Token becomes a 'money pit,' major firms' computing power surges

Author | Chuan Chuan

Editor | Da Feng

Have you noticed how fast the AI community's vibe has shifted in just half a year?

At this time last year, everyone was still celebrating the drop in computing power costs. Alibaba Cloud led the charge by announcing a maximum price cut of 60%, followed closely by Tencent Cloud, Huawei Cloud, and Baidu Cloud. The scene resembled a Black Friday sale. In entrepreneurial groups, people were showing off their bills daily: 'Look, I spent only a few cents for a million tokens!' Back then, it felt like spring had arrived for AI startups, with computing power being as cheap as cabbage. Who couldn't afford to develop large model applications?

But what happened? The slap in the face came quickly.

Just last month, the trend made a complete U-turn. Google, Amazon, Tencent, Alibaba, and Baidu all issued price hike announcements within 10 days. By how much? Generally 30% to 50%. The most aggressive was Tencent Cloud, with one core product skyrocketing by 400%.

From 'fire-sale discounts' to 'rocket-like price hikes,' less than a year passed. What exactly happened? Who is pushing prices up from behind the scenes? More importantly, in this wave of price increases, who is suffering the most, and who is secretly smiling?

Last year, they were offering rock-bottom prices; why the collective price hike this year?

Let’s briefly recap this 'plot twist.'

In April 2025, Alibaba Cloud dropped a bombshell: a maximum price cut of 60% on core products. This wasn't just a minor discount—it was a real 'slash-and-burn' strategy. Shortly after, JD.com said, 'Go ahead and lower your prices, we'll match them,' while Tencent Cloud, Huawei Cloud, and Baidu Cloud all followed suit. Within no time, the computing power market was filled with smoke from fierce price wars that created quite a spectacle.

What was the slogan back then?'Make AI affordable' and 'universal access to computing power.'Many ventures truly believed it, starting to aggressively burn through tokens and run models.

However, there’s no such thing as a free lunch that lasts forever.

In January 2026, Amazon AWS quietly took an action — without any press conference or prior notice, they directly raised the price of EC2 servers by approximately 15%. Though the increase seems modest, its significance is enormous: this marks the first price hike in the cloud service industry in nearly two decades. Keep in mind, over the past twenty years, AWS had cut prices more than a hundred times, always moving downward, never upward.

This move was like knocking over the first domino.

On March 11, Tencent Cloud followed suit, increasing the input price of its Tencent HY2.0 Instruct model from 0.0008 yuan per thousand tokens to 0.004505 yuan per thousand tokens — a staggering 463% rise, more than fourfold. On March 18, Alibaba Cloud announced price increases of 5% to 34% for its computing power card products, while Baidu Intelligent Cloud also raised prices by 5% to 30%. Those previously free beta large models, such as GLM 5, MiniMax 2.5, and Kimi 2.5, all ended their 'free trial period' and transitioned to official billing.

From 'rushing to cut prices' to 'rushing to raise prices,' why has the change been so rapid?

On the surface, it appears that cloud providers can no longer bear the burden. GPU chips are becoming increasingly expensive, with electricity costs for data centers accounting for 40% to 60% of operational expenses. Additionally, storage chip prices began to rise in the second half of 2025, bringing real cost pressures.But what truly made raising prices a 'necessary evil' is a more fundamental reason — computing power is genuinely running out.

Who is voraciously 'consuming' tokens? The truth behind the 140 trillion figure.

Wasn’t there talk of an oversupply of computing power before? How did it suddenly become insufficient?

The answer is:Tokens are being consumed too quickly.

According to data disclosed by Liu Liehong, director of the National Data Administration, in March this year: By March 2026, China's daily token usage will have exceeded 140 trillion.

How staggering is this figure? Here are two reference points:

· At the beginning of 2024, this number was only 100 billion. In two years, it increased more than a thousandfold.

· By the end of 2025, this figure stood at 100 trillion. This means that in just three months, it surged another 40%—the additional volume in these three months alone (40 trillion) is 400 times the total daily volume at the beginning of 2024.

This is not linear growth; it’s a tsunami.

So, the question arises: Who is consuming tokens at such a frenetic pace?

