Big Techs' Covert Battle for the 100-Billion-Yuan Market: Positioning and Uncertainties in the AI Coding Track
In the leaked closed-door exchange record between Liang Wenfeng and investors that was later removed, it is mentioned that the top priority for DeepSeek at the current stage is the Coding Agent.
Just a month ago, at the Volcano Engine Conference, while the public was immersed in praise for Seedance reaching the global SOTA level, ByteDance seemed somewhat absent-minded. Tan Dai, President of Volcano Engine, stated that the internal team has always regarded Coding as a core direction.
What is more thought-provoking is that a few days ago, IDC suddenly released a market share data on AI programming, which was widely reported by the media: Alibaba's intelligent programming platform Qoder, launched for less than a year, captured 47.6% of the revenue share. In May this year, Liu Weiguang, President of Alibaba Cloud Public Cloud, publicly said that Coding is our most important direction, and it is almost for everything.
All signs indicate that almost all major domestic tech companies are accelerating their layout of AI Coding, which has become one of the most fiercely competitive fields.
ByteDance Trae, Alibaba Qoder, Tencent CodeBuddy, Zhipu CodeGeeX, SenseTime Raccoon, Baidu Comate, Huawei CodeArts, Moonshot AI Kimi Code... This list can keep getting longer.
The reason behind this is not complicated: AI Coding has become one of the largest consumption scenarios for model inference computing power, and its commercial value is easier to measure and reflect.
Claude Code is a learning sample for global large model companies. This pioneering popular Coding application has seen its ARR (Annual Recurring Revenue) skyrocket from 17 million US dollars in April 2025 to 2.5 billion US dollars in February 2026, with ARR exceeding 1 billion US dollars only 6 months after launch. Although Anthropic is a private company, its executives have been actively disclosing information to the public, which stimulates the sensitive nerves of the market every time.
Even OpenAI, the leader of large model companies, is following Anthropic's footsteps and has reaped considerable benefits.
In 2024, OpenAI did not have an independent Coding product, which was embedded in ChatGPT Plus/Team. In mid-2025, Codex was launched independently and began to gain momentum. At the end of January this year, its ARR exceeded 1 billion US dollars, and it became one of the core growth drivers of the company.
For domestic companies that have not yet found a definite profitable direction, this is a huge temptation and encouragement.
However, there is another important reason why major tech companies are betting heavily on Coding: Coding capability is the foundation, and the key to determining the upper limit of Agent effect. Therefore, the current significance of AI Coding to the industry has long gone beyond the scope of a productivity tool.
However, in that long list of companies betting on Coding, not everyone can stay. Major tech companies are doing everything possible to achieve a curve overtaking as they did in the past.
Domestic development started relatively late
There is no doubt that AI Coding is a key direction that both Chinese and American large model companies are focusing on. The industry has evolved from the previous situation where Claude Code dominated the market to a current pattern of diversified development.
In February 2025, Anthropic quietly released Claude Code, taking the industry into the Agentic Coding stage for the first time. The focus of industry competition has also shifted from "whether AI can write code" to "whether AI can independently complete software engineering".
This AI agent that can directly operate computers and independently complete programming tasks brought more than 1 billion US dollars in annualized revenue to Anthropic in just six months.
Even leading global companies are customers of Claude Code. Meng Xing, a partner at 5Y Capital, shared his observation of Silicon Valley in April this year: as a trillion-dollar market value company, all Meta employees are using Claude Code. "Half a year ago, this was completely unimaginable, because code is the core asset of the company. How could you let the API of an external company access it?" The reason why they have to use it is very simple: the self-developed coding product of Meta "is not easy to use, and no one uses it."
When Claude Code formed a dominant position, OpenAI took action.
Before 2025, OpenAI did not have an independent Coding product, and coding requirements were mixed in ChatGPT subscriptions and APIs. From the middle of that year to this year, its product Codex has shown a fierce counterattack momentum.
In September 2025, the usage of Codex was only 5% of that of Claude Code. By January 2026, this ratio had approached 40%. Especially after the launch of GPT-5.5, OpenAI has won back the trust of many developer users in terms of code capability, tool invocation and Agentic Coding.
In mid-May, Sam Altman, founder and CEO of OpenAI, added fuel to the fire. He announced on X that in the next 30 days, the company will provide two months of free usage for companies that want to switch from Claude Code to Codex.
"We can see that Codex is gaining obvious momentum this quarter, and many developers are switching from Claude Code to Codex, especially after Anthropic itself encountered traffic restrictions, price adjustments and reputation fluctuations, OpenAI seized this window period," Henry Yin, founding partner of MoE Capital, said in a recent interview with LatePost.
He also pointed out that Coding is no longer just an application scenario. It is not only the most important source of revenue at present, but also the basic capability for many cutting-edge breakthroughs in the future. This is a track that no large model company can afford to fall behind on.
This awareness quickly spread to China.
In 2025, Hong Dingkun, Vice President of ByteDance Technology, also said that Coding, as a highly structured and logically rigorous task, has high requirements for the model to understand complex semantic structures, logical reasoning, algorithm design and accurate expression, and can help explore the upper limit of model intelligence.
