From Alibaba to Baidu, major tech giants are competing for the dividends of AI cloud "new infrastructure"

Facing the vast blue ocean of intelligent transformation across all industries, cloud providers have once again lined up at the starting line, kicking off a race to compete in implementation capabilities.

In 1698, Thomas Savery invented a steam pump composed of a boiler, piston, and valve. It created a vacuum through steam condensation and then used atmospheric pressure to draw water from mines.

In 1712, blacksmith Thomas Newcomen improved the steam pump and created the atmospheric steam engine, which could operate continuously for 24 hours, keeping mines as deep as 150 meters free of water.

In 1765, James Watt invented the separate condenser, increasing the efficiency of the steam engine by six times. Over the next 20 years, Watt successively invented the flywheel and gear system, enabling the steam engine not only to pump water up and down but also to drive machines in rotation.

In 1785, the first Watt steam engine was put into operation in a cotton mill, doubling the spinning efficiency and thus opening a new chapter in the "Steam Age" for human society.

Looking back at the evolution of cloud computing, it is highly similar to that of the steam engine.

Early cloud computing was known for virtualization and elasticity. Just as the steam pump replaced the windmill water - pump, cloud computing solved the most urgent problems in enterprise digitization: no need to build computer rooms, buy servers, or maintain infrastructure. Enterprises only needed to purchase cloud resources on - demand.

The wave of large models has further redefined the value of the cloud. Just as the atmospheric steam engine improved water - pumping capacity, cloud computing in the era of large models has gradually taken on services such as model training, inference calls, and AI application development, evolving into an intelligent foundation across industries.

The emergence of Agents has enabled AI to move beyond dialog boxes and begin to have the ability to break down tasks, call tools, connect systems, coordinate processes, and execute continuously. It is like adding "flywheels and gears" to the cloud, breaking free from the constraints of "selling servers" and rising to become the intelligent engine for all industries.

In the 19th century, the steam engine was quickly applied in industries such as metallurgy, flour, coin - minting, and textiles, becoming a "universal machine" suitable for various manufacturing industries. As intelligence becomes a social need, cloud computing, which undertakes the mission of intelligent transformation across all industries, is starting a new battle.

In the foreseeable future, AI cloud will be the "new infrastructure" in the intelligent era - not only the greatest dividend of the era but also a newly emerging blue - ocean market.

The demand has changed: from "renting resources" to "demanding results"

As we reach 2026, the evolution direction of the cloud market has long been rewritten by Agents.

In the past decade, there has been a well - established formula to measure the competitiveness of a cloud provider: look at resource scale, revenue growth, number of customers, and market share. The one with a larger scale, more resources, and wider infrastructure coverage is considered the leader in the market.

This logic held true during the mobile Internet and industrial digitization stages.

At that time, the core demand of enterprises was digital transformation: business systems needed to be migrated to the cloud, data needed to be centralized, organizations needed to collaborate, and processes needed to be online. Cloud providers provided the foundation, resources, and basic capabilities. When customers chose to move to the cloud, they were essentially buying elasticity, stability, and cost optimization.

After entering the Agent era, the demand has undergone a fundamental change.

McKinsey mentioned in its judgment on Agentic AI infrastructure that the IT infrastructure is entering a new stage. AI Agents are starting to orchestrate, govern, and scale work within enterprises. The infrastructure is no longer just a support layer but has further become the core framework for enterprises to capture AI value.

In a nutshell: Customers no longer just "rent resources" but "demand results".

Banks don't just need computing power resources. They need to know whether thousands of AI applications can run stably and support continuous iteration in different scenarios such as risk control, customer service, investment research, operation and maintenance, and compliance.

Automobile manufacturers don't need an isolated AI model. They need to know whether assisted driving can form a complete closed - loop from training, simulation, verification to mass production and on - road operation.

Energy companies don't need a demonstration application. They need to know whether AI can be integrated into business operations such as power grid dispatching, equipment inspection, fault prediction, and customer service, and truly impact production efficiency, safety, and stability.

Manufacturing enterprises don't need "an intelligent Q&A system". They need to know whether AI can be integrated into R & D, supply chain, quality inspection, equipment operation and maintenance, and production line management to help enterprises solve specific business problems...

As Forrester mentioned in its report on Google Cloud Next 2026: Enterprise AI is moving from the "pilot era" to the "scaled - management era". Last year, enterprises asked "Can we build an Agent?" Now the question has become "How to manage thousands of Agents?"

