In the spring of 2026, &#34;raising lobsters&#34; became a sensation in the AI circle. OpenClaw, which allows everyone to have their own Jarvis, not only gained popularity but also fattened up the players selling large model tokens.
Due to the massive token requirements for Agent operations, API call volumes skyrocketed exponentially. As &#34;shovel sellers,&#34; domestic large model manufacturers like Zhipu, MiniMax, Kimi, and Alibaba Cloud launched API packages for the Coding Plan to seize this windfall, with Zhipu's top-tier package even selling out.
During this period, &#34;shovel sellers&#34; made a fortune, with revenues surging, and the capital market experienced an unprecedented valuation frenzy.
Zhipu and MiniMax saw their market capitalization soar in the Hong Kong stock market, both surpassing HKD 300 billion, becoming the &#34;measuring unit&#34; for large companies' valuations. The yet-to-be-listed Kimi achieved a valuation exceeding USD 10 billion, setting a record for domestic large model unicorns.
Currently, as AI transforms into an automated machine capable of self-cycling and consuming tokens wildly for error correction and execution, the speed of computing resource consumption has exceeded everyone's imagination.
The &#34;shovel sellers&#34; are pulling out all the stops—adapting to the cloud, one-click deployment, platform-level compatibility... trying to deeply bind this hot &#34;lobster&#34; and seize market mindshare.
Meanwhile, behind the scenes, a divergence in the route dispute is brewing, as the purely &#34;token-selling&#34; business model faces potential crises. OpenClaw and other open-source frameworks expose significant security risks and unstable calling experiences, and their development barriers are not high.
Major companies are slowly realizing that the real moat lies not in open APIs, but in locking user data contexts and business scenarios through self-developed localized development tools and edge-side intelligent agents.
A covert ecological battle from &#34;selling water in the cloud&#34; to &#34;building cities locally&#34; has already begun.
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/7307b76b-2176-4f87-b122-8d5c62381877.jpeg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/7307b76b-2176-4f87-b122-8d5c62381877.jpeg"/>
<h2>Spring Carnival</h2>
In this sudden surge of traffic, leading domestic large model manufacturers became the most direct beneficiaries.
Alibaba Cloud's Qianwen, Kimi from Moonlight, MiniMax, and Zhipu's GLM all experienced unprecedented revenue spikes during this Spring Festival.
&#34;I didn't watch the Spring Festival Gala on New Year's Eve; I was focused on the token consumption in the backend. The speed of those numbers jumping was both exciting and terrifying,&#34; an independent developer told Wall Street Insight.
It is understood that the small team he belongs to is trying to build a fully automated cross-border e-commerce customer service and shipping system using OpenClaw. To ensure the smooth operation of the system, the team tested multiple top large models from both domestic and international sources.
The developer admitted that the latency and high pricing of overseas models made it difficult for them to bear, and to get OpenClaw to run a fully automated customer service and web scraping system, they continuously purchased high-tier computing packages from Alibaba Cloud and Zhipu
What makes it magical for him is that this digital virtual product has actually turned into a &#34;Spring Festival train ticket&#34; that needs to be snatched up.
During the Spring Festival, major companies like Zhipu launched Coding Plan computing power packages specifically for high-frequency developers, with some products selling out in a short time. Even when the platform later announced structural price increases for some advanced API interfaces, it still couldn't stop developers' enthusiasm for recharging.
This strong purchasing power emerging from the B-end and developer side has finally allowed all the large model manufacturers that have been struggling to find monetization paths over the past year to see the dawn of real profits.
The supply-demand imbalance in computing power has transmitted to the capital market at an unprecedented speed, directly giving rise to one of the craziest wealth myths at the beginning of 2026.
The shovel sellers not only benefited on the business side but also experienced a nuclear-level leap in valuation and stock price. Using real data to review the capital movements over these few days is enough to shock any traditional internet giant.
Zhipu, which just completed its IPO on the Hong Kong Stock Exchange in early January, staged a violent &#34;Welcoming the God of Wealth&#34; on February 20 (the fifth day of the Lunar New Year). Stimulated by the latest GLM-5 large model open-source release and the 30% price increase of the API programming subscription package that was still sold out, Zhipu's stock price surged nearly 43% in one day, closing at HKD 725, with a total market value directly breaking HKD 323.2 billion.
This AI company, established only a few years ago, instantly surpassed Kuaishou in market value and is now on par with the established e-commerce giant JD.com, just one step away from surpassing Baidu.
