Have you noticed that the painting style of the AI circle has changed a bit faster in the past six months?

At this time last year, everyone was still partying for price reductions in computing power. Alibaba Cloud took the lead in shouting "The highest reduction60%". Tencent Cloud, Huawei Cloud, and Baidu Cloud all followed suit. The battle was like a Double Eleven promotion. In the entrepreneurial group, people post their bills every day: "Look, I only spent a few cents on my million Token!" At that time, everyone thought, AIThe spring of entrepreneurship is here, computing power is as cheap as cabbage, who can't afford large-scale model applications?

What's the result? The slap in the face came too quickly.

Just last month, the wind turned 180 degrees. Google, Amazon, Tencent, Alibaba, and Baidu all issued price increase announcements within 10 days. How much? Commonly ranges from 30% to 50%. The most ruthless one is Tencent Cloud. One of its core products has risen directly by 400%.

From "jumping off the building sale" to "rocket-like price increase", it took less than a year. What happened? Who is pushing prices up behind the scenes? More importantly, in this wave of price increases, who is suffering the most, and who is laughing?

It was still "breaking bones" last yearWhy did the collective price increase this year?

Let’s briefly review this "reversal drama".

2025In March 20254Alibaba Cloud took the lead in dropping a blockbuster: the highest price reduction for core products60%. This is not a small fight, it is a real "cut in half and then discount". Immediately afterwards, JD Cloud said "drop it as you like, I will follow", and Tencent Cloud, Huawei Cloud, and Baidu Cloud all followed suit. For a time, the computing power market was full of smoke, and the price war was a lively one.

What was the slogan back then? "Make AIaffordable" and "computing power inclusive". Many start-up companies really believed it and began to sell Token and run models with great fanfare.

However, a free lunch never lasts long.

2026Year1Month, AmazonAWSDid something quietly - without any press conference or any notice, directly raised the server price of EC2 by by about 15%. Although the magnitude is not large, it is of great significance: This is the first price increase in the cloud service industry in the past two decades. You know, in the past twenty years, AWS has cut prices more than a hundred times. It has always been downward, not upward.

This time, it was like knocking down dominoes.

3MonthOn November 11Tencent Cloud followed up with its Tencent HY2.0 InstructThe model input price is from0.0008yuan/ThousandstokensRising to0.004505yuan/Thousandstokens——Increase463%, four times more. On March 3Month18Alibaba Cloud announced the increase in computing power card products5%to34%, Baidu Smart Cloud has also risen5%to30%. Those large models that were previously released for free public beta, such as GLM 5, MiniMax 2.5, Kimi 2.5, all the "free prostitution period" have ended and have been transferred to formal billing.

FromFrom "rushing to lower prices" to "rushing to increase prices", why is the change so fast?

On the surface, it seems that the cloud vendors cannot handle it. GPUThe more chips you buy, the more expensive the chips are. The electricity bill of the data center accounts for 40%to60%, coupled with 2025The price of memory chips will also begin to rise in the second half of the year, and the pressure on the cost side is really coming. But what really makes the price increase a "have to do" is another more fundamental reason - computing power is really not enough.

Who is crazy"eating"Token? 140The truth of trillions

Didn’t we say there was excess computing power before? Why is there suddenly not enough?

The answer is:Token was "eaten" too quickly.

According to the data disclosed by Liu Liehong, director of the National Data Bureau in March this year: to2026Year3Month, China's dailyTokenThe volume of calls has exceeded140Trillion.

How exaggerated is this number? Give you two references:

· 2024At the beginning of the year, this number was only 1000billion. In two years, it has increased more than a thousand times.

· 2025At the end of the year, the number is 100trillion. In other words, in the past three months alone, it has increased by 40% - the new volume in these three months alone (40Trillions), which is 2024400 times the entire day at the beginning of the year.

This is not linear growth, this is a tsunami.

Then the question is: Who is consuming Token crazily?

The answer is just one word: Agent (Agent).