The answer is one word:Agents.

Since last year, products represented by the open-source agent OpenClaw (commonly known as “Lobster”) have become wildly popular. AI has evolved from being merely a “chatting” robot to becoming an assistant that can “get things done”—booking flights, writing code, creating PowerPoint presentations, analyzing data… Sounds cool, right? But the cost is that performing even a simple task consumes 10 to 100 times the amount of tokens compared to a regular conversation.

For example: Let AI help you write a web crawler script. In a normal conversation, it gives you a piece of code, you copy it and leave, consuming a few hundred tokens. But if it’s an intelligent agent, it has to run the code itself, encounter errors, debug, rerun, readjust... back and forth for more than ten rounds, with token consumption skyrocketing to tens of thousands.

Not to mention video generation, which is a real 'money guzzler.' Some analyses point out that generating one minute of video consumes about 10 trillion tokens. Yet current video models charge only a few cents to a few dollars for generating five seconds of video — this isn't making money; it's clearly operating at a loss just to attract users. But with so many people using it — video, music, code, data analysis — every direction is voraciously consuming tokens.

Supply can't keep up with demand, so computing power prices naturally rise. This isn't a conspiracy; it's a straightforward case of supply and demand imbalance.

The computing power race: Big companies take the lion’s share, small players can’t even get a sip.

Price hikes mean completely different things to different people.

For cloud providers, the price increase is actually good news. Brokerage firms have done the math: For every 1% price hike by Alibaba Cloud, its profit margin increases by one percentage point. So the data you see shows that Alibaba Cloud’s market share has not dropped but instead risen, capturing 36% of China’s AI cloud market. And in the narrower track of AI usage, Volcano Engine (under ByteDance) holds nearly 50% — meaning half of all token usage in China goes through Volcano Engine’s pipelines.

Meanwhile, Huawei Cloud and Tencent Cloud are seeing slight declines in their market shares. The head effect is becoming increasingly evident.Big players are getting stronger, and resources are becoming more concentrated.

So who suffers the most?

Small and medium-sized AI ventures, as well as those new entrants to the game.

The reason is simple: the price hike has directly driven up their operating costs. When tokens were cheap, you could run experiments and tweak models freely since it didn't cost much. Now that prices have risen several or even a dozen times, every round of training and every inference task requires careful consideration.

To make matters worse, small players lack bargaining power. Large clients can sign long-term agreements with cloud providers to lock in relatively favorable prices. But if you're a startup spending only a few thousand dollars a year on computing power, who's going to offer you a discount? You just have to pay the increased price without any choice.

Many projects that originally planned to develop AI applications have quietly shelved their plans after crunching the numbers. Some ongoing projects have either scaled back or are struggling with losses. The problem is that competition in the end market is fierce—you can't easily raise prices for users when your competitor is offering services for free; otherwise, all your users will leave. In the end, all the cost pressures have to be absorbed by yourself.

A practitioner complained to me, 'We used to think that cheap computing power lowered the barrier to entrepreneurship. Now we realize that the threshold wasn't lowered—it let us in first before closing the door.'

This is actually a brutal ranking competition. Over the past two decades, cloud providers thrived using the strategy of 'low prices for high volume, securing market share before turning a profit.' But that era is over.Computing power has officially ended its subsidy period and entered the phase of commercial pricing. The future competition will no longer be about who is cheaper but rather who offers more stable services, who has a more complete ecosystem, and who can help enterprises truly make the best use of every unit of computing power.

In this ranking competition, small players are likely to be left behind.

Looking back at the roller coaster ride of the past year, you'll discover a rather painful truth.

The rise in computing power costs from 'rock-bottom prices' to 'skyrocketing prices' essentially reflects how the AI industry has transitioned from unbridled growth to maturity. The era of free services is over, and value-based competition has begun. Business models reliant on subsidies will fail, while products with real technology, applicable use cases, and actual users will survive—and perhaps even thrive—in an environment of rising computing costs.

The core competitiveness of AI startups has never been about how cheap computing power is, but what you do with it.

In the era of computing power, Token is indeed valuable. But what's even more valuable than Token is a mind that knows how to make good use of Token.

49K Views