In December 2025, ByteDance's Trae team shared a number that traveled quickly through China's developer community. Ninety-two percent of ByteDance's engineers were using Trae, the company's AI-native coding tool, for development work. A month earlier, at the Volcano Engine Force conference, ByteDance's technical vice president Hong Dingkun had disclosed another figure: in the Trae team itself, more than 90 percent of the code was being written by AI.
The two numbers are often cited together as evidence of how far ahead Chinese tech companies are in AI coding adoption. They are real. They are also measuring two different things.
The 92 percent measures how many engineers use the tool. The 90 percent measures how much code the tool produces inside one team. Neither number measures what most enterprise buyers actually care about: how much AI-generated code makes it into production, how much engineering time is saved, and how much of the output survives contact with a real codebase.
The gap between those questions is where the adoption story becomes more complicated.
Two Metrics, Two Meanings
The distinction between "engineers using AI" and "code written by AI" is not semantic. It changes what the data can support.
ByteDance's 92 percent figure, confirmed in multiple Chinese media reports in December 2025, describes adoption breadth. It means that nearly every engineer in one of the world's largest technology companies has access to an AI coding tool and uses it in their daily workflow. That is a significant organizational achievement. It does not mean that 92 percent of ByteDance's code is AI-generated.
The 90 percent figure describes code volume inside the Trae team — a group of engineers who build the AI coding tool itself. That team has every incentive to use its own product aggressively, and every capability to do so. It is not a representative sample of the broader engineering organization, let alone the broader market.
The distinction matters because the two numbers are often conflated in public discussion. A company that says "92 percent of our engineers use AI" and a company that says "92 percent of our code is written by AI" are making very different claims. The first describes tool penetration. The second describes production output.
ByteDance's own data suggests the relationship between the two is not linear. In July 2026, the Trae team published an internal analysis noting that while more than 90 percent of its code was AI-generated, per-capita requirement throughput had improved by only 60 percent. The gap between 90 percent code generation and 60 percent productivity gain became a talking point in China's engineering community. As one analysis put it, the numbers suggest that "writing code is only part of the development process" — and that focusing on code generation rate alone can obscure the larger bottleneck.
The Tencent Data Point
Tencent offers a second data set that complicates the picture further.
In February 2026, Tencent disclosed that 50 percent of its new code was AI-generated. A month later, InfoQ published a detailed account of a Tencent Cloud experiment in which the CodeBuddy 2.0 upgrade was completed by four engineers in four months, with 99.9 percent of the code generated by AI. The codebase covered more than 12,000 Tencent engineers.
The 99.9 percent figure is striking. But it describes a specific, bounded project: a version upgrade of a tool, executed by a small, highly capable team, using the tool itself. It is not a company-wide average. Tencent's own company-wide figure — 50 percent of new code — is more representative of what a large engineering organization actually achieves when it deploys AI coding tools at scale.
The difference between 99.9 percent and 50 percent describes the gap between what a dedicated team can do with a tool it built, and what a large organization can do with the same tool deployed across thousands of engineers with different skill levels, different projects, and different levels of comfort with AI-assisted development.
The Market Average
The company-level numbers are higher than the market average, and the market average is higher than the depth of adoption.
According to the China Academy of Information and Communications Technology's 2026 AI4SE survey, covering 2,109 valid questionnaires, the average AI code adoption rate — the share of AI-generated code that is actually accepted and merged — rose from 26.81 percent in 2024 to 42.61 percent in 2025, a year-over-year increase of 55.17 percent. The share of enterprises in which more than 90 percent of developers use intelligent development tools rose from 5.71 percent to 27.65 percent. More than half of enterprises achieved large-scale coverage of their development teams.
Those are strong adoption numbers by any measure. They describe breadth, not depth.
A separate 2026 analysis by Simatec, based on data from more than 200 enterprises, found that while 96 percent of enterprises had adopted AI-generated code, 87.6 percent of enterprises had AI code accounting for no more than 50 percent of the code merged into their main branches. Only 2.25 percent of enterprises achieved an adoption rate above 81 percent. The report described the pattern as "high penetration, shallow depth" — nearly every enterprise is trying AI coding, but very few have made it the dominant mode of production.
The implication is that the ByteDance and Tencent numbers are not outliers because they are fabricated. They are outliers because they describe a level of organizational commitment and tool integration that most enterprises have not reached. A company where 92 percent of engineers use an AI coding tool, and where a dedicated team can achieve 90 percent code generation, is not representative of the market. It is a preview of what the market could look like if every enterprise deployed the tool with the same intensity.
The American Comparison
The American data tells a different story about where adoption is happening.
According to JetBrains' 2026 Developer Ecosystem Survey, covering more than 15,000 professional developers, 90 percent of developers use AI agents weekly, and 68 percent use them daily. Claude Code's penetration reached 39 percent globally and 47 percent in the United States — nearly half of American developers. OpenAI's Codex saw its awareness rise from 27 percent to 65 percent in a year.
Those numbers describe individual developer adoption. They do not describe organizational deployment. An American developer who uses Claude Code at work is making a personal choice about which tool to use. A ByteDance engineer who uses Trae is operating inside an organization that has deployed the tool as part of its standard development workflow.
The distinction matters for the same reason it matters in China. American developers have access to a wider range of tools, and they switch between them more frequently. The JetBrains data shows that Cursor's adoption declined from 18 percent to 12 percent globally, and from 28 percent to 16 percent in China, as Claude Code and Codex gained share. The market is fragmenting into specialized workflows rather than consolidating around a single tool.
