Compute Router
Compute Router
The cheapest tokens money can buy at the price you set. We aggregate every source of underpriced compute and see every trace that runs through it. Need privacy? That's a tier. We don't serve models. We run the market that prices them.
We don't ask why the tokens are cheap. We see every trace that runs through us. That's the moat.
Cheapest and most-private look like opposite ends. They're the same marketplace, and the buyer picks where they sit. We don't pretend the cheap tokens are private. We don't pretend the private tier is cheap. Honest about both is the whole edge.
"Pick the right model per request" is dead. Workloads are homogeneous, caching punishes mid-thread switching, a model is just universally better or worse. The only real question:
Set a $/Mtok ceiling. The router takes the best model with a provider under that price.
Refuses to overspend. Rung prices move with live demand and supply. Best fill under your limit, the way any market clears.
Privacy is a paid tier, not a marketplace-wide claim. It runs only on enclaves we control, attested B200 / TEE, end to end. We never route a private-tier request through third-party supply we can't seal. Honest scope is the point: cheap tokens are watched, private tokens are sealed, and we don't blur the two.
Cheap tokens come watched, that's the trade. The default tier runs through us and we keep the telemetry. The flow that earns the spread is the same flow that gathers the data, and the data is worth more than the spread.
Traces across the cheap tier show what's heating up before the market does. Nobody else has the view.
The signal prices the next capacity deal. Buy cheap, sell dear, the spread funds the next.
Against a price, not a model. Supply shifts, cost falls, no code changes. Leaving raises their bill.
Self-serving GLM-5.2 on 8×B200 sets a hard cost floor. Knowing the true cost of a token is what lets us price against everyone who's mispricing.
Floor math, not a promise. The take is the spread between true cost and market clearing price, protected by privacy.
Anything that runs an attested enclave joins the pool. The router treats it all as offers under a ceiling.
We don't ask why the cheapest tier is cheap. The CVM seals the payload regardless of who owns the metal, so we don't have to.
Eliza already needs this layer to make agent economics work. Build it as a market, not a cost center. First customer is ourselves.
Cheapest tokens by default. Privacy when you pay for it. The traces are ours either way.