Following closely is MiniMax, which also went public in Hong Kong in January. With an exceptionally strong full-modal product line and an overseas revenue ratio of up to 70%, MiniMax surged over 14% on the first trading day of the Year of the Horse, with its market value also strongly breaking through the HKD 300 billion mark.
Goldman Sachs, in a subsequent research report, directly gave it a valuation expectation of nearly USD 39 billion, defining it as a scarce asset among the world's top foundational models.
In the unlisted unicorn camp, the actions of the Dark Side of the Moon (Kimi) are even more aggressive.
During the Spring Festival, Kimi completed a new round of financing exceeding USD 700 million, backed by giants like Alibaba and Tencent. Combined with the USD 500 million financing completed just over a month ago, its total financing amount in a short time has exceeded USD 1.2 billion.
This not only doubled Kimi's valuation to break through the USD 10 billion mark, making it the fastest large model enterprise in China to become a &#34;decacorn,&#34; but its newly released K2.5 large model also created a commercial miracle of &#34;total revenue in nearly 20 days exceeding the total revenue for the entire year of 2025.&#34;
Some large model practitioners pointed out to Wall Street that the market has initially accepted the business logic of AI as a core productivity tool.
<h2>Removing the Filter</h2>
Beneath the revelry, there are indeed concerns.
As OpenClaw's explosive popularity elevates underlying computing power to a pedestal, the industry and hardcore developers are beginning to calmly examine this ecosystem led by third-party open-source projects.
In forums, a large number of complaints wrapped in praise have unveiled the fragile side of the purely API-selling model
Multiple programmers who participated in the early testing revealed that OpenClaw's performance in actual office scenarios is far from as smooth as shown in the demonstration videos.
When it was asked to handle local projects containing dozens of nested files, the rough scheduling logic of the open-source framework often caused the large model to get stuck in a deadlock.
The key is that to maintain this high-frequency interaction, OpenClaw sends dense API call requests in a short period.
A testing team found that when they tried to use Kimi's API to drive local search tasks, they frequently triggered the platform's &#34;429 Too Many Requests&#34; anti-spam error mechanism, causing the entire automation workflow to collapse.
In addition to the fragmented experience, what frightens enterprise users even more is the security issue.
The core logic of frameworks like OpenClaw requires users to grant the system the highest read and write permissions.
An architect working at a leading security company told Wall Street Insights that handing over complete access to core confidential code or commercial databases to an open-source script that is modified daily by various anonymous hackers is akin to &#34;running naked&#34; on the company's servers.
&#34;Once the large model experiences hallucinations, or if a dependency package of the open-source framework is implanted with malicious code, the system could execute catastrophic deletion or data leakage commands in milliseconds.&#34; In their view, this determines that projects like OpenClaw can initially only be toys for the geek community and cannot truly enter the data centers of financial, governmental, and large enterprise clients willing to pay high fees.
The seemingly magical system has a low development barrier behind it. It is essentially just middleware that connects prompts, large model interfaces, and operating system command lines. The core intellectual suppliers still come from large models like those of Alibaba and Zhipu.
Since the technical threshold of the framework is not high and the open-source solutions have irreparable security risks, large companies naturally do not want to be just an obscure &#34;provider of computing power, water, coal, and electricity&#34; that could be replaced at any time.
This leads to the most critical trend prediction in the large model track after the Spring Festival computing power melee: those who sell shovels will inevitably step onto the field themselves, fully turning to the independent development of localized intelligent tools to build their own walls.
<h2>Preparing for Counterattack</h2>
Major companies have already keenly sensed that the future of AI competition does not lie in who has the lowest API price, but in who can root themselves most deeply in the user's local workflow.
A practitioner in the large model industry told Wall Street Insights that in the past, when users called APIs in third-party tools, the large model could only see the few lines of code or requests that the user was currently inputting, like peering at a leopard through a tube.
If a major company officially launches a deeply integrated integrated development environment (IDE) plugin or local computer manager, this tool can silently read the user's entire project architecture, historical coding habits, and even the company's internal document library in a compliant sandbox environment.
This deep understanding of local context will lead to a qualitative leap in the output quality of official tools, completely crushing those external open-source frameworks that can only conduct single conversations.