China's market is consolidating differently. Trae, Qoder, and CodeBuddy are backed by cloud platforms — ByteDance's Volcano Engine, Alibaba Cloud, and Tencent Cloud — that bundle the coding tool with infrastructure, enterprise support, and domestic model access. The adoption numbers inside ByteDance and Tencent reflect the integration of the tool into the company's own development pipeline, not just the tool's standalone appeal.
What the Adoption Rate Actually Measures
The 92 percent figure is real. The 90 percent figure is real. The 99.9 percent experiment is real. None of them are wrong. They are measuring adoption at different layers of the stack.
At the top layer, tool access is nearly universal inside China's largest technology companies. The tools are deployed, the licenses are provisioned, and the engineers are using them.
At the middle layer, code generation volume varies enormously. A dedicated team building the tool itself can reach 90 percent. A company-wide deployment might average 50 percent. A market-wide average sits closer to 42 percent.
At the bottom layer, production impact is harder to measure and less impressive than the top-line numbers suggest. ByteDance's own data shows a 60 percent throughput gain from 90 percent code generation. The Simatec data shows that most enterprises have not crossed the threshold where AI code becomes the majority of what ships. And the Harvard working paper on software production bottlenecks found that when code volume rises without a corresponding increase in review capacity, the productivity gain can be absorbed entirely by downstream constraints.
The adoption rate is not a single number. It is a stack of numbers, each measuring a different thing. The companies that cite 92 percent are describing the top of the stack. The companies that measure 42 percent are describing the middle. The companies that ship software are living at the bottom.
What the Numbers Mean for the Market
The Chinese AI coding market is growing faster than the global market, and its largest players are deploying tools at a scale that American competitors have not matched inside a single organization. IDC projects China's AI coding market will reach 1.173 billion yuan in 2026, nearly triple the previous year, with a 40 percent compound annual growth rate through 2030.
But the adoption rate that matters for enterprise buyers is not the one inside ByteDance or Tencent. It is the one inside their own engineering organizations. That number is lower. It is rising. And it is constrained by the same bottleneck that affects every organization deploying AI coding tools: the capacity of humans to review, integrate, and trust what the AI produces.
The 92 percent figure tells you that ByteDance has deployed the tool. The 42 percent figure tells you that the market has not yet figured out how to use it deeply. Both are true. The gap between them is where the next phase of the adoption story will be written.
Sources: Tencent News "The TRAE Used by 92% of ByteDance Engineers" (December 19, 2025); Geek Park "After Handing 90% of Code to AI" (July 8, 2026); InfoQ "Handing 99.9% of Code to AI: A Tencent Cloud Dogfooding Experiment" (March 16, 2026); CAICT 2026 AI4SE Industry Survey via sgpjbg.com (August 5, 2026); Simatec 2026 AI Software Engineering Impact Report via sgpjbg.com (July 27, 2026); JetBrains 2026 Developer Ecosystem Survey via InfoQ (August 31, 2026); Cailianshe "Domestic Mystery Model Strikes: Half of Tencent's New Code Is AI-Generated" (February 10, 2026); 36Kr "Tencent Executive: Most of Our Code This Year Is AI-Generated" (June 5, 2026).
Disclaimer
The information provided in this article is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this article. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. The performance data and benchmarks discussed are based on specific research studies and may not generalize to all use cases or environments. As of the publication date, the AI landscape continues to evolve rapidly, and readers should verify current information independently.
Limitations
This analysis is based on company statements, industry surveys, and public reporting available as of the article date.
Adoption-rate figures come from surveys with different sampling methods (CAICT 2,109 questionnaires; Simatec 200+ enterprises) and are not directly comparable.
Company-level figures (ByteDance, Tencent) describe specific teams or projects, not company-wide or market-wide averages.
The 92 and 99.9 percent figures measure different things; they should not be read as a single adoption rate.
American comparison figures come from a developer survey, which measures individual use rather than organizational deployment.
Reporting standards differ across jurisdictions; figures may not be directly comparable.
This sector is changing quickly; claims can become outdated soon.
Sources
- Tencent News "The TRAE Used by 92% of ByteDance Engineers" (December 19, 2025)
- Geek Park "After Handing 90% of Code to AI" (July 8, 2026)
- InfoQ "Handing 99.9% of Code to AI: A Tencent Cloud Dogfooding Experiment" (March 16, 2026)
- CAICT 2026 AI4SE Industry Survey via sgpjbg.com (August 5, 2026)
- Simatec 2026 AI Software Engineering Impact Report via sgpjbg.com (July 27, 2026)
- JetBrains 2026 Developer Ecosystem Survey via InfoQ (August 31, 2026)
- Cailianshe "Domestic Mystery Model Strikes: Half of Tencent's New Code Is AI-Generated" (February 10, 2026)
- 36Kr "Tencent Executive: Most of Our Code This Year Is AI-Generated" (June 5, 2026).
The information provided in this article is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this article. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. The performance data and benchmarks discussed are based on specific research studies and may not generalize to all use cases or environments. As of the publication date, the AI landscape continues to evolve rapidly, and readers should verify current information independently.
Limitations: This analysis is based on company statements, industry surveys, and public reporting available as of the article date.; Adoption-rate figures come from surveys with different sampling methods (CAICT 2,109 questionnaires; Simatec 200+ enterprises) and are not directly comparable.; Company-level figures (ByteDance, Tencent) describe specific teams or projects, not company-wide or market-wide averages.; The 92 and 99.9 percent figures measure different things; they should not be read as a single adoption rate.; American comparison figures come from a developer survey, which measures individual use rather than organizational deployment.; Reporting standards differ across jurisdictions; figures may not be directly comparable.; This sector is changing quickly; claims can become outdated soon.