More than 100 years ago, when electricity first entered factories, many factory owners made what seemed like a reasonable decision: keep the complex steam-engine shafts and belt systems, and simply replace the steam engines with electric motors. The result disappointed them. Power increased, but productivity barely rose with it.

Over the next 20 years, they gradually realized that the real transformation brought by electricity was not a simple swap of power sources. It came from embedding electricity into different business units, using different types of motors to drive drills, conveyor belts and other equipment. Only then did productivity surge.

That is the kind of capability Alibaba Cloud wants to give companies as they adopt AI more deeply. It believes customers in the AI era do not want only a single model or a cloud capability. They want a more flexible, integrated experience that lets them use stronger models at lower cost. Like electricity when it finally lifted industrial productivity, AI cloud should provide services across multiple layers and embed itself throughout enterprise workflows.

Business data supports that view: among customers calling large-model APIs on Alibaba Cloud, through its MaaS service, 70% are also using its GPU computing power services.

Liu Weiguang, senior vice president of Alibaba Cloud Intelligence Group and president of its public cloud business, said the first wave of customers using AI deeply will divide scenarios into different tiers. They will not simply call large-model APIs, but will also fine-tune or post-train foundation models with internal data, or train their own models from scratch.

In Liu’s view, the changes AI will bring to cloud computing are only beginning. The entire cloud computing architecture must be rebuilt for AI. “MaaS has huge growth potential, but the key is winning the full incremental AI cloud market,” he said. That means building a full-stack AI cloud capability that integrates hardware and software, helping companies call stronger AI models at lower cost and solve problems across different layers and scenarios. “That is what will decide the competition.”

In-Depth Research With 146 Industry Customers: Token Quality Matters More Than Quantity

“If every AI application today could only be used 100 times a day for free, what would you use it for?”

That is the question Liu Weiguang kept returning to after visiting 146 customers in 2025. His conclusion was blunt: no one would waste those uses on jokes or small talk. They would use them for the most important things, such as decisions most useful to their work or advice most critical to family life.

In his view, this is the fundamental difference between enterprise use of large models and consumer use. Consumers may spend tokens for entertainment, but for efficiency-driven companies, every token exchange carries a cost. They pay not only for the tokens, but also invest labor and time across business lines.

It is hard to imagine a young engineer facing equipment failure needing several rounds of dialogue before getting a solution. What is needed is a faster response and a guide that helps solve the problem quickly. Likewise, when a trader at a fund company uses AI to help capture trading signals, it is hard to tolerate a model that falls into prolonged “thinking,” produces a long answer, and then asks, “Would you like me to make this more polished?”

Companies in traditional industries are using different methods to avoid AI’s weaknesses and improve efficiency as much as possible. An automotive diagnostics company, for example, built a large model for remote auto repair assistance using 30 years of accumulated industrial inspection reports, and only then applied it to diagnostic reports. Fund companies combine models with more than 20 years of accumulated data and trading behavior, turning unstructured data in various forms, such as text, audio and images, into standardized information that can support investment decisions.

Enterprises also want to fully tap AI’s potential. Two Chinese agricultural and animal husbandry giants are both using Qwen to do similar work. They are not only trying to use VL, or vision-language large models, to identify the number of pigs, but also to detect abnormal behavior, monitor pigs’ health and vitality, and develop veterinary large models to address shortages of specialized talent. A leading lighting company, after integrating Qwen, is no longer limited to simple light switching or color-temperature control. It is using AI to understand ambiguous user commands and build a smarter on-device language model through smoother dialogue, making interaction between people and lighting more natural.

Once companies find the right way to embed AI into business processes, their use becomes serious and consistent. Almost every company in online recruitment is introducing AI resume screening, intelligent interviews and automatically generated interview records. Once recruiters get used to AI-assisted work, a new workflow forms. They use it every day, regardless of personal preference.

“Consumer use of AI will fluctuate, but the enterprise market will only keep growing,” Liu said. Both breadth and depth of use will continue to increase, and many scenarios have not yet been unlocked. “If AI can change auto damage assessment, that would absolutely be a ‘revolution.’”