Since last year, products represented by open source intelligent agentOpenClaw (known as "lobster" in the world) have become very popular. AIFrom a robot that can only "chat", it has become an assistant that can "work" - helping you book flights, write code, and do PPT, analyzing data...sounds cool, right? But the price is that when the agent does a simple thing, it consumes Token10 times as much as a normal conversation. leaf="">100 times.

For example: letAI help you write a crawler script. If it is a normal conversation, it will give you a piece of code, you copy it and leave, consuming hundreds of Token. But if it is an intelligent agent, it has to run the code, report errors, debug, run again, adjust again... and go back and forth for more than ten rounds. TokenThe consumption is tens of thousands.

Not to mention the video generation"gold-eating beast". Some analysis points out that generating 1minutes of video consumes approximately 10trillionToken. However, the current video model only charges you a few cents to a few dollars to generate a 5 seconds of video - this is not making money, this is clearly making money at a loss. But there are so many people who can't stand it. Video, music, code, data analysis...every direction is "eating" like crazyTokenToken

Supply cannot keep up with demand, and the price of computing power will naturally increase. This is not a conspiracy, but a naked supply and demand imbalance.

Computing power ranking competition: big companies eat meat, small players can’t even drink soup

The price increase means completely different things to different people.

For cloud vendors, price increases are actually a good thing. A brokerage firm has calculated an account: every time Alibaba Cloud raises its price by 1%, its profit margin will increase by 1%. So the data you see is that Alibaba Cloud's share has increased instead of falling, accounting for 36% of China's AI cloud market. In the more detailed track of AIVolcano Engine (owned by Byte) accounts for nearly 50%——In other words, half of the Token calls in China go through the pipeline of the Volcano Engine.

At the same time, the shares of Huawei Cloud and Tencent Cloud have declined slightly. The head effect is becoming more and more obvious: Large companies are getting stronger and stronger, and their resources are becoming more concentrated.

Who is the worst?

Small and mediumAIStartup companies, as well as those small players who have just entered the game.

The reason is simple: rising prices directly push up their operating costs. In the past, when Token was cheap, you could run experiments and adjust models at will, and it wouldn’t cost much anyway. Now that the price has increased several times or even more than ten times, every round of training and every reasoning must be weighed and weighed.

What’s even more troublesome is that small players have no bargaining power. Large customers can sign long-term agreements with cloud vendors to lock in a relatively favorable price. You are a start-up company and you only spend tens of thousands of dollars a year on computing power. Who will negotiate a discount with you? You can only pay the increased price honestly.

Many projects that were originally intended to be AIwere shelved silently after settling the accounts. There are also some companies that are already doing it, either downsizing or carrying on with losses. However, the competition in the terminal market is fierce, so you don’t dare to increase the price for users easily - the one next door is still free, but as soon as you charge, all the users will run away. In the end, all the cost pressure can only be swallowed by oneself.

A practitioner complained to me:"I used to think that computing power was cheap and the threshold for starting a business was low. Now I realize that the threshold is not low. It lets you in first and then closes the door."

This is actually a cruel qualifying match. In the past two decades, cloud vendors have thrived on the strategy of "increasing volume at low prices, first acquiring territory and then making profits". But that era is over. Computing power officially bids farewell to the subsidy period and enters the commercial pricing stage. In the future, the competition will no longer be about who is cheaper, but whose service is more stable, whose ecology is more complete, and who can help enterprises make full use of every bit of computing power.

In this ranking match, there is a high probability that small players will be left behind.

Looking back at the roller coaster ride of more than a year, you will find a heartbreaking truth:

The computing power has gone from "cabbage price" to "rocket price", which is essentially the AIA microcosm of the industry's transition from barbaric growth to maturity. The free era is over, and value competition has begun. Those business models that rely on subsidies will die, while those products that truly have technology, scenarios, and users will survive in an environment of rising computing power costs, or even survive better.

AIThe core competitiveness of entrepreneurship has never been how cheap computing power is, but what you use it for.

In the age of computing power, Token is indeed expensive. But what is more expensive than Token is a brain that knows how to use Token well.