Coding | TAM
Self-reflection on Coding’s Real TAM
A year or two ago, I thought coding’s TAM was a simple P × Q exercise.
P: ~$200/month per developer, already a premium to GitHub / GitLab mid-tier pricing at ~$20–30/user/month
Q: 5–6M developers in the U.S., possibly lower if Big Tech mostly uses internal tools
That gets you to a rough $10–15B TAM.
But that framework was clearly wrong. Anthropic is now reportedly adding something close to that amount of ARR in a very short period of time. The issue was not the math — it was the mental model.
The first correction is that P should not be benchmarked to software seats.
If coding agents meaningfully augment developer labor, then the right anchor is not $20–30/month, or even $200/month. It is a percentage of developer salary. At $100–150K annual comp, even modest value capture pushes the TAM to $700B+.
But I think even that is still too narrow.
The bigger shift is that coding is becoming a general interface for digital work. Code runs software. Software runs companies. And many knowledge-work tasks are ultimately software operations: retrieve context, make a decision, update a system, generate an output, trigger the next workflow. So as agents get better, the market expands from developer productivity to digital labor automation.
Another way to frame it:
Global software spend is already $1.4T+
IT services add another $500B–1T
The broader digital economy is $15–16T+
Knowledge-work labor is potentially $35–50T, using 1B+ knowledge workers globally and $35–50K blended annual cost
Obviously, not all of this is addressable.
But the point is that the ceiling is not “developer seats.” The ceiling is the share of digital work that can become software-executable.
Most of what we do every day — searching, summarizing, writing, analyzing, reconciling, coordinating, updating systems — can increasingly be done by code, if the interface is easy enough and the agents are reliable enough.
So coding’s real TAM may not be “coding tools.” It may be the automation layer for the digital economy.
Upon reflection, it was clearly not the right framework to think of Coding and Excel in parallel: Everything runs on code, and therefore the best coding agent/ LLM would be able to make the best excel agent.
