The majority of code at leading tech companies is now AI-generated, and the shift is still accelerating
Shopify, Anthropic, Spotify, and Databricks Neon have each crossed the threshold where AI writes most of their code, and top engineers have stopped writing it themselves. The questions that remain are not about whether this is happening but about what comes next operationally.
The numbers Tobi Lütke puts on it are the ones that reframe the conversation. Over 50% of code at Shopify is now AI-generated, and the proportion, in his telling, is converting to much higher numbers. That alone would be a significant data point. What sits beside it is more striking: many of Shopify’s best engineers have not written code since December, when, as Lütke described it, everything changed.
That pattern is not unique to Shopify. Krishna Rao states that over 90% of Anthropic’s own code is written by Claude Code. Wesley Huff and Daniel Priestley both report that Spotify’s engineers have not written a single line of code since December. Jason Lemkin notes that over 90% of databases at Databricks Neon are built by agents, not by humans. These are not projections. They are current operating conditions at some of the largest and most technically sophisticated organizations in the industry.
The velocity data is, in some ways, more telling than the percentage figures. Walden Yan reports that Devin’s commit percentage on its own repositories rose from 16% in January to 80% in March. Andrew Feldman documents that token spending per engineer at Cerebras climbed from under $1,000 to $25,000 or $30,000 over eight months. Kyle Daigle notes that GitHub is now doing more activity in a single month than it did across an entire prior year. These are measures of pace, not just level, and they suggest the transition is still accelerating rather than plateauing.
8 months ago we weren't spending $1,000 an engineer on tokens and we're probably at 25 or 30,000 right now. Andrew Feldman
The shift is also widening who counts as a code contributor. Cole Murray describes teams where product managers are no longer filing issues for minor bug fixes. Instead, they prompt directly through Slack, and pull requests are generated automatically. Evan Spiegel reports that many designers at Snap now ship code, which he calls extraordinary. Cliff Weitzman describes Speechify’s approach to new products: no designer is assigned at the start, AI-generated defaults are used throughout early development, and designers are brought in later to improve the result. The boundary between technical and non-technical roles, which once required years of specialization to cross, is dissolving in practice.
The bottleneck has shifted accordingly. Aaron Levie offers a precise account of where friction now accumulates: when AI builds 80 to 90% of a feature, the thing that slows release is the security review. Code injection risks do not disappear because the code was generated by a model. They may, in some configurations, multiply. Levie’s observation points to a structural gap that organizations are only beginning to price in: the tooling and process for reviewing AI-generated code at scale does not yet match the speed at which the code is being produced.
Dan Shipper adds a related dimension from the data side. He describes a shift among data scientists whose jobs have moved from doing analysis to reviewing AI-generated work, most of which he characterizes as bad data science. The review burden, in other words, is not just a security problem. It is a quality problem that falls on whoever has enough domain knowledge to catch errors the model does not flag.
What the evidence, taken together, describes is not a gradual adoption curve but something closer to a threshold crossing. The question of whether AI would write the majority of professional code has been answered in the affirmative at multiple organizations simultaneously. The questions now open are operational rather than speculative: how review processes scale, how quality is maintained when the reviewer-to-output ratio inverts, and what the engineering role looks like when the primary skill is directing, debugging, and auditing rather than writing. Those questions do not have settled answers yet. But the organizations that are farthest into the transition are already living with them, and the gap between their operating reality and the assumptions baked into most engineering organizations is widening every month.