Alibaba's Steam Engine Era

Author | Huahua

Many people may not know that the first batch of steam engines put into practical use during the Industrial Revolution had nothing to do with trains. 

In 1712, British engineer Thomas Newcomen built an atmospheric steam engine that worked around the clock to pump water at the mouth of a coal mine. The thermal efficiency of this machine was only about 1%. Tons of coal were burned in carts, but it only did the work that miners and horses had already been doing. 

It could operate mainly because coal mines had no shortage of coal, and cheap fuel temporarily covered up the efficiency problem. 

From the perspective of that time, it was an expensive mine equipment, far away from textile factories, steamships and railways. 

More than 300 years later, Wu Yongming, CEO of Alibaba, placed today's AI Coding next to the electric light of 1882 at the Yunqi Conference: machines can write code, generate reports, sort out meeting minutes, and still undertake the existing work of human beings, while the representative product of the machine intelligence era has not yet appeared. 

Next, he tried to calculate the account of this yet-to-appear productivity. Today, the total amount of machine thinking is less than 3% of that of human beings, and it may reach more than 1000 times of that of human beings in the future. 

Alibaba has not announced the specific statistical caliber, and this number cannot be directly converted into Tokens. It is more like a strategic assumption about the market size. 

I. Steam Engines Could Only Pump Water at the Very Beginning

When Newcomen's machine appeared, the main power in Britain still came from water and wind. 

Economic historian Nicholas Crafts estimated based on historical statistics that the installed steam power in Britain in 1760 was about 5000 horsepower, accounting for only 5.9% of the total power of wind, water and steam. 

In 1830, steam power was roughly equal to water power; by 1870, this figure had reached 2.06 million horsepower, accounting for nearly 90% of the total power of the three types. 

It took the steam engine more than a century to grow from 5.9% to nearly 90% share. 

Later generations are used to associating steam engines, trains and the Industrial Revolution together, as if the train drove out of the station the moment Watt tightened the last screw. 

But history advanced much more slowly. 

Early steam engines had low efficiency and narrow uses. They first pumped water in coal mines, then entered textile and metallurgy industries; factories were rebuilt around power, railways and ports expanded the scope of use, and a single equipment in the mine gradually became the basic capability of industrial society. 

Today's machine intelligence is also at the stage where the power has emerged but the organizational mode has not changed yet.

AI Coding can shorten a project that originally takes eight hours to complete to two hours, but the number of projects, the scale of personnel and working hours are still constrained by the original boundaries. 

Writing reports, making PPTs and sorting out meeting minutes follow the same logic: the machine improves the speed, but the total amount of tasks has not changed by orders of magnitude accordingly. 

In his speech, Wu Yongming described another kind of demand: millions of Agents conduct long-term research on rare diseases, or divide the construction of a Mars spacecraft into tens of millions of subtasks. Humans only need to give the goal once, and the machine will continuously call tools, check results and arrange subsequent steps. 

The computing demand thus extends from how many people use AI to how many machine tasks a single goal can unfold.

II. Wu Yongming Found a "Horsepower" for Machines

After Newcomen, Watt improved the steam engine. 

In 1769, Watt obtained the patent for the independent condenser, which moved the steam condensation process outside the cylinder. According to the data from the Science Museum in the UK, this improvement can reduce coal consumption by about two-thirds. 

In the 1780s, Watt solved the problem of rotary motion, so that the steam engine could drive looms and other factory equipment. 

When the machine became easy to use, Watt still had to make mine owners and factory owners understand how much it was worth. Customers were familiar with how much work a horse could do in a day, but it was difficult for them to judge the working capacity of a steam engine. 

Watt then converted the machine output into "horsepower", using the most common power in the old era to explain a strange commodity. 

"Horsepower" later became a physical unit, and it also served as a sales tool at the very beginning.

"Machine thinking volume" plays a similar role. 

It takes human thinking as a reference to mark the possible demand scale for an immature machine productivity. Two experiments disclosed at the Yunqi Conference left countable records for this work. 

Qwen3.8-Max built the training process, constructed data, designed experiments and located defects without human intervention, running continuously for more than one month and completing 33 rounds of effective iterations. 

Another chip design experiment lasted for more than 60 hours, the model called EDA tools more than 10,000 times, and finally reduced the physical area of a chip bus module by 42%. 

These results come from Alibaba, and more external verification is still needed. But "33 rounds" and "10,000 calls" have already recorded a change: after humans issue the goal, the machine can continue to complete a large number of steps in the background. 

Alibaba also announced that Qwen4 based on the new architecture has entered the training phase, and Qwen4.5 and Qwen5 are planned to be scaled to 5 trillion to 10 trillion parameters. Parameters cannot be directly converted into productivity, and this plan points to the complex task capability, continuous running time and room for improvement based on feedback. 

Tokens record the calls that have occurred, while "machine thinking volume" describes the scale that Alibaba expects these calls can still expand to. 

III. Power Spreads First

Once a long task lasts for several hours, days or even months, it is difficult to complete only by the model capability. 

The task may be interrupted or go in the wrong direction. The system needs to save progress, resume operation, check permissions, and keep computing, storage and network in coordination. 

T-Head announced the new generation of training and inference integrated AI chip Zhenwu V900 at the conference. The official said that the performance of a single chip is 3 times that of the previous generation M890, and it is planned to be mass-produced and commercially released in the first quarter of 2027. 

The conference also demonstrated server CPUs, interconnection chips, smart network cards and storage controllers. The upgraded super node architecture can connect these components into a single cluster with a maximum scale of 500,000 cards. 