However, the overall pace of domestic development in the Coding scenario is still a little slow. Tan Dai explained in an interview with 36kr that in addition to the fierce global competition of LLM (Large Language Model), the most important reason is that the Coding direction was first defined and invested in by Anthropic and OpenAI.
Now, all of this is accelerating.
An important part for major tech companies to complete their AI strategy
Since the beginning of this year, with the rapid improvement of the Coding capability of self-developed models, major domestic tech companies are gradually reducing their dependence on overseas models.
The most typical example is Alibaba. On July 3, Alibaba issued an internal notice, announcing that starting from July 10, it will completely ban internal employees from using all Anthropic products including Claude Code in the office environment, and recommend using the self-developed Agentic Coding platform Qoder as an alternative. Shortly afterwards, Ant Group also issued a similar internal notice.
"Newberry" learned that Tencent and ByteDance have not explicitly banned the use of Claude Code internally for the time being, but the general direction is to tilt towards self-developed models.
CodeBuddy, the AI Coding tool promoted by Tencent, supports a variety of mainstream models such as Hunyuan, Claude, GPT and Gemini. With the continuous improvement of Hunyuan's Coding capability, Tencent continues to strengthen the integration of self-developed models and Agent capabilities, but overall it still maintains a product strategy of multi-model compatibility.
However, after the Cybersecurity Threat and Vulnerability Information Sharing Platform of the Ministry of Industry and Information Technology released the risk warning of "Claude Code has potential security backdoor hidden dangers" in early July, Claude Code has become a "red line" issue within ByteDance.
This may also be related to the fact that ByteDance regards Coding as one of its core strategies.
After the release of Seed (Doubao Large Model) 2.1 on June 23, Tan Dai commented that "we have officially joined the game in the Coding field." He added that this is a very important matter.
The official statement from ByteDance is that the Coding capability of Seed2.1 Pro can already match that of Claude Opus 4.6.
The programming evaluation benchmark Terminal Bench even believes that the Coding capability of Seed 2.1 Pro is basically on par with Claude Opus 4.7, while the comprehensive usage cost is nearly 80% lower than that of the Claude Opus 4.6 to 4.8 series of models. At the same time, Doubao Large Model 2.1 Turbo for high-frequency invocation scenarios was launched synchronously, with the price only half of that of 2.1 Pro.
It can also be seen from Tan Dai's statement that Coding is one of the important AI propositions of ByteDance. But before that, compared with the popularity of Seedance, the Coding business line did not seem to bring much revenue increment to ByteDance.
Compared with OpenAI and Anthropic, ByteDance started its investment in AI Coding later, with a different path: ByteDance first built the IDE (Integrated Development Environment), and then supplemented the model. Trae is ByteDance's earliest attempt. As the first native AI IDE in China, Trae initially accessed multiple models including DeepSeek, instead of relying entirely on self-developed Coding models.
The underlying reason may be that the Coding capability of the self-developed model was not good enough — until Seed-Code was officially released to the public in November 2025, its overall performance was still inferior to Claude Sonnet 4.5.
"The reason why ByteDance has difficulty making breakthroughs in Coding effect is the lack of data reflux," an insider pointed out, as quoted by "Intelligent Emergence". Due to the limited model capability, relevant internal businesses of ByteDance are unwilling to use Seed-Code.
However, since the beginning of this year, many ByteDance employees have felt that various business parties are increasing their support for the Seed model. A Seed employee revealed that ByteDance did not originally restrict the business side from using third-party Coding models for development, but since 2026, multiple application departments have been forced to use the Seed model.
However, in the context of ByteDance's entire AI strategy, internal usage alone is not enough. The core lies in whether there are customers willing to pay real money for usage.
An important driver of AI cloud growth?
Major domestic tech companies started late in the Coding field, and lost their inherent advantages in terms of individual users and developers. But as we all know, we are good at curve overtaking, the premise of which is to find the right curve.
At present, almost all these domestic AI Coding products are free for individual users, except Kimi Code which is purely paid. There is a phenomenon that if you use Qoder, the inference consumption will definitely run on Alibaba Cloud; similarly, the computing power consumption of CodeBuddy runs on Tencent Cloud, and Trae is based on ByteDance Volcano Engine.
In other words, if AI Coding is encapsulated as part of the organizational production infrastructure in cloud services and turned into enterprise services, new growth points can be found.
This is also the steepest growth source for Claude Code and Codex. According to media reports, OpenAI's revenue in the first quarter of 2026 is about 5.7 billion US dollars, and toB enterprise business revenue accounts for 40% of total revenue, with Codex as one of the core growth drivers. 70%-80% of Anthropic's revenue comes from enterprise customers and API business, and its gross profit margin has jumped from 38% a year ago to more than 70%.
LatePost reported that after ByteDance's senior management visited Anthropic in April this year, the company began to adjust the allocation of AI resources, shifting the focus from mass-oriented products such as Doubao to enterprise-oriented products. When the internal team believes that Seed2.1 has the capability to participate in the game in the Coding field, it will inevitably be the moment for Volcano Engine AI Cloud to exert its strength.