At the 2024 Apsara Conference, Alibaba Group CEO Eddie Wu said in a speech that the greatest potential of generative AI is not creating one or two new super apps on phone screens, but taking over the digital world and changing the physical world.

The development of China’s enterprise AI market over the past year has validated that judgment. AI is no longer just an app inside a phone. It is appearing in more carriers, including glasses, earbuds, learning tablets, toys, fitness equipment, cars, robots and the full range of hardware devices.

These needs across different layers and scenarios cannot currently be met by a single model API service. In the U.S. market, the SaaS industry, which already provides tools to enterprises, has seen continued growth in large-model call volumes and offers a relatively standardized solution. In China, where the SaaS industry has not developed in the same way, traditional industries tend to rely on customized services to solve specific scenario problems, often requiring post-training or fine-tuning of large models. By providing such services, cloud computing companies are, to some extent, offering enterprises something similar to SaaS.

“Even if you add up all MaaS services today, they still account for a small share of China’s cloud computing market, and even of the AI cloud market. MaaS clearly has enormous room, but not today,” Liu said. Counting only large-model API calls in the public cloud market cannot represent the whole picture of AI cloud. Real token consumption must include MaaS platform API calls, tokens generated by public-cloud GPU inference clusters, tokens generated by private model deployments, tokens generated by device-side models and more.

“Token consumption below the surface is very large, but it cannot be counted. And enterprise AI adoption is still in the early stage of transformation. More than 90% of companies have not really begun to act, so future growth will certainly be hundreds of times larger.”

What can be observed, however, is that as long as foundation model performance keeps improving, and cloud providers go deep into every layer of the technology stack to provide services, improve inference capabilities and lower costs, they can bring more customers in more industries to use AI to solve problems.

Building Infrastructure for the AI Era to Serve Needs at Every Layer

NVIDIA CEO Jensen Huang once made a well-known argument: GPU clusters are “token factories,” taking in energy and producing tokens. That is a classic chip company perspective, simplifying the AI production process into energy conversion at the physical layer.

For cloud providers, simply reselling computing power makes it difficult to offer usable AI services today. They must use systems engineering capabilities to improve the efficiency of existing computing power as much as possible, and provide model services for enterprises in different industries and at different layers.

That is Alibaba Cloud’s choice: to become infrastructure for the AI era. In Liu Weiguang’s analogy, Alibaba Cloud is building a modern water plant, not merely transporting water, or large-model APIs. It also has to maintain water sources, meaning open-source models; build purification workshops, meaning data cleaning and model training platforms; lay water transmission networks, meaning high-performance networks; and treat wastewater, meaning security governance.

Within this system, Alibaba Cloud can serve different types of current “water use” needs:

MaaS, or direct water supply: just as households can turn on a tap and use water, enterprises and developers do not need to worry about the complex underlying network. They can call APIs directly, use them out of the box, and pay on demand. This is the lightest form of access.

PaaS, or industrial water supply: similar to factories that need specific water sources, enterprises can obtain foundation models and directly use open-source models on Alibaba Cloud’s platform for fine-tuning or their own post-training, then deploy them in the appropriate environment.

IaaS, or water-treatment infrastructure: this is like sending water that has been preliminarily purified and extracted to beverage or beer factories. Enterprises can use the computing power and basic software provided by Alibaba Cloud to train proprietary “beverages,” such as autonomous driving models and various vertical large models.

Alibaba Cloud has already made early progress. According to market research firm Omdia, China’s overall AI cloud market, including AI IaaS, PaaS and MaaS, reached 22.3 billion yuan in the first half of 2025. Alibaba Cloud held a 35.8% share, more than the combined share of the second- through fourth-ranked players.

Building this comprehensive infrastructure requires not only heavy investment, but also strategic resolve. In February 2025, Alibaba announced it would invest more than 380 billion yuan over the next three years to build cloud and AI hardware infrastructure, exceeding its total investment over the previous decade. In the first three quarters of 2025 alone, Alibaba’s capital expenditure on AI data centers and related infrastructure reached 95 billion yuan.

