If There's a Bubble, What Survives the Pop?

Whether current AI investment is a bubble is unknowable from here, and arguing about it is less useful than a different exercise: assuming a correction happens, which parts survive it and which don't.

Previous technology corrections have a consistent shape. The technology keeps working and keeps being adopted; the capital structure around it resets, and a lot of companies whose economics depended on continued easy funding stop existing. Separating the technology from the capital structure is most of the analysis.

What survives a correction

Capability that's already useful at current prices. If something is cost-effective today at real prices, it stays cost-effective. Coding assistance, document processing, classification and extraction, retrieval — the workloads with clear returns don't stop having them.

Companies with revenue exceeding cost. Obvious, and the discriminator. A correction is a re-pricing of future promises; businesses funded by present revenue are much less exposed.

Skills. Understanding how to build reliable systems around models doesn't depreciate when valuations do. If anything it appreciates, because the cohort chasing the hype thins out.

Infrastructure that's genuinely used. Serving, tooling, and platforms with real workloads on them persist through ownership changes.

What doesn't

Products priced below cost to buy growth. If unit economics only work with subsidy, a funding contraction ends them. ⚠️ This is the practical risk for a buyer: a tool your workflow depends on can disappear or reprice sharply, which is an argument for the portability discipline rather than for avoiding tools.

Companies whose value proposition is a thin layer over a model API. The thinner the layer, the less survives when the market stops paying for potential.

Roles created for positioning rather than for work. Hiring driven by needing an AI story rather than by needing the output.

Valuations premised on near-term transformation. Whatever the technology does, the timeline priced into a lot of current investment is aggressive, and timelines are what corrections re-price.

💡 The historical pattern worth remembering

In previous cycles, the technology that caused the excitement mostly turned out to matter roughly as much as claimed — over a longer period, and after the capital structure reset and most of the companies involved failed.

Both halves are usually true simultaneously: the enthusiasts were right about the technology and wrong about the timeline and the winners. Which suggests the useful posture is to take the capability seriously and the market structure sceptically.

✅ What to do with the uncertainty

As a builder: get to real economics sooner than feels necessary. Something that works at current prices with a defensible margin survives a re-pricing; something that requires continued cheap capital doesn't.

As a buyer: keep the portability discipline — tools behind standard protocols, eval cases and prompts in your own repository, traces in your own store, one seam for the model. Costs a day, and it's what makes a vendor disappearing an inconvenience rather than a rebuild.

As an employee: weight roles by whether the work is needed rather than by whether the title is fashionable. A position that exists because output is required survives; one that exists because a company needs an AI narrative doesn't.

As a learner: invest in the durable parts — reliability engineering, verification, systems thinking, domain knowledge. Those are useful in every scenario. Tool-specific proficiency is the depreciating asset regardless of whether there's a correction.

The takeaway

Separate the technology from the capital structure. The capability that's useful at real prices today stays useful; the companies and roles funded by expectation rather than revenue are what a correction removes. Get to real economics early if you're building, keep your tooling portable if you're buying, and invest in the skills that are valuable whether or not anything pops — because those are the same skills either way.