Microsoft CEO says AI now writes up to 30% of company code; Meta’s Zuckerberg aims for “most” within 18 months

The chief executive of Microsoft said up to 30% of the company’s software code is now being generated by artificial-intelligence tools, at Meta’s LlamaCon.

Meta’s wager appears to be scale first, optimize later. The company plans to invest between $60-65 billion in 2025 to bolster its AI infrastructure. It recently hiked its 2025 capital-expenditure guidance to $64–72 billion, up from $60–65 billion, specifically “to support our artificial-intelligence efforts,” as it stated in its Q1 2025 earnings release. While Zuckerberg didn’t offer a point-in-time metric comparable to Microsoft’s, he projected a rapid increase in AI contribution for Meta’s own AI development: “Maybe half” of the engineering work on future Llama models would be handled by AI agents over the next year, he said.

As Zuckerberg elaborated in a recent interview with Dwarkesh Patel, this stems from a significant internal effort: “We also have a significant coding effort and are developing coding agents inside Meta… specifically, a coding and AI research agent to advance Llama research, fully integrated into our toolchain… I’d guess that sometime in the next 12 to 18 months, we’ll reach the point where most of the code going towards these efforts is written by AI… writing higher quality code than the average very good person on the team.”

A February 2023 Copilot telemetry update showed developers accepting AI suggestions for 46% of all code lines across languages and 61% in Java. This trend continues to accelerate, with GitLab’s 2024 DevSecOps survey revealing that 78% of 5,000 developers are already using or plan to use AI for code within two years. GitHub Blog, Feb 2023

On a related note, Salesforce CEO Marc Benioff told analysts in February that AI tooling has lifted engineering output by roughly 30%. Result: no new software-engineer hires in 2025. Gartner now projects that 75% of enterprise software engineers will use AI coding assistants by 2028, up from less than 10% in early 2024, according to a Gartner Report.

Evaluating the true capability of these models remains challenging. Public benchmarks can be misleading, as highlighted when Meta sparked a mini-backlash after submitting a hand-tuned “Maverick-03-26-Experimental” build that briefly topped the popular Chatbot Arena leaderboard. The company later open-sourced a different, less-performant sibling model. That prompted the benchmark maintainers (LMSYS) to tighten submission rules, as The Register noted. Zuckerberg acknowledged this difficulty, stating that Meta’s internal “North Star is the product value reported by users,” as optimizing too heavily for benchmarks like Chatbot Arena “has sometimes led us astray from product quality” and that such benchmarks “are often skewed towards specific use cases not typical for average users” and can be “easily gameable.”

Microsoft CTO Kevin Scott said on a recent podcast that “95%” of code could be AI-generated by 2030, with humans shifting to architecture and intent. Scott, however, argued that instead of replacing engineers, AI would elevate the role of human engineers to higher levels of abstraction (authorship, design). Despite this optimism, a Microsoft Research study (similar to findings from a joint Microsoft-Meta paper MSR paper) found even the best LLM agents solved less than 50% of debugging tasks on the SWE-bench Lite benchmark, underscoring remaining gaps compared to human engineers.

“Even if the code writes itself, you still need to pour concrete and string fiber. Building a gigawatt-scale data center just takes time.” — Mark Zuckerberg Dwarkesh Patel podcast

Scott also noted that AI was forcing tech companies to rethink their offerings, drawing parallels with early disruptive technologies. “So it was super confusing in the early days of the internet and I think it was super confusing in the early days of mobile where everyone had ideas about what was valuable, and very few of those ideas were actually the durable ones that proved all of the way through,” he said. Meanwhile, Amazon’s Q Developer agents migrated 30,000 Java applications in-house, a move that CEO Andy Jassy claims will save $260 million annually, according to a Amazon Blog.

This is the best time to be alive if you have an entrepreneurial spirit. —Scott