At the foundation model level, Alibaba has continued to invest in training models of different sizes, types and modalities, while committing resources to bring them into the top tier. For example, the Wanxiang 2.6 visual generation model has performance comparable to OpenAI’s Sora 2; Qwen-Image-Layered is the industry’s first model capable of precise layered image editing; and Qwen3-Max ranks near the top of global model performance rankings.

Alibaba has chosen to open source these models and make them available to teams and companies across industries. Qwen now has more than 180,000 derivative models, the largest number globally.

As infrastructure, Alibaba Cloud supports more than Alibaba’s self-developed models. Moonshot AI also trains its Kimi series models on Alibaba Cloud, and many intelligent driving teams use Alibaba Cloud to train models as well.

At the same time, Alibaba Cloud provides a full set of system capabilities to support the rapid growth of new products born in the AI era. Beyond Alibaba’s own Qwen app, these include Ant Group’s Lingguang and Afu, Moonshot AI’s Kimi app, and MiniMax’s Hailuo.

Although AI adoption across industries is still in an early stage, Alibaba Cloud, positioned as infrastructure, is also exploring new product forms to prepare for the next wave of AI applications. The most typical example is the agent version of the Qwen app that Alibaba is developing. It will not be limited to answering user questions. It will be able to call Taobao for price comparisons, use Amap for navigation, and potentially use all Alibaba services as plugins.

Ultimately, these capabilities, tested within Alibaba’s internal businesses and among leading companies in various industries, will be absorbed into Alibaba Cloud and become products offered externally. The aim is to give customers the ability to generate and use intelligence over the long term and in a sustainable way, rather than locking them into a single metering method.

AI Is Accelerating Customers’ Move to the Cloud

AI is giving cloud providers new growth momentum. AWS, Microsoft Azure, Google Cloud and Alibaba Cloud are all expanding rapidly.

But that momentum is not coming only from GPU usage or large-model API calls. The Alibaba Cloud team has observed that customers using these services on Alibaba Cloud are increasing their use of products such as compute, storage, networking and big data faster than the overall market.

“AI will accelerate customers’ move to the cloud,” Liu said. To use AI well, customers have to move their data fully to the cloud. For an agent to create value, the foundation model is only one part of the equation. High-quality business data matters just as much.

Microsoft Azure follows a similar growth logic. Selling API services for OpenAI’s foundation models is only one part of it. A stronger driver comes from enterprises that, in order to apply more powerful models in their businesses, migrate content and data scattered across local systems and different platforms into cloud products that are easier for models to call.

Traditional cloud computing architectures designed for high-concurrency Web and HTTP requests struggle to support such needs efficiently. A major reason Oracle regained growth momentum was its deployment of RDMA, or remote direct memory access, high-performance networking and autonomous databases, which fit the needs of large-model training and inference.

This has directly changed the outlook for public cloud computing services in the Chinese market. For years, Chinese cloud computing companies did not resemble overseas cloud platforms such as AWS, whose public cloud customers span industries and include the New York Stock Exchange, large oil companies and banking giants, allowing scale to translate into profits.

In China’s cloud computing market, the electricity, bandwidth and other infrastructure costs that platforms depend on are not controlled by the companies themselves. Some traditional enterprises, due to data security, compliance or historical inertia, still prefer to build their own data centers.

“Alibaba Cloud’s basic cloud architecture has been rebuilt for AI,” Liu said. AI infra is not a specific vertical; it is cloud computing itself. It requires not only scale, security and stability, but also the ability for services on the cloud to flow across one another, including new vector databases, efficient big-data cleaning platforms, flexible development frameworks and supporting software systems that meet enterprises’ AI needs across different layers and scenarios.

In Alibaba Cloud’s view, the competitiveness of cloud computing platforms in the AI era lies in an integrated hardware-and-software system capability. Hardware is not just chips, but the entire high-performance underlying architecture built around GPU computing power. Software is the ability to understand, optimize and schedule models.

“Alibaba Cloud’s goal is to capture 80% of the incremental Chinese AI cloud market in 2026,” Liu said. But even 10% of next year’s incremental market will be larger than the entire market of the previous year. What has been achieved in the past is therefore not important. The changes are only beginning.