Something is changing in the unit economics of software

Most of the useful pushback was not "AI is fake". It was that the old world was never as costless as the post implies, and the new world may settle into something more familiar than it sounds. Enterprise software has always carried per-customer costs through support, security, compliance, integrations, and custom work. What changes with AI is that the variable compute bill becomes large enough to show up directly in product design and packaging. People pointed out that users are already trained to accept quotas and metered plans, so AI pricing is not alien if the value is clear. The more interesting line of argument was that inference cost is only half the story. Several commenters argued the bigger shock is that LLMs lower the cost of building bespoke software, which raises the floor for weak SaaS products. If a customer or employee can generate a good-enough internal tool, script, or workflow automaton in hours, a lot of thin wrapper software loses pricing power even if inference gets cheaper. That did not translate into "all products disappear" though. Collaboration, reliability, support, compliance, and shared workflows still make packaged software valuable. Good-enough one-off generation can kill some table-stakes work, but it does not replace products that coordinate teams or sit inside real business processes. A recurring theme was that falling model costs will not automatically fix margins. Even if inference gets dramatically cheaper, products tend to spend the savings on more calls, deeper integration, longer context, and autonomous behavior rather than pocketing the difference. That leaves founders with a familiar but harsher problem: build products whose value comes from workflow ownership, proprietary data, network effects, or integration depth, because code alone is getting easier to reproduce while model usage remains an expense someone has to pay.