Over the past two years, the large model industry has witnessed one of its most extraordinary developments. While numerous companies were still vying to see who could craft more lyrical poems or generate more lifelike images, Anthropic capitalized on a programming tool named Claude Code—initially perceived as “just a niche vertical market”—to achieve an annualized revenue exceeding $1 billion in a mere six months. By February 2026, this figure had soared to approximately $2.5 billion. How remarkable is this achievement? It has left nearly all large model companies, still struggling to turn a profit while aggressively acquiring users, in awe.

Around the same time, the domestic market quietly underwent a significant transformation. ByteTrae, Alibaba Qoder, Tencent CodeBuddy, Zhipu CodeGeeX, SenseTime Raccoon, Baidu Comate, Moonshot Kimi Code—the list of contenders continues to expand. Liu Weiguang, President of Alibaba Cloud Public Cloud, even proclaimed publicly: “Coding is our most critical direction; it’s pivotal for virtually everything.” How did a domain once considered a “programmer’s specialized tool” suddenly become the battleground for all large model enterprises?

01. Code: The Ideal Training Ground for AI

To comprehend this collective shift, we must first address a fundamental question: In what scenarios do large models learn most rapidly?

The answer lies in two words: immediate feedback. Code inherently possesses a comprehensive “validation” system encompassing compilation, testing, and runtime results—every code snippet generated by a model can be swiftly verified for correctness. Compilation failure? Revise. Test failure? Revise. A “hypothesis-verification-correction” loop is completed in mere seconds. In contrast, assessing the quality of copywriting is subjective, and the aesthetic appeal of an image varies based on individual taste—but whether an interface functions correctly, the compiler provides an unambiguous verdict.

Huang Tiejun, Chair of the Beijing Academy of Artificial Intelligence (BAAI), disclosed a revealing detail: During model training, Anthropic utilized 4.2 trillion code tokens, accounting for over one-third of the total, with roughly half sourced from commercial software code. From the outset, the company prioritized programming as a strategic focus—its inaugural product in 2021 was a VS Code programming assistant. Meanwhile, most domestic large model companies were benchmarking against OpenAI: “Pursue general capabilities, text-to-image, text-to-video, consumer and enterprise markets—whatever OpenAI does, we follow.” At that time, OpenAI was valued at $300 billion, and “programming seemed like just a niche vertical market.” In retrospect, that assessment may have resulted in valuation gaps amounting to hundreds of billions.

02. Yourself: The Most Reliable Indicator

If technological rationale dictates “what is feasible,” commercial rationale determines “what is worthwhile.” AI Coding’s emergence as the first “deep-water zone” for large model commercialization stems from its exceptionally concise and transparent value chain.

Claude Code’s meteoric rise exemplifies this point. In February 2025, Anthropic discreetly launched an AI agent capable of directly operating terminals, autonomously reading code, executing tests, fixing bugs, and completing tasks. In just six months, its annualized recurring revenue (ARR) surpassed $1 billion. How impressive is this pace? Consider that Cursor—another popular AI programming tool—achieved $0 to $100 million ARR without a dedicated sales team, already hailed as “one of the fastest-growing companies in history.” Claude Code elevated that trajectory to unprecedented heights.

OpenAI soon recognized the looming threat. Prior to 2025, it lacked a standalone Coding product, bundling programming capabilities into ChatGPT subscriptions and APIs. By mid-2025, Codex was introduced as an independent offering, reaching $1 billion ARR by January 2026. From an overlooked niche to a $1 billion growth engine in under a year. According to AI research platform Funda, OpenAI and Anthropic’s combined annualized revenue could approach approximately $120 billion, with AI Coding serving as a key catalyst.

Domestic market figures are equally compelling. IDC reports that China’s AI programming market reached RMB 399 million in 2025, with projections to grow to RMB 1.173 billion by the end of 2026. Another IDC measurement noted a 2025 market size of RMB 2.45 billion, up 187.3% year-over-year, with 2.8 million active users. IDC surveys also revealed that developers utilizing AI coding assistants experienced an average productivity increase of 35%, with over 20% reporting efficiency gains exceeding 50%. When productivity enhancements directly translate into commercial value, no company can afford to remain passive.

03. From “Assistant Tool” to “Productivity Cornerstone”

However, if AI Coding were merely a revenue generator, its strategic significance would be underestimated. More profoundly, programming capabilities are becoming crucial for large model companies to define the upper limits of next-generation AI agents.

Zhipu articulated a keen insight when explaining its substantial investment in Coding Agents: The latter half of the large model era isn’t about “writing code” but “getting the job done.” Early AI programming assistants merely suggested code snippets, leaving the implementation to programmers. New-generation Coding Agents aim higher—first comprehending projects, modifying multiple files, executing compilations and tests, troubleshooting errors via logs, and iterating code until completion.

This “plan-execute-check-correct” closed-loop capability is precisely the foundational skill required for general-purpose AI agents. Software engineering provides a natural “training ground”: code, documentation, tests, and version histories all reside within computers, and enterprises are accustomed to paying for R&D tools. An AI capable of independently completing tasks in the code realm is merely a “physical interface” away from doing so in the real world.

This explains why nearly all tech giants are accelerating their deployments—AI Coding transcends being a mere productivity tool. It’s a rehearsal for “whether AI can truly perform your job,” not just “whether AI can assist you in writing code.”

04. A Belated but Crucial Lesson

Of course, this collective pivot reflects a sense of urgency. Anthropic prioritized programming from the outset, while most domestic firms took nearly three years to follow suit. In October 2024, U.S. AI coding startup Bolt.new launched, achieving $4 million in annualized revenue in just four weeks—“maxing out Anthropic’s GPUs.” Around the same time, a Chinese tech giant’s AI programming team was “relatively idle, with no need for overtime.”

The data disparities are even more pronounced. On developer platform Vercel, Anthropic accounted for 24.9% of token consumption and 71.8% of spending. As of August 6, 2026, no Chinese tech giant had disclosed revenue from coding products. Yet the competition is fierce. IDC data shows Alibaba’s Qoder commands 47.6% of China’s market share, followed by Zhipu CodeGeeX (11.5%), SenseTime Raccoon (10.5%), Tencent CodeBuddy (6.9%), and Baidu Comate (6.0%). Alibaba Qoder now serves over 5 million global users, including FAW Group and CITIC Securities.

On the capital front, the enthusiasm is undeniable. Anysphere, the creator of Cursor, raised $2.3 billion in November 2025 at a valuation of $29.3 billion—nearly triple its Series C valuation just six months prior. Cognition, the parent company of autonomous AI software engineer Devin, secured over $1 billion in May 2026 at a valuation of $26 billion, up from $10.2 billion just eight months earlier. When capital votes with its wallet, the signal couldn’t be clearer.

The large model industry has narrowed its focus from “pursuing everything” to “concentrating on Coding” within a year. The rationale is straightforward: In a tech wave brimming with uncertainty, whoever first identifies a “definitely profitable” scenario will survive and thrive. Today, that scenario is called Coding. Whether it evolves into Research, Work, or something else tomorrow is another battle. But for now, all large model companies comprehend one fundamental truth—first, master coding. Then, contemplate changing the world.

- End -