On AI Coding Part 1: Death of the Free Lunch
This is the first in a series on AI coding: how I use it, where I think it’s headed, and what worries me. I’m starting with the negatives. Specifically, the economics that everyone seems eager to ignore.
Setting the Stage
The AI coding market, as currently advertised, makes no economic sense. My thinking here owes a heavy debt to Ed Zitron, who has done the definitive reporting on the industry’s financial chicanery. I want to lead with that up front so that readers of this blog can go see a much more informed take than mine.
Zitron is bearish on AI coding capabilities; I’m not. I believe AI-assisted development is a legitimate tool. But that doesn’t mean guaranteed survival. Getting here required burning gigantic piles of venture capital to acquire users with no clear path to profitability. Between the shifting political climate and an increasingly strained energy market, these companies are about to face a stress test that I expect will topple more than a few of them.
On the Topic of Free Lunches
AI as a technology can be boiled down to a pipeline that takes vast quantities of chips and gigantic inputs of energy plus all the infrastructure required to support those two inputs and churns out sophisticated statistical predictions about textual data. Code just happens to be text you can verify automatically. It’s not surprising that AI coding emerged as one of its best use cases.
In a sense AI was exactly the technology big companies dreamed of. The giants of industry have started to reach market saturation for a lot of their products (like Amazon Web Services), and were flush with cash but at a loss for what the “next big thing” they could sell was going to be. AI is a technology that in theory scales almost purely on infrastructure and cash spend. They threw more and more money and training data at it, scraped the entire internet and processed books in an attempt to feed the beast. AI on paper is the perfect solution to the problem of having too much cash and not enough avenues to spend it.
The cash has been burned by now, the return on investment has hit diminishing returns, and there is no more organic data left to be processed. The beast has run out of food and the VC cash is drying up. This is important context to understand when we examine the topic of AI coding subscriptions like Claude Code and Gemini and ChatGPT. These tools are heavily subsidized; Claude and ChatGPT are losing about 10-20x the subscription cost for each user of plans like Claude Code Pro or Max. If you’re paying $100 dollars a month for Claude Code and you use the service even somewhat regularly, Claude is paying $1000–2000 dollars per month for your usage.
The era of the free lunch is over. Recent events have revealed the truth: Claude Opus 4.6 appears to have recently undergone a lobotomy and performs much worse than at launch. Usage limits and rate limits across the board have gone up significantly. Prices are rising, energy costs have gone up, the helium critical for production of chips has been choked off by the Iran war and the stockpile for production will run out soon. The AI players are going to have to start charging users what the tech actually costs them, and at those prices users are going to be unwilling to spend money on a tool that doesn’t always get it right the first time. The economics were already precarious for the American frontier models and the current state of the world is going to tip it over the edge.