Bottlenecks are the seed of great startups. The founders we canonise rarely invented anything from scratch. They watched a constraint, waited for it to break, and made sure they were standing exactly where the value would land. Newton said we stand on the shoulders of giants.
The pattern is old
Steve Wozniak did not invent the microprocessor, and he certainly did not intuit memory from scratch. Two decades of business computing created the demand for those components, and economies of scale made them cheap. His genius was assembling them into the Apple II at the exact moment the constraint dissolved.
The AI cycle is running the same script
The same thing is happening right now, and it is so obvious that everyone should internalise it. GPT-2 and GPT-3 shipped as APIs, and the world mostly shrugged. The capability existed, but the interface was the bottleneck, so the smartest models on earth sat around as a toy for developers. Then someone put the same model in a chat box, and ChatGPT became the fastest-growing consumer app in history. The model did not change. The bottleneck did.
The next constraint was data. Better models needed genuine experts at scale to fine-tune for consumer and enterprise use, and someone had to find those experts and prove they were real. Who solved that? Mercor did, by using AI to assess whether an expert actually is an expert, then recruiting them to train the frontier labs' models. What is their ARR? It just passed $2bn, on one of the sharpest run rates in the industry. That is a very good business built on a bottleneck almost nobody was staring at.
Then came agents. The models alone could not support agentic usage or justify enterprise deployment, and it took coding agents to prove the loop actually worked. Notice that there is still no CFO agent or legal agent roaming the application layer. Agentic functions live inside apps behind heavy guardrails, because nobody wants a hallucinating intern with admin access. So who is crushing it right now? The scaffolding companies are. Braintrust builds evals. Browserbase lets agents use the web. LangChain handles orchestration and context, MCP gateways are being built so agents can call the right tools, and engineers are racing to solve context management and the pipelines that feed agents. Each of them spotted a bottleneck six months before it emerged, and the market has paid them handsomely for their eyesight, with Braintrust at $800m and LangChain at $1.25bn.
Cycles stack on cycles
None of this layering is new. The datacentre buildout gave us AWS and the cloud primitives. The developer layer followed, as Stripe, Shopify, and MongoDB turned payments, commerce, and data into APIs. Then Figma, Notion, and Linear matured on the rails beneath them. Each wave commoditised the inputs of the next one. On that map, the AI cycle is still deep in its datacentre and developer buildout, which should make you feel early rather than late.
So where is the bottleneck of tomorrow?
Run the thought experiment with Harvey and Legora, the flagship application companies in legal. If they win and 90 per cent of drafting is commoditised, what gets freed? Where does a law firm actually compete once the document writes itself?
The answer is everything around the work. Client service, advice, and the generation of legal demand become the differentiators. Adoption itself becomes a battleground, and firms will compete on digitally native client experiences that lawyers have never had to build before. The quality bar lifts, and new primitives emerge for discovery and the operational side of legal work altogether.
Now generalise it. When production is commoditised, the bottleneck migrates to distribution, trust, and verification. Someone has to prove that an agent's output is correct, safe, and compliant. Someone has to own the client relationship when the artefact itself is a commodity. And the category leaders still do not exist in construction, finance, HR, or recruitment, so the dental OS and the recruitment OS are sitting there waiting for someone impatient enough to build them.
The selection rule
Do not build for the bottleneck of today. That trade is crowded and the returns are priced in. Ask where an industry will compete once its current constraint dissolves, and build for the one that shows up next. The winners of every cycle answered that question six months before it became obvious, and then they let everyone else call them lucky.