The cost of AI-written code is not the same everywhere

Since ChatGPT arrived, one of the clearest uses of artificial intelligence has been writing code. Every new version of these tools programs better than the last. We went from models that completed small fragments to a point where Anthropic reports that more than 80 percent of the code it merged into its production codebase in May was written by Claude, its own model.

Dear reader, this is a paradigm shift. Anthropic says that same jump multiplied by eight the volume of code each engineer ships. Several companies already run on agentic code, AI agents that build software of any scale inside the organization.

For a long time the mantra in Silicon Valley was to spend as much as possible. Some engineers were generating between 500 and 2,000 dollars a month in AI usage, and the heaviest users went far above that. The practice earned a name, tokenmaxxing, for the extreme use of the token, the smallest unit a model uses to process text. Meta even built an internal leaderboard to push its employees to consume more.

That same mantra was used to justify layoffs. Companies like Block, Jack Dorsey’s payments firm, and the web tool Tailwind cited AI directly when they announced their cuts. Pinterest did the same. Meta notified around 8,000 employees on May 20, close to 10 percent of its workforce, while Intuit cut 3,000 jobs.

Meta’s case deserves its own paragraph. Beyond the layoffs, it reassigned 7,000 employees to work on its AI tools. The company began monitoring its workers’ clicks and keystrokes to train its models, with no way to opt out. More than 1,500 signed a petition to stop it. Its chief technology officer summed it up by saying the agents now do the work and the human role is to direct and review them.

But the tide began to turn. That indiscriminate consumption is producing staggering costs. Axios reported the case of a company that supposedly spent 500 million dollars in a single month after failing to set usage limits on Claude across its employees. The figure deserves caution because it comes from a single anonymous source. What is verifiable points the same way. Microsoft canceled much of its internal Claude Code licensing and Uber burned through its entire 2026 AI budget by April.

This is no small thing. If US companies with generous software budgets already face this problem, it is worth asking what it means for the global south, where infrastructure budgets do not come close. The cost of a token is the same in any country. It comes from an international compute cost and not from a local salary.

Here is the point that interests me. A top-tier senior engineer in Silicon Valley can cost more than 300,000 dollars a year counting salary and equity. In Mexico that same profile costs between 60,000 and 90,000 dollars. If replacing that person with tokens costs 200,000 dollars a year, the math works in the United States. In Mexico it does not. The arbitrage that makes it attractive to swap people for AI disappears when the local salary is low.

That said, it is a fact that coding agents speed up software development. That is why we will see a clear gap between countries where token access is affordable relative to their budgets and those where it is not.

With that in mind, I expect that over at least the next year we will see the rise of agents built on open-weight models like Kimi or DeepSeek. They do not match the capabilities of Claude or GPT on the hardest problems, though they are already good enough to maintain a codebase. The economics favor them. DeepSeek set a permanent price near 0.43 dollars per million input tokens, a fraction of what frontier closed models charge.

I am excited about the future of agentic software. I see it as one of the great use cases companies need to start paying attention to. Getting there takes teams that specialize in deploying and running it, plus a discipline around costs that until recently no one thought was necessary.