expectedwrong hindsight

The Jewelers Are Fine

DeepSeek just handed the application layer a margin windfall while everyone panics about Nvidia.

2 min read 407 words #ai #investing #deepseek #inference #economics
hindsight — nailed it

The jewelers are fine. Application-layer companies benefiting from cheaper inference costs was the correct read. Perplexity's fundraising timing was not an accident. The picks-and-shovels trade got complicated exactly as described.

The chips are in trouble. The data centers are in trouble. The thing that is not in trouble is every company that was paying OpenAI eighty cents out of every dollar they made.

That's the actual story here — not the Nvidia sell-off, not the hand-wringing about whether DeepSeek means we were all wasting money. The picks-and-shovels trade just got complicated, and the people who were quietly building use cases on top of expensive inference got a gift they didn't ask for.

Aravind going hard on funding right now is not an accident. Perplexity has a use case. It has users. It's been slowly bleeding out to OpenAI on inference costs, and now the story you tell investors is completely different — the commodity layer got cheaper, which means the margin structure of every application-layer company just improved without them doing anything. That's a good week to be fundraising.

Every ChatGPT wrapper on the planet just got validated in the same motion. There's a version of the last three years where those companies were permanent tributaries to OpenAI — structurally unable to escape the toll booth because they had no leverage over the underlying cost. That version is now less true. Maybe a lot less true. The refiners and the jewelers get to keep more of what they make, which means they're actual businesses now instead of OpenAI distribution subsidiaries with good UX.

The part that doesn't get discussed enough: the next-generation data centers are already mostly built. The concrete is poured. The power contracts are signed. Whatever those H100 clusters and whatever comes after them are going to do — that's happening regardless of what happens to funding in the next eighteen months. The slowdown, if it's real, bites into the generation after that. Which means the compute step-change that validates or invalidates the entire current wave of AI companies is already locked in. We just don't know what it does yet.

And if DeepSeek is right about reinforcement learning — if the actual unlock isn't raw compute but the way you use compute to teach a model to reason — then the next wave doesn't require the data centers we're panicking about anyway. It requires cleverness. Which is historically harder to fund and easier to find.

The alternative, of course, is that all of this slows down enough to not kill us. Which at this point might be the most optimistic scenario on the table.