A surge in computing power pricing is impacting the global market, with major industry players initiating a new round of price adjustments. Domestically, Alibaba Cloud issued three separate product price increase announcements within just four days. On the evening of April 15, Alibaba Cloud released a notice regarding price adjustments for certain model unit services on its large model service platform, Bailian. The company stated that to ensure the stable supply of underlying hardware, enhance platform operation and maintenance service quality, and respond to changes in computing power market costs, it will moderately adjust the service prices for some MU (Model Unit) model units. The price increases range from 2% to 5% and will take effect on May 15.
On the same day, Alibaba Cloud also announced price adjustments for its DDoS protection products. To standardize product pricing, it decided to adjust the pricing for the elastic 95 feature of DDoS Native Protection 2.0 (subscription), DDoS Advanced Protection (Mainland China), and DDoS Advanced Protection (non-Mainland China) products, effective July 15, 2026. Specifically, the price for DDoS Advanced Protection (Mainland China) elastic 95 will increase from 100 RMB/Mbps/month to 150 RMB/Mbps/month. Earlier, on April 13, the company announced adjustments to the API free tiers for DataWorks Standard and Professional edition users and introduced pay-as-you-go billing. For these users, the daily API call limit was removed. The free API call quotas were adjusted to 100,000 calls/month and 500,000 calls/month for Standard and Professional editions, respectively, with usage beyond these limits billed via OpenAPI pay-as-you-go.
Simultaneously, overseas AI giant Anthropic has also adjusted its pricing. However, compared to Alibaba Cloud's clear percentage increases, Anthropic's approach is more subtle. According to The Information, Anthropic has modified the subscription model for its enterprise product, Claude Enterprise. It has shifted from a fixed fee of up to $200 per user per month to a model based on actual computing power consumption. Additionally, users are now required to pay a fixed fee of $20 per month. Fredrik Filipsson, co-founder of Redress Compliance, a firm that assists with software licensing negotiations, indicated that for heavy users, the new pricing could lead to costs doubling or even tripling. This change was not previously publicly disclosed but has been rolling out over recent weeks, with some customers only feeling the impact upon contract renewal. An Anthropic spokesperson confirmed the new pricing model.
Looking further back, other leading companies like Tencent and Baidu have already raised prices for their large models. Concurrently, rental prices for AI chips in the B2B computing power leasing market are also climbing. Data from SemiAnalysis shows that the one-year lease contract price for an H100 GPU has increased from a low of approximately $1.70 per hour/GPU in October 2025 to $2.35 per hour/GPU in March 2026, a rise of nearly 40%.
Since the beginning of 2026, a global shortage of computing power has emerged, driven by rapidly growing demand with a steep growth curve. In April, the weekly cumulative token consumption on OpenRouter, the world's largest API aggregation platform, was about 7-8 times higher than a year ago, with domestic large models being the primary driver of this surge. CITIC Securities pointed out that the explosion in token usage has led to a massive spike in computing power demand. Meanwhile, supply-side growth is constrained by various hardware limitations, resulting in limited short-term marginal increases. This has led to a severe computing power shortage both domestically and internationally. The firm expects that the domestic computing power crunch, corresponding to the AI application boom, will accelerate the volume shipment of domestic chips, with a more direct impact on inference chips. While the performance of individual domestic chips is catching up rapidly, it still lags behind Nvidia by about 1-2 generations. They are optimistic that against the backdrop of the domestic computing power shortage, domestic AI chips, especially inference chips, will see explosive growth opportunities. It is estimated that the localization rate of the domestic AI chip market is currently around 30-40% and is expected to reach 60-70% by 2030.
Soochow Securities believes that the computing power leasing industry will undergo restructuring in the "Token Era." In Q1 2026, leading cloud providers and large model companies significantly increased their purchases of high-end computing power from leasing companies. Only top-tier cloud providers have been able to secure relatively sufficient high-end computing power, while the demand from second-tier cloud providers and large model companies for such resources remains far from satisfied, indicating a massive gap. Computing power leasing companies capable of providing scarce high-end computing power services are becoming core strategic resources fiercely competed for by cloud providers, large model companies, and application enterprises in this new Token-centric era.
Analysts suggest that the business model for computing power leasing is likely to shift from "selling computing power" to "selling tokens." Given the current high degree of computing power scarcity, leasing companies are expected to gain stronger pricing power. Their business models are evolving from simply leasing raw computing power to offering model services or token-sharing models—essentially transitioning from selling computing power to selling tokens. This shift has the potential to significantly enhance the profitability of computing power leasing companies and could drive a change in their valuation metrics from price-to-earnings (P/E) ratios towards price-to-sales (P/S) ratios.
Comments