By Xie Zefeng, Edited by Yang Xuran
Like almost every other business sector, AI, as an industry, has reached its own watershed moment this year.
Doubao and WorkBuddy have begun testing paid models, while ChatGPT and Anthropic have already validated the token subscription model overseas. Companies such as UBTECH, Unitree, and Tashi Intelligent Navigation are also exploring the scenario-based deployment of humanoid robots.
From the consistent actions of key players, it is evident that the critical period for AI commercialization is approaching. Especially after the industry collectively crosses the "innovation death valley" of large models, generating revenue quickly has become a shared priority.
Entering 2026, the Scaling Law, originally adhered to by the industry, has hit a high wall: the returns from simply hoarding computing power and increasing data volume to enhance model capabilities have begun to diminish. Spending billions of dollars to train the next generation of super large models has not yielded a significant leap in performance.
"For a startup, betting massive resources to train ultra-large parameter models with extremely low cost-effectiveness is certainly not a pragmatic choice," Kai-Fu Lee once reflected.
In the capital market, investors no longer blindly flock to offer equally high valuations. The diverging stock price trends of Zhipu and MiniMax, as well as Kimi's high-frequency financing, all illustrate this point.
At this juncture, the AI industry chain is bidding farewell to "0 to 1" and seeking the commercial singularity of "1 to 100". All criteria for judgment and evaluation are changing.
This article is an in-depth value piece from the content team of Wave of the Tide. We welcome you to follow us across multiple platforms.
The Brick Wall
In the early hours of November 30, 2022, Sam Altman quietly posted a tweet: ChatGPT is officially launched. At that time, OpenAI was merely a small startup known only to a few insiders, and Altman did not hold high expectations for the initial form of GPT. However, just five days later, the goal of 1 million users was easily surpassed; two months later, over 30 million users flocked to ChatGPT.
As early as 2020, OpenAI proposed the Scaling Law: the larger the training parameters and the more computing power, the better the model performance. Over the next three years, AI large models advanced along this path. Whoever had more parameters, greater computing power, and higher benchmark scores could secure more financing and higher valuations. From GPT-3's 175 billion parameters to GPT-5.6 Sol's approximately 3 trillion (3T) total parameters, "stacking parameters" almost became the development belief of the entire industry.
Anthropic took this to the extreme. Its Mythos, described as "too powerful to be made public," boasts 10 trillion parameters, with estimated training costs reaching $10 billion.
The tech giants simultaneously unleashed their "money power," striving to hoard chips. Estimates show that Meta alone possesses GPU computing power exceeding the total of all AI companies in China. Google's total AI computing power is equivalent to 5 million NVIDIA H100s, accounting for a quarter of the global total.
However, the issue lies in the fact that when funding and computing chips hit bottlenecks, the performance growth curve of large models will inevitably flatten. For Chinese companies, which face restrictions on high-end chips and lack sufficient financial reserves, it is even more critical to re-examine the Scaling Law.
Early last year, Kai-Fu Lee, founder of 01.AI, reflected: "We have realized the diminishing marginal returns of the Scaling Law. Betting massive resources to train ultra-large parameter models with extremely low cost-effectiveness is not a pragmatic choice for startups."
01.AI subsequently shifted its strategy. Yi-Lightning adopted a MoE (Mixture of Experts) architecture, activating only over 20B parameters, yet achieving superior model performance. DeepSeek also leveraged MoE architecture optimization, open-source strategies, and distillation techniques to reduce training costs to less than one-twentieth of that for ChatGPT.
Abandoning the pursuit of super large models in favor of building "small but beautiful" pragmatic products may be the consensus among Chinese large model companies after experiencing the "hundred models war" and the "price war". The shift from "building larger models" to "using models to build profitable businesses" signifies that the innovative heroism of the geek era is giving way to business pragmatism.
Closed Loop
Although the competition strategies for large models in China and the US differ, they converge at the commercialization level.
On June 24, ByteDance's Doubao officially launched a paid model with three subscription tiers: Standard Edition at 68 RMB/month, Enhanced Edition at 200 RMB/month, and Professional Edition at 500 RMB/month. Behind the paywall lies exorbitant cost expenditures: as of March 2026, Doubao's daily token usage exceeded 1.2 quadrillion, doubling in the past three months and reaching a full 1,000 times that of its initial launch in May 2024.
Zheshang Securities estimates that ByteDance's capital expenditure in 2025 was approximately 150 billion RMB, of which 90 billion was spent on procuring AI computing power, resulting in a daily expenditure of 438 million RMB. Meanwhile, the "AI + e-commerce" model explored by Doubao can currently only generate a limited scale of revenue. Therefore, revenue from the Professional Edition becomes even more crucial for Doubao.
Almost simultaneously, Tencent also pressed the acceleration button for commercialization. Its Buddy AI (including WorkBuddy, CodeBuddy, etc.) announced a subscription upgrade, effective July 1, with three pricing tiers: Standard Edition at 99 RMB/month, Advanced Edition at 199 RMB/month, and Flagship Edition at 999 RMB/month.
The direct driver for the paid model is the explosion in invocation volume. WorkBuddy's per capita token consumption increased tenfold within three months. Although Tencent emphasizes that AI is still in a "strategic investment period," the launch of the paid model confirms that as AI services shift from novelty to high-frequency use, their cost expenditure characteristics will force the formation of a commercial closed loop.
Across the ocean, OpenAI has adopted a multi-pronged approach, launching a combination of "advertising + subscription + token payment". It introduced the Plus subscription ($20/month) in February 2023, and this year launched the even cheaper ChatGPT Go ($8/month). Paying users increased from 47 million at the end of last year to 55 million in the first quarter.
