "scrambling for hardware" to "competing for application scenarios." Domestic computing power chips are already demonstrating cost-performance advantages in specialized fields such as inference tasks, industrial quality inspection, and smart government applications.

With the continuous traffic impact of the Spring Festival, the brief outage of Tencent Yuanbao might be just the beginning. Storage products, which require large capital outlays and have seen rapid price appreciation, have become new targets for speculation. Some graphics card dealers have already switched to trading memory modules.

The table below summarizes the current supply status and price trends for NVIDIA's high-end GPUs:

In this computing power supply-demand contest intensified by the Spring Festival, market speculation intertwines with genuine need, import restrictions coexist with domestic substitution efforts, collectively steering China's AI computing power market toward a more diversified and self-reliant development path.