Density
Saw a few more people getting into the token reseller business this week, the relay/proxy kind. Even Fu Sheng and Justin Sun are doing it now. The setup is to wholesale tokens from OpenAI or Anthropic and resell them with a markup to whoever needs API access. The more volume that flows through, the more the operator makes.
It made me think for a second. Most AI apps people are shipping right now are also basically reselling tokens. The app wraps a prompt around a model call and bills the user for it. Margin is whatever can be squeezed between input cost and what users are willing to pay. So in some sense the whole AI app layer is one big reseller market, and the question is just who resells more.
If that is all there is, then it comes down to volume and bundling, same as any commodity. Not very interesting.
A few years ago I started and invested in Sineaptic (https://sineaptic.com) as a hobby. Guys from Sineaptic are about to launch a new ribbon driver headphone: SE-2. One version made for audiophile amp setups (we finally have it, audiophiles) and one wireless with bluetooth. The thing about ribbon drivers is they use a much stronger magnetic field to drive the ribbon, so the sound behaves more like a string instrument than what a traditional cone driver puts out. The result is cleaner highs and detail that a regular dynamic driver tends to smear over. On top of that, the SE-2 uses a patented multi-ribbon array driver, so multiple ribbons stack up in the same physical space and push the density of the sound even higher.
That logic carries over to what I have been thinking about tokens. What makes a token dense is the work compressed behind it together with the range it has to cover. A wrapper reply is just the model thinking out loud once. A dense reply sits on top of agents that went off and pulled sources, filtered them and cross-checked the results, then collapsed everything into a short answer that also has to span more meaning per word than the wrapper version. Same as the SE-2’s ribbon array reaching both finer detail and wider frequency response than a single cone driver could in the same space. The word count on the page can look identical, but the token is heavier behind it and broader across.
Reselling tokens is a fine business and people made real money selling minutes and gigabytes too. The thing I want to build sits on the other side though. Higher density per token, the way a multi-ribbon driver gets more out of the same cubic centimeter of air.
Funny that a headphone launch ended up clarifying a software question for me.