Anthropic CFO says AI writes 90% of code, shifting white-collar roles. The disclosure raises margin questions for software firms.
Anthropic CFO Krishna Rao disclosed that artificial intelligence now writes 90% of the company’s code, automating a swath of white-collar tasks and pushing employee roles toward management of AI outputs. The statement, reported Tuesday, marks one of the most specific internal metrics from a leading AI lab and immediately resets the conversation about how quickly generative AI can reshape software development economics.
The 90% figure is not a vague aspiration; it is a current operational reality at Anthropic. This means that the majority of code production–from boilerplate to complex logic–is generated by AI models, with human engineers shifting to prompt engineering, code review, and architecture decisions. The implication is that a company can scale its product output without a linear increase in engineering headcount. For Anthropic, which builds large language models, this internal efficiency likely accelerates its own model development cycle, creating a flywheel where better AI leads to faster AI improvement.
The CFO noted that white-collar roles are moving toward management. This suggests that the traditional software engineer job, centered on writing lines of code, is being replaced by a role that oversees AI agents, validates outputs, and integrates systems. The demand for junior developers who primarily write code could decline, while demand for senior architects and AI-savvy managers rises. This has direct implications for tech employment, education pipelines, and wage structures across the industry.
For publicly traded software firms, the Anthropic disclosure provides a real-world benchmark. If a company can achieve 90% AI-generated code, its cost of goods sold for software development could compress dramatically. Gross margins for software companies already run high, however operating margins could expand if engineering headcount growth slows or reverses. Companies that successfully integrate AI coding tools–such as Microsoft’s GitHub Copilot, Google’s Gemini Code Assist, or Amazon’s CodeWhisperer–may report lower R&D expense ratios in coming quarters. Conversely, firms that lag in adopting these tools risk competitive disadvantage, as rivals ship features faster with leaner teams.
Anthropic itself is private, yet its usage validates the product category. The company’s own Claude model likely powers much of this code generation, demonstrating a tight loop between AI development and AI consumption. This could strengthen the narrative for AI platform companies that sell coding assistants, and for cloud providers that host the inference workloads.
The next concrete data points will come from quarterly earnings calls of major software and cloud companies. Investors should listen for management commentary on AI-driven productivity gains, headcount plans, and R&D efficiency. Companies like Microsoft, Google, and Amazon may disclose adoption metrics for their coding assistants. Any mention of internal code generation percentages similar to Anthropic’s 90% would validate the trend and could move software sector valuations. Additionally, AI labs like OpenAI or Cohere might share their own internal metrics, providing further benchmarks.
For traders, the Anthropic disclosure crystallizes a thesis that AI is not just a cost center but a margin driver for early adopters. The next decision point is whether software earnings in the current cycle begin to reflect these efficiencies, or if the benefits remain confined to AI-native firms. Watch for any divergence between companies that highlight AI-driven productivity and those that do not. The broader stock market analysis suggests that software valuations are increasingly sensitive to productivity narratives, making this a key theme for the second half.
Drafted by the AlphaScala research model and grounded in primary market data – live prices, fundamentals, SEC filings, hedge-fund holdings, and insider activity. Each story is checked against AlphaScala publishing rules before release. Educational coverage, not personalized advice.