Capability: Is the AI Bubble Real? What Builders Should Watch Instead
A Hacker News post says it isn't convinced the AI bubble exists. For builders, the better question is what breaks in your stack if valuations fall.
Is the AI bubble real? If you ship software, that is the wrong question. The right one is what breaks in your stack if it pops.
A post titled “I’m not convinced the AI bubble exists” is climbing Hacker News. It is the latest round of an argument that keeps coming back. The rounds are getting shorter. The answer matters less to you each time.
Here is the tempo.
November 2022: OpenAI launches ChatGPT, and the public conversation is about what the product can do.
March 2023: OpenAI releases GPT- and the conversation turns to whether developers can build real products on top of it.
January 2025: DeepSeek releases its R model, and Nvidia’s stock sells off sharply as investors ask whether the spending behind all this was too much.
Read those in order and the pattern is not “hype, then crash.” Capability shipped first. The valuation argument arrived later and kept reversing. Each shock hit prices, and each shock left builders with better models, cheaper access, or both.
That is the mechanism. A bubble is a claim about what investors paid for expected future profits. It is not a claim about whether the model you call from your code works. Those two can come apart. Fiber laid during the dot-com boom stayed useful after the companies that paid for it failed. A correction in AI valuations would hurt the people who own the stock. It would not make your API call return worse answers.
The real exposure is narrower and it is yours. You are at risk if you depend on one vendor’s pricing, one vendor’s model behaviour, or one vendor’s survival. That is a concentration problem. It is not a macro problem.
My position: stop trying to call the top. Build so that you don’t have to.
- Put a thin abstraction between your product and any single model provider.
- Keep one tested fallback. A small model with guardrails can be enough for many agentic tasks, and the Forge write-up shows how far that goes.
- Judge demand by revenue, not by valuations. The Solo Founder Revenue Atlas maps small AI-native companies that are actually earning.
The falsifiable part: by December 2026, every frontier lab that offers a public API today will still offer one. If the bubble talk is right about the market, it will still be wrong about the tools. If a major lab pulls its API entirely before then, I am wrong and you should have built the fallback sooner.
I cannot tell you whether the market is overheated. The pace of these arguments suggests the market is not the part you control.
Want the next shift as it lands? Subscribe and the field notes come to your inbox.