Hacker News essay argues AI coding agents are eroding the customers AI labs depend on

An essay arguing that AI labs' push to replace software engineers is backfiring on their own customers reached the front page of Hacker News this week. The piece, published on the techbronerd Substack on October 24, contends that OpenAI and Anthropic have built a product that erodes the economics of the software companies they were counting on as paying customers, while subsidising that erosion through below-cost consumer pricing.

The toll-booth theory of software

The essay's central framing is that incumbent software vendors were never factories but toll booths. Adobe, in the author's telling, collected recurring revenue not because Photoshop was magical, but because replicating it required hundreds of well-paid engineers, years of work and heavy capital investment — barriers few companies could clear. The labs' plan, the essay argues, was to insert themselves into that cash flow: persuade the toll booths to lay off expensive engineers and hand a share of the savings to model providers instead.

The flaw, according to the author, is that the labs did not sell their coding agents exclusively to those incumbents. They sold them to everyone at consumer prices.

Clone economics

According to the essay, a hobbyist can now build a credible Photoshop competitor over a weekend, and AI tools are being used to reverse-engineer large commercial products and republish them as open source. A free clone pays nobody: when a designer drops a $60 Adobe subscription for a GitHub project, the author argues, that revenue does not migrate to OpenAI or Anthropic — it evaporates, and the hobbyist who built the clone earns nothing from it either. The difficulty that made software valuable, in this reading, is being dissolved by cheap AI, and the incumbents' greatest advantage, their overhead, has become trivial for a single person to match.

Subscriptions priced below cost

The essay then turns to unit economics. The $20 monthly coding plans are described as a marketing expense rather than a price: by the author's account, the heaviest users burn through compute worth several multiples of what they pay. A $200 premium tier works out to $2,400 a year — less, the essay notes, than a single month of a senior engineer's salary at Google.

Meanwhile the infrastructure bill keeps climbing. The essay points to plans by Amazon, Alphabet and Meta for roughly half a trillion dollars of capital spending in 2026, much of it for AI data centres, and frames every one of those dollars as a bet that someone will eventually pay full price for machine intelligence.

A pricing trap with no exit

The author's proposed alternative — locking the models down, pricing an AI engineer at $150,000 a year and selling only to large enterprises — is, in the essay's view, no longer possible. It points to Meta's release of Llama in February 2023, followed by free open-weight models from Chinese labs, as the moment artificial scarcity broke: open models are already good enough to clone software, and downloaded weights cannot be recalled. Raise prices and users switch to free models running on cheap rented servers; keep prices low and the labs continue losing money while their product dismantles the customer base that was supposed to fund them.

The essay supports this with historical analogies: Spanish silver from the Americas devaluing Spain's own windfall, lab-grown diamonds ending De Beers' century of supply control, and the parable of mining a golden asteroid only to crash the price of gold. It closes by arguing that OpenAI and Anthropic are structurally exposed in ways Nvidia, which sells the underlying hardware, and Meta, which has advertising revenue to fall back on, are not.

Why it matters

The essay is an argument rather than a report, and its claims about per-user losses and clone quality reflect the author's assessment. But it frames a question the AI industry has not convincingly answered: who, exactly, is the customer? If cheap coding agents compress software vendors' margins faster than those vendors' budgets migrate to AI subscriptions, the labs' addressable market shrinks as their product succeeds. That dynamic bears directly on the sustainability of hundreds of billions of dollars in planned data-centre spending, on enterprise software valuations, and on how much pricing power closed-model providers can retain against free open-weight alternatives — a debate investors, engineers and executives are increasingly having in the open.