From the perspective of enterprises, AI Coding has almost all the ideal characteristics of toB products: unlike relatively vague indicators such as AI customer service experience improvement and how much time AI meeting minutes can save, the value of AI Coding can be verified naturally — whether the development cycle is shortened, the number of Bugs is reduced, the test coverage is improved, and the delivery time of a single requirement is compressed, all of which can be directly converted into R&D efficiency and labor cost.
The main users of AI Coding, software developers, are the most mature paying group in the AI era. They have long been accustomed to paying for tools that can improve efficiency. Therefore, when AI really enters enterprises, the R&D department is often the first group of customers willing to purchase Agent products.
Liu Weiguang, Senior Vice President of Alibaba Cloud Intelligence Group and President of Public Cloud Division, once said in an interview that in the cloud computing era, there is a long-term pain point: when counting customers' IT budgets, we cannot get the share of enterprise internal software development and labor outsourcing. Now the situation is just the opposite, these budgets can be 100% covered by AI Coding.
For this reason, almost all leading cloud vendors have put AI Coding at the forefront of their Agent strategies.
Microsoft deeply integrates GitHub Copilot into GitHub and Azure; Google incorporates Gemini Code Assist into the Google Cloud development system; Alibaba continues to strengthen the Alibaba Cloud development ecosystem around Qoder; Tencent promotes the coordinated evolution of CodeBuddy and Tencent Cloud; ByteDance also continuously strengthens the linkage between Trae, Volcano Engine and the Seed model.
"AI Coding is evolving from a 'code generation tool' to a 'software productivity platform'," Li Haoran, an IDC analyst, pointed out. "With the continuous enhancement of large model inference capability, Agent autonomous execution capability and enterprise code asset management capability, AI Coding is no longer limited to code completion, but gradually covers the complete software life cycle including requirement analysis, coding development, test verification, operation and maintenance delivery."
As a team that develops AI Coding tools, more than 90% of the code of TRAE in the past six months was completed by AI. Hong Dingkun publicly stated that the per capita requirement throughput rate has increased by 60%, reaching 1.6 times.
Therefore, Coding Agent is most likely to become the first core scenario for enterprises to continuously use Agent, which is the basis for it to become an important source of AI cloud growth. Almost all existing customers with cloud services are considered potential users of Coding.
However, for cloud vendors, the truly important significance of AI Coding is not to sell an AI development tool, but to take the lead in occupying the enterprise R&D workflow. Whoever masters the development entry will have more opportunities to carry model inference, Agent operation and more AI applications in the future.
In Liu Weiguang's view, Agent has become the biggest driving force driving the model market and even the existing cloud market.
The most crowded track for making money?
In addition to the competition in front-end applications, Coding is also the most fierce and crowded track for major tech companies to compete at the model level. One of the very important reasons is that AI Coding is one of the few application scenarios with definite Token consumption at present.
According to Gartner's analysis, as of April 2026, the annualized scale of the global enterprise-level AI programming market is estimated to reach 9.8 billion to 11 billion US dollars. It is expected that by 2028, more than 70% of enterprise software engineers will rely on AI Coding Agent to complete daily development tasks, which is expected to bring 30% to 50% productivity improvement to software engineering teams.
Since Coding Agent needs to continuously read the code base, call tools, execute tests and repeatedly modify according to feedback, the model in this process consumes not only the Tokens required to generate code, but also a large number of Tokens required to understand project context, process tool feedback and maintain long-link inference.
A 2026 study on Agentic Coding Token consumption shows that the Token consumption of complex Coding Agent tasks is significantly higher than that of traditional code Q&A and code reasoning tasks, which can reach 1000 times the magnitude of the latter. The main cost of Token consumption comes from continuous context input, rather than the final output.
In this open war at the model level, Anthropic has undoubtedly reaped the biggest dividends.
According to the Wall Street Journal, Anthropic's revenue is expected to more than double in the second quarter to 10.9 billion US dollars. According to the latest analysis from research institution SemiAnalysis, Anthropic's annualized revenue has skyrocketed from 9 billion US dollars to more than 44 billion US dollars in a few months.
Alibaba, which occupied nearly half of the domestic AI Coding market share in 2025, also benefited from this.
In the first 5 months of 2026, the Token revenue of Alibaba Cloud's MaaS business increased by 15 times, and the monthly Token revenue reached the level of hundreds of millions of yuan. Wu Yongming, CEO of Alibaba Group, predicts that the annualized recurring revenue (ARR) of Alibaba's AI model and application services will exceed 10 billion yuan in the second quarter, and exceed 30 billion yuan by the end of the year.
At the same time, domestic startups have also joined the competition one after another.
Companies including Moonshot AI, MiniMax and Zhipu have begun to strengthen their Coding capabilities, and compete for the market through more competitive API prices, free quotas or developer programs; the three major telecom operators have also successively launched large model platforms and Agent services for developers, hoping to drive cloud