Cloud Providers Collectively Shift Strategies

Image Source | Internet (Please contact for deletion if infringement occurs) Partially AI-generated

As we look back on the first half of 2026, an interesting development has taken place in China's domestic cloud computing industry.

At the beginning of the year, discussions were still centered around when the price war for large models would end.

By mid-year, the wind had suddenly shifted.

Alibaba Cloud replaced its growth engine from ECS to MaaS, Tencent Cloud publicly stated that 'competing on price is meaningless,' ByteDance's Doubao exceeded 180 trillion daily Token calls, and Huawei Cloud redefined itself as an 'Agentic Cloud.'

Several leading cloud providers almost simultaneously pivoted their strategies around Agents.

How Much Cloud Does an Agent Consume?

Data from the National Data Bureau shows that China's average daily Token calls were around 100 billion in early 2024. By March 2026, this figure had surged to 140 trillion.

A thousandfold increase in just two years.

Behind this exponential growth lies the true driver: Agents.

Traditional Chatbots engage in single-round dialogues, responding to user queries one at a time, consuming roughly 1,000 Tokens per call before ending the model invocation.

Agents, however, operate differently. Upon receiving a task, they autonomously break it down, invoke models multiple times, call external tools, and perform self-validation—all without human supervision.

The Token consumption per task for an Agent is 5 to 30 times that of a standard chatbot.

More critically, Agents can run 24/7 in the background, with multiple Agents interacting at high frequencies.

This means Token consumption growth no longer depends on 'how many people are using it' but rather on 'how many tasks are running.'

According to IDC forecasts, the global number of active Agents will grow from 79.4 million in 2026 to 2.216 billion in 2030, with a compound annual growth rate (CAGR) of 129.8%.

In the Chinese market, this growth is even more pronounced, with a CAGR of 151.2% for active intelligent agents and a staggering 3499.3% CAGR for Token consumption during the same period. Data from OpenRouter corroborates this trend.

After February 2026, the gap between Agent-related Token usage and human dialogue Token usage widened rapidly. By August, Agent usage surged from approximately 0.51 trillion to 7.3 trillion—a 14-fold increase—while human usage only grew 2.8-fold.

McKinsey's 2026 survey also revealed that 40% of large organizations have already deployed AI Agents at scale, up from just 27% a year earlier.

On cloud infrastructure consumption bills, a new major customer has emerged: 'machines.'

The surge in cloud usage is not a future scenario—it's happening right now.

What Are Leading Cloud Providers Doing?

Alibaba Cloud: Rebuilding the Entire Tech Stack

The Alibaba Cloud Summit in May 2026 was described by many attendees as 'one of the most disruptive launches in 17 years.'

Alibaba Cloud announced the completion of a full-stack Agent-based upgrade across 'chip-cloud-model-inference,' simultaneously launching a new AI product portal 'Qianwen Cloud,' a super-node server powered by its self-developed AI chip Zhenwu M890, and its flagship model Qwen3.7-Max.

Several key points are worth noting.

At the chip level, the Zhenwu M890 delivers three times the performance of its predecessor, the Zhenwu 810E. Paired with the self-developed interconnect chip ICN Switch 1.0, 128 AI chips can form a single computer with P2P latency below 150 nanoseconds, specifically addressing bottlenecks in massive concurrent inference scenarios for Agents.

More fundamentally, Alibaba Cloud transformed its cloud products into directly callable components for Agents. It revamped its entire cloud product lineup with Skill-based, MCP-based, and CLI-based modifications, enabling every cloud product to become a standardized module that Agents can invoke 'like calling a function.'

The logic is clear: if the primary users of future cloud platforms are Agents rather than human engineers, then the interface design philosophy of the cloud must evolve accordingly.

In May 2026, Alibaba Cloud's revenue from AI-related products exceeded 30% for the first time, reaching RMB 8.971 billion, marking the eleventh consecutive quarter of triple-digit year-over-year growth.

Alibaba management expects AI revenue to account for over 50% of external commercialization revenue within about a year. Agent-driven MaaS revenue is set to replace ECS as Alibaba Cloud's largest product line.

The growth engine has fundamentally shifted from selling computing power to selling Tokens.