A single Agent pilot tests the model's capabilities and demonstration effects. The stable operation of thousands of Agents tests the cloud provider's system engineering capabilities, which require computing power scheduling, model services, permission systems, data governance, security auditing, cost control, and so on.

There is a similar consensus in China.

In May, the Cyberspace Administration of China, the National Development and Reform Commission, and the Ministry of Industry and Information Technology jointly issued the "Implementation Opinions on the Standardized Application and Innovative Development of Intelligent Agents". AI is now officially regarded as industrial infrastructure. The cloud needs to evolve from a platform that "hosts applications" to an engine that "hosts intelligent decision - making and intelligent execution".

This corresponds to a change in the market competition logic: in the past cloud wars, the competition was about who occupied more territory; in the new cloud wars, the competition is about who has deeper roots.

The so - called "occupying more territory" refers to competing in resource scale, number of computer rooms, customer coverage, and market share, answering the question of "how big"; the so - called "deeper roots" refers to competing in industry understanding, scenario accumulation, engineering capabilities, delivery capabilities, and continuous operation capabilities, answering the question of "whether it has truly penetrated into the customer's business".

The key points of the race have changed: accelerating the evolution towards the "intelligent factory"

Reflected in cloud providers' press conferences, the narrative style has become more practical.

In the past, they talked about the number of global nodes and the number of customers. Now they talk about model call volume, Token consumption, and MaaS revenue, trying to prove to the market that "I not only have a good large model but also turn model calls into a steadily growing business".

Surprisingly, Baidu Smart Cloud stands out.

According to Baidu's Q1 financial report, Baidu's AI cloud revenue reached 8.8 billion yuan, a year - on - year increase of 79%, of which GPU cloud revenue soared by 184% year - on - year.

In the past month, many people have offered explanations: some think Baidu has caught the Agent wave, and the market demand for AI cloud is accelerating; some believe that Baidu Smart Cloud focused on "intelligence" 10 years ago, and now it's finally the pay - off period; others admit that Baidu's AI cloud has performed well, but the overall cloud business scale is still not large enough...

Perhaps what is more worthy of in - depth exploration than the results is - what exactly did Baidu Smart Cloud do right? The answer lies in the strategic upgrade in May 2026. Baidu will build a new full - stack AI cloud around "chip, cloud, model, and agent", which can be specifically summarized into two parts:

One is AI Infra. The original "MaaS model service" has been upgraded to the "Token Factory", and the product architecture has been reconstructed with the Agent - first concept. The goal is to minimize token redundant calculations and provide faster generation speed and more cost - effective token services.

The other is Agent Infra. Through optimization schemes such as hierarchical pooling, increasing KV Cache hit rate, PD separation, and cache scheduling, as well as the adaptation of mainstream models by the super - node product, the intelligent level of each Token is maximized, enabling agents to better complete tasks.

For example, the positioning of AI cloud is changing from a "training site" to an "intelligent factory". The former solves the problem of creating models from scratch, focusing on parameters, computing power, and training efficiency; the latter solves the problem of transforming AI capabilities into production capacity, focusing on inference cost, task orchestration, data closed - loop, and industry adaptation.

It's not just Baidu. The entire cloud market is evolving towards the "intelligent factory".

Microsoft launched Project Solara at Build 2026, emphasizing the construction of a platform from chips to the cloud for "agent - first" enterprise devices. Devices are no longer organized around traditional apps but around Agents. The cloud hosts Agent services, state management, and task scheduling.

Nvidia has been constantly emphasizing the concept of the "AI Factory" and has reached multiple AI infrastructure cooperation agreements with enterprises such as SK Group, Naver, LG, and Hyundai. Among them, SK Telecom will build a GW - level AI cloud, and Naver plans to build a GW - level AI factory to meet the needs of AI services and Physical AI.

Alibaba announced on June 8 that it would merge the Tongyi large - model business unit with the Future Life Laboratory and established a new Token Foundry business unit. By bringing together the complete puzzle from chips, models to applications in one department, it aims to control the "full life cycle of Tokens".

In other words, the scale - based narrative in the cloud market is a thing of the past. The key is not just "who has more computing power" but whether AI capabilities can be matched with industry needs, such as translating "general technology" into "industry capabilities", turning "model capabilities" into "business results", and enabling customers to use the service continuously rather than just trying it once.

If cloud providers' thinking remains at the resource level, they are doomed to be unable to answer the above questions.