In fact, this strategy of deeply embedding large model capabilities into local and national-level applications has long been in the making.
For example, Alibaba has already begun to deeply integrate Qianwen into DingTalk's &#34;D Plan,&#34; as well as into everyday high-frequency applications like Amap.
The core logic of this strategy is consistent: no longer simply relying on users to actively ask questions on the web, but instead surrounding users' scenarios with their own end-side tools and national-level applications.
In the development field, whether it's ByteDance's MarsCode local native environment that is gaining momentum, or Alibaba Cloud's accelerated iteration of the Tongyi Lingma privatized version, the ambitions of the big companies are evident.
They aim to create a secure, compliant, and customized &#34;body&#34; for their large models.
Through the officially launched localized intelligent agent tools, large companies can provide enterprise clients with comprehensive operational auditing, permission isolation, and data de-identification services, completely eliminating the security pain points of open-source frameworks.
More importantly, once developers become accustomed to this set of official tools, the entire operational logic of their projects and API calling standards will be tightly bound within the ecosystem of that large company.
This is akin to the competition between Apple and Android in mobile operating systems; large model vendors are vying for control of the operating system in the AI era through local tools.
Once a path dependency is formed, large model companies will truly possess a moat that cannot be easily replaced.
<h2>The Heat of Tokens</h2>
Since large companies are racing towards localization and ecosystem development, how long can the current hot situation, which relies solely on selling tokens and even sees API quotas sold out, continue?
Industry insiders believe that, considering recent industry trends and the long-term logic of the capital market, the period of exorbitant profits from simply selling basic APIs may be shorter than expected.
As the underlying model capabilities gradually converge and the large-scale deployment of computing power infrastructure (such as domestic GPU alternatives) occurs, the prices for basic instruction API calls will inevitably lead to a price war.
In the recent earnings expectations of major companies, it can also be seen that although the current demand for computing power exceeds supply, the long-term improvement of gross margins still faces challenges.
The future API market will inevitably move towards stratification.
For extremely complex tasks that require hundreds of billions of parameter models for deep logical reasoning, high-level APIs will still maintain a high premium, becoming a strategic resource similar to industrial electricity. However, for massive and high-frequency simple automation operations, large companies will undoubtedly adopt an &#34;end-cloud collaboration&#34; architecture.
Therefore, many developers believe that future localized tools will deploy lightweight small models on users' computers or mobile devices, using local computing power to handle basic loops and retrieval tasks, thereby significantly reducing cloud token costs and improving response speed.
This Spring Festival frenzy triggered by OpenClaw's breakout seems to be a victory for the underlying model's computing power, but in reality, it is a rallying call for the industry to undergo a major reshuffle at the application layer.
Those unicorns whose valuations have skyrocketed in the capital market now hold over ten billion dollars in ammunition
Their next core task is not merely to purchase more graphics cards to stack the next parameter monster, but to quickly encapsulate this intelligence into indestructible localized tools and edge applications.
For large companies, only by turning the shovel into an indispensable industrial-grade excavator and firmly controlling the entrance to the construction site can they laugh last in the true showdown of the AI era.
In the future business landscape, merely having a &#34;smart brain&#34; is not enough. Those who can create the safest and most user-friendly &#34;localized embodiment&#34; for this brain, and those who can embed this set of tools deeply into human daily workflows, will be the ones who can truly sit firmly on the Iron Throne of artificial intelligence infrastructure in the next decade.
The awakening of the shovel sellers heralds the brutal second half of a big fish eating small fish and ecological unification, which has just begun

MSFT

GOOGL

NVDA

The second half of the battle for the Agent entry among major companies

- In spring 2026, the AI sector experienced a surge in demand for OpenClaw, driving up sales for major model providers like Zhipu and MiniMax, whose market valuations soared.  
- As companies responded to escalating token consumption from AI agents, a shift from relying solely on API sales to developing localized intelligent tools became apparent to establish stronger user relationships.  
- This transformation indicates an impending competitive landscape where firms must integrate AI capabilities deeply into user workflows to maintain their market positions amidst evolving security and operational challenges.

A "Spring Brawl" of OpenClaw selling shovels