In recent times, the AI landscape has witnessed a meteoric rise, with domestic internet behemoths aggressively deploying large-scale AI models. The recent surge in popularity of 'Lobster' (OpenClaw) has enabled these firms to reap significant profits from their AI ventures. However, as anticipated, the wave of AI price hikes has arrived as scheduled. With overseas behemoths like Amazon and Google, along with domestic giants such as Baidu, Alibaba, and Tencent (BAT), collectively raising prices, domestic large model enterprises like Zhipu AI have also followed suit. This leads to the question: Will the cost surge render 'Lobster staff' unaffordable?

I. Tech Titans Unite in Price Hikes

According to a report by Haibao News, concept stocks related to optical modules and computing power hardware have recently surged, bucking the market trend. This surge is fueled by the 'Lobster' (OpenClaw) craze that swept from abroad to China this year, leading to a sharp increase in Token usage and directly driving up demand for computing power. The Token supply chain has emerged as a new focal point.

Alibaba Cloud and Baidu Intelligent Cloud have successively announced price increases for AI computing power-related products. Alibaba Cloud's official website stated that due to the global explosion in AI demand and supply chain price hikes, prices for its AI computing power and storage products would rise by up to 34%. Insiders reveal that Alibaba Cloud's MaaS business platform, Bailian, achieved record-high growth from January to March this year.

Baidu Intelligent Cloud cited the rapid development of global AI applications, leading to a continuous rise in demand for computing power and significant increases in costs for core hardware and related infrastructure. Consequently, it has decided to raise prices for AI computing power-related product services by 5% to 30%, with parallel file storage increasing by about 30%.

Earlier this year, Amazon AWS and Google Cloud successively announced price increases for some services. On January 22, AWS announced a 15% price hike for EC2 instances used for large model training. On January 27, Google Cloud significantly adjusted prices for data transmission services such as CDN Interconnect, Direct Peering, and Carrier Peering, with increases of up to 100% in North America.

Domestically, Tencent Cloud announced a price increase for its large model services on March 11, adjusting the billing strategies for some models. Taking the Tencent HY2.0 Instruct model as an example, its input price surged from the original 0.0008 yuan per thousand Tokens to 0.004505 yuan per thousand Tokens, a staggering increase of 463.13%.

Moreover, the wave of price increases has not subsided. On April 8, Zhipu AI announced its third price increase plan for the year, raising prices by 10% while releasing its new flagship model, GLM-5.1, just one month after its previous price increase of at least 30%. Besides Zhipu AI, Tencent, Alibaba, and others have also successively halted free public testing and raised API call prices. Among them, Tencent's Hunyuan large model saw a maximum price increase of 463%. In contrast to the price reductions and free trials in 2025, the current clear signal is that domestic AI large models are entering an era of collective price increases.

II. The AI Cost Boom is Here: Will 'Lobster Staff' Become Unaffordable?

Recently, major large model giants have collectively announced price increases for some of their cloud services and related products, offering hope to those previously worried about being 'overshadowed.' Compared to expensive Tokens, we seem to be a more cost-effective option. So, what industrial logic underpins this wave of price hikes?

Firstly, the explosion in popularity of AI agents has triggered a surge in demand for computing power. From a structural shift on the demand side, the popularity of next-generation AI agents represented by OpenClaw has completely reconstructed the underlying logic of computing power consumption. In the early stages of large model applications, user behavior was mostly limited to single-round conversations or simple text generation, with relatively limited and predictable Token consumption. However, with the maturation of autonomous agent technology, AI is no longer a passive question-answering machine but a digital employee capable of autonomous planning, tool invocation, and executing complex tasks.

The per capita Token consumption of a mature agent in a single day's operation often exceeds that of traditional chat users by tens or even hundreds of times. This exponential increase in demand is not linear business growth but a dimensional leap. When a vast number of agents are simultaneously online, engaging in high-frequency logical reasoning and data interaction, the computing power infrastructure originally designed for human interaction instantly faces immense throughput pressure.

This explosive growth in demand has directly disrupted the old supply-demand balance, causing computing power resources to rapidly shift from 'relatively abundant' to 'extremely scarce.' When marginal utility sharply rises and supply elasticity is insufficient, price increases are not only an inevitable manifestation of market laws but also a necessary means to screen high-value application scenarios and curb inefficient computing power waste.