Tencent Cloud: Focusing on Scenarios, Not Token Volume

Unlike Alibaba Cloud's full-stack reconstruction approach, Tencent Cloud took a more 'pragmatic' path.

In an interview earlier this year, a Tencent Cloud vice president stated bluntly: 'The first half of AI was more about selling resources.' He described the typical state of the past two years: enterprises chasing GPU computing power with simplistic sales approaches, but many clients lacked clarity on how AI could truly reduce costs and improve efficiency.

Tencent Cloud does not believe that merely selling API calls is a sustainable business model, arguing that 'while the volume looks large, the stickiness is extremely low.' Their strategy, therefore, focuses on application scenarios.

This judgment is reflected in their business structure. Currently, about a quarter of Tencent Cloud's revenue comes from SaaS, with an overall '4-4-2' breakdown (40% IaaS, 40% PaaS, 20% SaaS), markedly different from many cloud providers still focused primarily on infrastructure.

At the product level, Tencent Cloud has been aggressively launching offerings tailored for Agent scenarios. In July, it released Agent Bucket, the industry's first native storage service for AI intelligent agent scenarios, elevating multi-Agent isolation from code logic to system capability through the introduction of 'Space' independent spatial units.

Its Intelligent Agent Development Platform (ADP) covers the full lifecycle from building, connecting, and distributing to governing Agents, supporting multiple development frameworks and mainstream models.

Its Efficiency Intelligent Agent Toolkit now spans over 20 vertical scenarios, with products like Qidian Marketing Cloud's MAGIC Agent and Operations Intelligent Agent CloudQ gradually being deployed.

Tencent Cloud's bet, in essence, is that in the Agent era, the true value lies not in Tokens but in the deep integration between Agents and enterprise business systems.

Whoever enables Agents to 'get real work done' will secure a seat at the next round of the game.

Huawei Cloud: Transforming 'Industry Know-How' into Intelligent Agent Assets

Huawei Cloud's strategic approach differs significantly from its internet-based counterparts, reflecting its strong 'government and enterprise DNA.'

In March 2026, Huawei Cloud explicitly adopted an industry-focused intelligent agent strategy, partnering with industry clients and partners to build 'Industry AI Dream Factories.' It plans to create specialized zones for smart healthcare, embodied intelligence, smart manufacturing, and scientific computing, opening up industry models, datasets, and tools.

At the infrastructure level, Huawei Cloud introduced the 'Agentic Infra' paradigm, centered around 'efficient Token factories + continuous learning + integrated general and intelligent scheduling + autonomous security,' and launched the Lingqu Ascend 950 intelligent computing cluster service, boosting Token throughput performance by 20% compared to the previous generation, aiming for 40 days of stable operation for training tasks on the cloud and fault recovery within 10 minutes.

Addressing the memory capacity and access efficiency demands of Agent long-term tasks, Huawei Cloud unveiled the CMS Memory Storage Solution, constructing a petabyte-scale memory space with double the storage capacity of similar industry products and supporting terabyte-scale memory high-speed reads.

Its enterprise-grade Intelligent Agent Platform 'Zhiguo AgentArts' has served over 100 enterprises, including clients like Shenzhen Longgang Government, China Southern Power Grid, and Kingsoft Office. The 'Industry AI Dream Factory' has accumulated over a thousand industry assets, with over 1,000 projects deployed.

Huawei Cloud's logic is clear: for Agents to truly operate within enterprises, generic models alone are insufficient. Industry knowledge, business processes, and data assets must be transformed into capabilities that Agents can invoke.

This path is resource-intensive, but once established, the barriers are high.

Baidu Intelligent Cloud and Volcano Engine: Two Different Paths to 'Speed'

Baidu Intelligent Cloud and Volcano Engine have adopted distinct strategies but are both moving rapidly.

Baidu Intelligent Cloud has built the Qianfan Agent Infra, a four-layer architecture covering 'Token Factory-Tool Services-Orchestration Engineering-Intelligent Agents.'

At the application layer, the general-purpose intelligent agent 'Baidu Dazi' collaborates with vertical intelligent agents like Famou, MiaoDa, Yijian, and Shengsuan, using Dazi as a unified portal to orchestrate models and tools.