At the mid - point of the race, implementation ability becomes the "decisive factor"

The "reshuffle" of cloud providers indicates that the growth point has shifted from "moving to the cloud" to "AI implementation".

The biggest opportunity in the mobile Internet era was the digitization of all industries. Enterprises moved their businesses online, transformed processes into data, and turned connections into entry - points, giving rise to trillion - level markets. In the AI era, the keyword has changed from "digitization" to "intelligence". Every industry hopes that the cloud and AI can enter the business scene together to re - activate key processes such as R & D, production, marketing, risk control, customer service, and operation and maintenance.

According to the general view in the industry, for AI cloud to be implemented, it needs to cross at least four thresholds.

The first threshold is to understand the industry.

Implementing AI is not just about handing a general model to customers. The financial industry has its own risk - control logic, the energy industry has its safety boundaries, and the automotive industry has its engineering verification cycle... Without industry know - how, AI is likely to remain at the surface - level application of "seeming intelligent".

The second threshold is to have full - stack capabilities.

In the Agent era, cloud services cannot just provide a single model or a certain amount of computing power. The operation of Agents requires the coordinated work of computing power, cloud platforms, large models, toolchains, intelligent agent frameworks, data governance, and security systems. Missing any link will affect the final implementation effect.

The third threshold is to have real - world scenario verification.

The most important thing for AI implementation is not the demos at press conferences but whether customers are willing to use it in the long - term and whether it can continuously generate value in real - world business. Especially in serious industries such as finance, energy, automotive, and government, customers will not easily migrate their core systems just because of new concepts.

The fourth threshold is cost sustainability.

One of the biggest variables in the process of AI applications moving from pilot to scale is cost. It's not difficult to make a single Agent work, but it's challenging to ensure that when thousands of Agents run for a long time, the inference cost, data call cost, operation and maintenance cost, etc. can be accepted by enterprises. Many projects end up stuck at the ROI stage.

Let's make a bold prediction: The ranking of AI cloud will not simply replicate that of the traditional cloud market. Scale is just an entry ticket, and the depth of implementation is the new dividing line. The business model of AI cloud will undergo a more in - depth transformation, moving from resource - based billing to capability - based billing, application - based billing, and even result - based billing.

The entire cloud market has entered a new "mid - race moment", and the players on the field seem to be ready to "charge forward".

Robin Li proposed the DAA concept at the Create Conference, believing that in the future intelligent agent era, the measure should not only focus on input but also on output, that is, how much is actually getting the job done and delivering results. Correspondingly, Baidu Smart Cloud is continuously strengthening its penetration into industries.

In the automotive industry, Baidu Smart Cloud has covered the entire chain from chip manufacturing to vehicle manufacturers and parts suppliers; in the financial industry, it has reached in - depth cooperation with more than 800 institutions, covering 80% of central - state - owned enterprise customers; in the field of embodied intelligence, it ranks first in the AI cloud market. A recent report from IDC shows that it holds the first market share in the decision - making tools and services market in the field of China's retail credit intelligent risk control, and is considered an important promoter of the intelligentization of the entire retail credit risk control.

Wu Yongming, the CEO of Alibaba, defined the Agent era as "a revolution in the computing paradigm" during the earnings conference call. Alibaba Cloud has started to make changes to chips, cloud platforms, models, and MaaS inference platforms simultaneously, trying to use a complete technology stack to meet the challenges of Agent scenarios.

Zhou Yuefeng, the CEO of Huawei Cloud, said bluntly that he "doesn't care about the total amount of Tokens" and wants to penetrate into the "black soil" of industries related to national economy and people's livelihood. He no longer regards the cloud as a simple storage and computing resource pool but as an industrial assembly line capable of producing Tokens on a large scale and with high efficiency...

Facing the vast blue ocean of intelligent transformation across all industries, cloud providers have once again lined up at the starting line, kicking off a race to compete in implementation capabilities.

Conclusion

From the perspective of enterprises, when the AI wave hits, it's not that they don't have a budget for AI or the willingness to move to the cloud. The real pain point is not knowing how to implement AI.

If cloud providers want to seize the dividends of intelligent transformation across all industries, they must make a fundamental change from "competing in scale" to "competing in depth": scale represents resource capabilities; depth represents implementation capabilities. "Scale" is the threshold to stay in the game, while "depth" is the "amulet" to stay in the customer's business in the long - term.

This article is from the WeChat official account “Alter Talks about Technology” (ID: spnews), written by Zhang Hefei and published by 36Kr with authorization.