The next horses
AI capex and the SWE replacement theory
I was listening to David McWilliams’ podcast last week, “Revenge of the Nerds”. I like McWilliams, he is one of the few economists who can hold your attention for 40 minutes. In this podcast he made the argument that software engineers are the new horses. When the steam engine arrived, the economic value of horses in coal mining collapsed. McWilliams thinks AI is doing the same thing to SWEs. His thesis is that after the last 3 years of capex and data center build up, there isn’t a lot of evidence that AI is adding to the economic pie in an incremental sense. Therefore, the only possible return on this investment is if firms can replace SWEs with AI, reducing opex because tokens are cheaper than SWE salaries.
I think going straight from “AI is getting very good at writing code” to “therefore SWEs face structural elimination” is an aggressive extrapolation. So I run the numbers and checked the economics of this thesis: can the capex be explained by substitutability of labor with AI?
In 2025, the six largest US tech companies (Amazon, Alphabet, Meta, Microsoft, Oracle, Apple) spent approximately $400B in total capital expenditure. Roughly 75% of that, around $310B, went to AI infrastructure and data centers. Amazon alone spent $131.8B, Alphabet spent $91.4B. Meta, which has no cloud business and must monetize every dollar through advertising (or firing SWEs, who knows), spent $72.2B. For 2026, the guidance points toward $500B in AI capex, with Goldman Sachs estimating cumulative hyperscaler spending at $1.15T between 2025 and 2027.
There are roughly 750,000 software engineers employed in the US tech sector, earning an average total compensation (salary, equity, bonus) of around $230,000. Fully loaded with benefits and payroll taxes, the entire US tech sector software engineer wage bill is approximately $200B per year.
The annual AI capex run rate already exceeds the total annual compensation of every software engineer these companies employ. Assuming a depreciation period of 5 years, annual opex to run the data centers of 6% of installed capex base, and a cost of capital in the 8%-10% range, the capex investment over 5 years would be $1.7T. However, the present value of permanently eliminating $200B per year in SWE compensation is $778B. In other words, firing all 750,000 of them (every single one, tomorrow, permanently) would justify only 44% of the 5 year AI infrastructure costs.
What if we count every SWE in the US? AI infra doesn’t just help tech companies replace their own engineers. It gets sold as a service to every company in the economy, theoretically allowing finance, healthcare, manufacturing, and government to eliminate their SWEs too.
If we expand the pool, there are roughly 1.8 million SWEs across the entire US economy, per BLS data, earning an average all-in compensation of around $175,000. Total annual labor cost: approximately $315B. Eliminate every one of them, every developer at every bank, hospital, and government agency in the country. We are still short, only 67% of AI capex over 5 years.
One could argue that not all AI capex is tailored to coding and functional to SWE replacement. There is foundation model training, consumer inference, advertising optimization, and any other non coding LLM use cases.
It’s hard if not impossible to disentangle that. An OpenAI large-scale usage analysis found that computer programming is ~4.2% of consumer ChatGPT messages. Let’s assume that 10% to 15% of AI capex goes to coding-specific automation. In this case, breakeven requires displacing 15% to 25% of all US SWEs within five years. Aggressive, but not impossible. But also a radically different claim than “software engineers are the new horses”.
So far the evidence about AI coding tools is that they are productivity boosters, not SWE replacements. A GitHub Copilot controlled study showed a 55.8% speed improvement on a standardized coding task. To be fair, it was done in 2023, so a geological era ago. Another trial by Microsoft and Accenture in 2024 found an improvement between 7.5% and 22% of more PRs per week. Google reports roughly 10% velocity improvement.
On top of that, productivity gains don’t automatically translate to headcount reduction. Every previous wave of software automation (compilers, IDEs, cloud computing, DevOps) expanded total engineering employment by lowering the cost of building software and therefore increasing demand for it. In a more recent article by Goldman Sachs they estimate a displacement of 6% to 7% of the US workforce, but only transitory, as new job opportunities created by AI put people to work in other capacities. The evidence from use cases where AI is driving productivity gains suggests that at most 2.5% of employment is at risk of automation today, mainly financial specialists and customer service reps, and only 0.35% for SWEs.
Steam engines didn’t make horses 10% or 20% more productive at hauling loads, they replaced a specific physical function entirely and at scale. AI, so far, does not appear to be doing the equivalent for software engineering1. McWilliams is right that the profession is changing, but the actual return must come from a combination of incremental revenue, broad productivity gains across all knowledge work, and new capabilities not yet imagined, rather than one-for-one replacement of SWEs.
Do you like this post? Of course you do. Share it on Twitter/X, LinkedIn and HackerNews
Although I have to admit that the latest release of Claude Code with Opus 4.6 is shaking this assumption.