The release of the Intelligent Agent Sandbox AX provides Agents with a secure, isolated execution environment supporting browser, code, and custom sandbox forms.

Baidu Intelligent Cloud has also performed strongly in the AI cloud bidding market. In the first half of 2026, it ranked first among cloud providers with RMB 1.385 billion in bid wins, accounting for nearly 60% of the total bid amount across five cloud providers.

This indicates that in industry markets requiring deep customization, such as government, finance, and energy, Baidu Intelligent Cloud's Agent solutions have secured substantial orders.

Volcano Engine, meanwhile, has taken a different route. At its FORCE conference in June, Doubao's large model exceeded 180 trillion daily Token calls, growing over 10-fold in the past year. IDC data shows Volcano Engine leading China's public cloud MaaS service market with a 49.5% share.

Volcano Engine's MaaS strategy is clear: maximize model capabilities to enable enterprises to access Agent capabilities at the lowest cost.

Doubao 2.1 Pro reduces comprehensive usage costs by nearly 80% compared to Claude Opus 4.6 and offers an Evolving version with rapid iterations two to four times per month.

At the development tool level, the Fangzhou CLI command-line tool allows developers to integrate Agents with a single command, while HiAgent and Kouzi rank first in China's intelligent agent development platform market with 17.8% and 19.3% shares, respectively.

Tan Dai, President of Volcano Engine, predicts: 'The new workload built on models and Agents could be 10 to 20 times larger than traditional clouds.' From this perspective, Volcano Engine's strategy of capturing Token market share through cost-effective MaaS services lays the groundwork for future growth.

The Turning Point of the Token Economy

While the five cloud providers have taken different paths, their simultaneous strategic adjustments are driven by several common underlying logics.

The 'Consumers' of the Cloud Have Changed

Traditionally, cloud customers were enterprises and developers purchasing computing, storage, and networking resources. Now, Agents have become major cloud consumers with far larger 'appetites.' Cloud interface designs, billing models, and security frameworks must all be reengineered around Agents.

Alibaba Cloud's assertion that 'the primary consumers of cloud and AI are shifting from humans to Agents' directly encapsulates this trend.

The Fundamental Logic of Business Models Has Changed

Traditional cloud business models sold resources on a usage-based basis, offering low commercial certainty. However, Token consumption in the Agent era is predictable, with task volume determining consumption and consumption driving revenue.

AI pricing is transitioning from usage-based billing to long-term subscription agreements, while cloud providers' capital expenditures are becoming smoother and more predictable. This means cloud providers can finally operate like utility companies, running a 'predictable' business.

The Sources of Competitive Barriers Have Changed

As MaaS becomes infrastructure, model invocation alone offers little differentiation. The true competition lies in embedding Agent capabilities into enterprises' core business workflows.

Tencent Cloud's reference to 'extremely low stickiness' highlights this point: selling Tokens alone creates no moat. However, if Agents operate within a company's marketing systems, government workflows, or production line quality inspection processes, the cost of replacement becomes prohibitively high.

This is why all providers are elevating Agent businesses to higher organizational levels, separating them from legacy product structures.

The adjustment logic can be summarized as: scale mature businesses while agilizing emerging ones. No one wants to be reduced to a mere conduit in the next layer of value distribution.

The year 2026 has not yet ended, but the general direction of this round of 'reshuffling' in the cloud computing industry is already clear enough.

Agents are becoming new consumers of the cloud, which is the most certain fact.

Around this fact, Alibaba Cloud has chosen full-stack restructuring, Tencent Cloud has opted for in-depth scenario cultivation, Huawei Cloud has selected industry penetration, Baidu Intelligent Cloud has chosen Agent Infra, and Volcano Engine has prioritized ultimate cost-effectiveness.

Despite their different paths, they are all answering the same question: When the primary users of the cloud shift from humans to Agents, what should cloud providers look like?

There is no standard answer to this question. However, it is certain that approaches that merely treat Agents as a functional module and attach them to existing product lines are unlikely to fare well in the upcoming competition.

This is because Agents are not just a new feature, but a new 'operating system-level entry point.'

They redefine who uses the cloud, how the cloud is used, and what the cloud itself should provide.