Secondly, the shortage of core hardware has led to tight computing power supplies. From the perspective of supply-side constraints, the shortage of high-performance computing chips and HBM high-bandwidth memory and other core hardware constitutes the physical foundation of this round of price increases. Despite significant investments by domestic cloud providers in building a domestic computing power ecosystem in recent years, global capacity bottlenecks remain severe in the field of high-end training and inference chips. In particular, HBM memory, the 'lifeblood' of large models, with its high technical barriers and long expansion cycles, has become a key bottleneck restricting computing power release.

Currently, computing power is no longer simply about stacking servers but a precision system composed of advanced process chips, high-speed interconnection networks, and high-bandwidth storage. The shortage of core hardware has significantly increased the marginal cost of computing power supply, preventing cloud providers from diluting costs through simple economies of scale. This rigid constraint on the supply side forces the industry to re-examine the pricing mechanism for computing power. When 'computing power is power' becomes a consensus, providers with stable, high-performance computing power supply capabilities naturally gain stronger bargaining power. The current price increases are, in essence, a reasonable revaluation of the value of scarce hardware resources and an inevitable result of cost pressures from the upstream of the industrial chain being passed downstream.

Recently, many friends of mine in tech companies, especially CTOs, have been lamenting that the current prices of storage chips and servers are truly prohibitive. The battle for computing power seems to have instantly become a battle for costs, which is the most noteworthy aspect at present.

Thirdly, the industry's pricing logic of 'trading volume for price' has undergone a fundamental transformation. Reviewing the development of the cloud computing industry over the past decade, we can easily identify a vicious cycle: cutthroat price wars. To compete for market share, major providers have resorted to 'price-cutting tactics,' not only squeezing the living space of competitors but also significantly compressing their own profit margins. At certain times, cloud service prices even fell below their operational costs, resulting in a typical phenomenon of 'involution.' This 'trading volume for price' model may have been effective in the mobile internet era, as the marginal cost approached zero, and the path to monetizing traffic was clear.

However, the advent of the AI era has completely shattered this logic. Computing power is no longer a cheap, general-purpose commodity but an expensive, specialized production factor. If low-price strategies are maintained, cloud providers will be unable to cover the high costs of GPU procurement and electricity operation, let alone continue to invest substantial R&D funds for model iteration. Healthy industrial development must be built on a reasonable profit foundation. Only when prices return to their value can enterprises engage in reproduction and innovation.

The collective price increases by tech giants are, in fact, a 'collective return to rationality' in the industry. This marks that the Chinese cloud computing market is bidding farewell to the 'money-burning subsidy' era and entering a value competition era centered on technological strength and service quality. This is extremely beneficial for the construction of a healthy industrial ecosystem, shifting the focus of competition from 'who is cheaper' to 'who is more stable, who is more intelligent, who can better solve problems,' which is undoubtedly a positive signal for industrial upgrading.

Fourthly, the rise of tokenomics will make tiered pricing the norm. Once, data centers were seen as 'warehouses' for storing data, with their value primarily reflected in space rental and data storage. In the AI era, data centers have evolved into 'factories' for producing intelligence, with their core output being high-value Tokens. This role transformation has directly given rise to a new pricing logic. Future AI services will no longer follow traditional annual or monthly subscription or pay-as-you-go models but will adopt tiered pricing based on dimensions such as Token throughput, response speed, and inference complexity.

This refined pricing strategy can more precisely match computing power demands in different scenarios, allowing tasks with high real-time requirements and complexity to pay higher premiums while enabling offline batch processing tasks to enjoy lower costs. This is not only a commercial model innovation but also a significant improvement in resource allocation efficiency. Through price leverage, the industry will guide computing power resources to areas that create the greatest social value, avoiding resource misallocation and waste.

Under such circumstances, many companies have even started using Token provision as a new employee benefit. Of course, we have been discussing whether Tokens are production factors or employee compensation and benefits, but there is no doubt that Tokens have become a significant computing power bottleneck restricting the development of AI companies in this era. Moreover, some friends of mine have complained to me that earlier this year, the tech giant they worked for required all employees to use 'Lobster' and their own digital twins. However, under recent massive consumption, usage restrictions have been imposed.

Ultimately, facing the increasingly expensive computing power resources, what should the future of AI 'Lobster' be? Can we still afford it?