The change behind this is very straightforward. 

Alibaba has begun to shift its attention from a faster AI chip to a complete computing system.

When the model runs continuously, the CPU processes tasks, the network transports data, and the storage is responsible for reading and writing. If any link cannot keep up, the expensive AI chip can only wait. 

In addition to hardware, Alibaba Cloud's released Agentic Cloud is responsible for the execution and management of long tasks. AgentCore manages failure retry, breakpoint recovery, identity permissions and operation audit, and Agent Sandbox provides an isolated environment. 

What enterprises actually purchase is a production environment that allows machines to run continuously, act according to permissions and leave complete records.

This system needs to be built in advance. 

In February 2025, Alibaba announced that it would invest at least 380 billion yuan in the construction of cloud computing and AI infrastructure in the next three years; at this Yunqi Conference, Wu Yongming set the data center target for 2032 at 20GW. 

Models, chips and clouds have been connected along the demand of long tasks. Today, these production capacities need to be digested by the machine tasks generated by enterprises every day. 

IV. The Electric Lights Are On, But Factories Have Not Changed Yet

On September 4, 1882, Edison's Pearl Street Station was put into operation in New York. It initially served 82 customers and only lit about 400 lamps. The electrical age later written into textbooks started from a small network in lower Manhattan. 

Electric lights soon made people see the use of electricity, but the changes in factories were much slower. 

Economic historian Paul David recorded a period of history. After early factories got electricity, the owners removed the central steam engine, replaced it with a large electric motor, and kept all the transmission shafts, belts and plant layout on the top of the workshop. 

The electric motor still dragged the original production system to run, and the productivity hardly changed. 

After small electric motors entered each machine, the central transmission shafts and belts gradually disappeared, the equipment could be rearranged according to the production sequence, and the factory buildings also shifted from multi-layer to more flexible single-layer layout. 

The efficiency improvement brought by electricity comes from the rearrangement of the production process together with the power system.

When AI enters enterprises, it also encounters similar thresholds. 

Many companies have integrated AI into daily office work, but few let it access contract approval, customer accounts, R&D release and financial systems. These tasks involve review, authorization, testing and responsibility confirmation. 

Without data, tools and permissions connected, the model can only stay outside the business process to provide assistance. 

What Alibaba demonstrated at the Yunqi Conference is how a task runs through these links. 

After a meeting ends, QwenNote Eva desktop robot or QwenNote A2 recording card leaves the on-site content, DingTalk and MyContext then complete the context from messages, documents and historical meeting records, and AgentCore subsequently manages execution permissions, progress and records. 

Tasks can also be initiated from mobile phones. The Honor Magic9 series will become the first batch of models equipped with Qwen Intelligence. From a sentence in the meeting room to specific actions in the enterprise system, Alibaba is trying to connect the previously disconnected links. 

What really determines how much work an Agent can do is no longer just the model parameters. 

It needs to know what happened in the company, who the customers are, where the files are, which systems can be called, and which things must be confirmed by humans. 

Context determines what the machine knows, tools determine what it can do, and permissions determine how far it can go.

When these are connected, AI will begin to evolve from a software that answers questions to a productivity that can continuously execute tasks. 

V. Machines Begin to Have Their Own "Working Hours"

When machines really start to work, what is missing is never just computing power. 

After an Agent gets a task, it needs to know the company's information, who the customers are, what happened before, which systems can be operated, and who to confirm with if something goes wrong. 

What enterprises give to AI is also changing: from a piece of Prompt, a document, to the complete business context, and then to tools, permissions and execution environment. 

Context determines what the machine knows, tools determine what it can do, and permissions determine how far it can go. When the three are connected, a goal can be split into dozens or even hundreds of actions: looking up materials, calling interfaces, writing files, running tests, waiting for feedback, and then proceeding to the next step. 

Thus, the workload of machines begins to have the concept of "working hours". 

A set of data disclosed by Ping An of China just provides a section: its daily average Token usage increased from about 30 billion to more than 300 billion in less than a year, and the computing power increased from 800PF to 1500PF. The computing power increased by less than double, but the Token usage expanded ten times. 

This is only data from a single enterprise, but it shows a growth path: 

The growth of enterprise AI can be manifested as that in the same set of businesses, machines begin to undertake more and more steps.

Another set of data disclosed by Ping An is also very interesting: AI has entered specific links such as customer service, medical treatment, and R&D, and more than 100 million AI services have been completed in less than half a year. 

Machines have obtained a batch of business scenarios that continuously generate tasks. 

This is exactly the difference between Agent and traditional AI tools. 

Model capability determines whether the machine can do the work, enterprise context determines what it knows to do, and tools and permissions determine whether it can keep going. As these links are gradually connected, enterprises have the opportunity to arrange a real "working hour sheet" for machines for the first time. 

Alibaba's 20GW computing power will eventually be applied here. 

No matter how many GW the data center has and how many cards the chips have, they all have to answer a more simple question: how much work can be handed over to machines every day? 

VII. When the Growth Unit Changes, the Company Will Also Change

If machines really become a new type of productivity, the first thing that may change is not the products, but the internal organizational structure of the company. 

After Wu Yongming took over as CEO in September 2023, he quickly listed "AI-driven" as Alibaba's strategic priority. By February 2025, Alibaba announced that it would invest at least 380 billion yuan in the construction of cloud computing and AI infrastructure in the next three years. 

The money invested in computing power has also begun to change the organizational