Furthermore, ChatGPT even plans to offer advertising services to enterprise clients, charging free and Go users based on impressions.
Even so, its revenue struggles to offset the massive cost expenditures, leading to long-term huge losses.
Anthropic has developed a unique business philosophy: focusing on government and enterprise clients, with no free tiers, no ads, and no price cuts. Starting in April, it banned subscription quotas for third-party tools like OpenClaw, mandating API pay-as-you-go, filtering high-value users out of the subscription model and into enterprise channels.
By the end of 2025, Anthropic's ARR was approximately $9 billion, swelling to $30 billion by April 2026. In February, only over 500 enterprises were willing to pay more than $1 million annually; by early April, such clients exceeded 1,000.
At the Code with Claude conference, Dario Amodei stated that the company originally prepared for a maximum10x annual growth, but in the first quarter, revenue and usage growth briefly reached 80x (on an annualized basis).
Superior product strength and unique business strategies have propelled Anthropic's Q2 revenue (estimated) to $10.9 billion, while achieving an operating profit of $560 million. It is set to become the benchmark for the first profitable large model enterprise globally, with its valuation approaching a trillion dollars.
Overall, competition in the AI industry chain has shifted from a pure performance war to an era that places greater emphasis on commercialization.
Divergence
The valuation anchor for AI companies in both primary and secondary markets is shifting from technical capability to commercial capability, leading to a divergence in valuations among enterprises.
The most typical cases are Zhipu and MiniMax. These two large model companies, which went public one after the other, have recently drawn widespread attention as 1 Zhipu now equals approximately 10 MiniMax.
Initially, the market actually favored MiniMax's multimodal and C-end AI applications, while Zhipu's story of localized deployment for government and enterprise clients was less favored by capital.
Just three months later, their performances completely reversed. MiniMax faced its "lock-up expiry" test on July 9, with its stock price plunging 18% that day. It subsequently continued to plummet, falling 83% from its mid-March peak, wiping out approximately340 billion HKD in market value. Zhipu, whose lock-up expired the previous day, remained remarkably resilient, rising 30% over two consecutive days.
Looking back now, the turning point in their market performances occurred on June 1 this year.
That day, MiniMax released the M3 model, switching from the previous subscription model to token-based billing. Negative public opinion subsequently fermented: the company canceled the 29 RMB/month Starter package without communicating with users; developers testing it reported that token consumption for identical tasks far exceeded expectations, with costs in high-token scenarios rising from 49 RMB to 175 RMB.
Users dubbed this behavior a "backstabbing" price hike, and MiniMax's stock price plunged 16% that day. Yan Junjie immediately announced he would no longer draw a salary and allocated 4% of his personal shares for talent incentives.
However, short-term promotions attempting to retain users with low prices failed; after the promotions ended, the M3 model's results fell short of expectations, leaving MiniMax in the dual dilemma of "unable to raise prices and unable to retain users."
Since June, the narrative logic in the large model track has accelerated its shift towards Vibe Coding and Agents. Led by Anthropic, B2B genes have begun to be favored by investors. In contrast, Zhipu's GLM-5.2 is not only powerful, but its Q1 API invocation pricing increased by 83%, yet invocation volume still grew by 400%, even facing supply shortages.
The degree and effectiveness of commercialization in the AI field are now directly linked to capital market valuations. From this perspective, those who still stubbornly equate AI with the "geek vibe" need to gain a renewed understanding of the entire industry.
In AI terminals, commercialization is equally important.
Beyond large models, autonomous driving is regarded as the commercial track where AI will land first and fastest, opening the door to physical AI. In reality, commercial applications in this field are rapidly maturing.
Compared to humanoid robots, autonomous driving has more advantages for deployment. First, market demand is clearer, and payment scenarios are well-defined. High-frequency travel demand drives the willingness to pay for advanced driver-assistance features, and ride-hailing platforms are also willing to deploy autonomous driving technology. Second, hardware and software possess mass-production standardization characteristics, allowing for cost reduction through later scaling, with a clear commercial closed loop. Moreover, once vehicles are on the road, they naturally form a data collection flywheel, where real-world road data helps iterate AI algorithms, continuously improving product experience.
Currently, high-level advanced driver-assistance systems (L2/L2+) for C-end pre-installed vehicles have almost become a must-have for new forces, with Tesla's FSD and Huawei's ADS (HUAWEI ADS) standing out in the track. Meanwhile, L4-level autonomous driving (Robotaxi, unmanned freight, etc.) is also on the verge of commercialization, with Baidu's Apollo Go, WeRide, Momenta, and Waymo actively laying out their strategies.
The maturity of smart glasses products is also deepening. Previously, the Humane AI Pin and Rabbit R1 were hugely popular but ultimately failed due to a lack of a commercial ecosystem. Today, Meta Ray-Ban smart glasses have won out with a mature ecosystem, selling over 7 million units in 2025, with this year's production capacity target raised to a maximum of 20 million units.
Conclusion
From "showing off skills" to "selling goods" is the shared expectation of capital, the market, and consumers for the entire AI industry chain.
We see that from the primary market paying for "imagination" to ARR becoming the valuation anchor, the entire AI track is undergoing a profound transformation from a "geek vibe" to "business integration."
2026 is destined to be a pivotal year for AI to move from technical validation to large-scale commercial promotion. This is also the inevitable path that all emerging technology industries have traversed throughout business history.