Thesis · why inference needs a tape

Same model,
a dozen sellers,
and no tape.

The same open-weights model is sold by a dozen providers at prices ~10× apart and latencies that vary several-fold. Corporate bonds traded the same way until a tape made the fills public. This page is the full mapping, the map of where the venues sit, and where the analogy breaks.

Thesis

Why us

The same model (say Llama-3.3-70B) is served by 15+ providers at up to ~10× price dispersion and 5–20× latency variance. That's a commodity market structure. We think it should be priced and executed like one. The people who execute in a commodity market best are the ones who've spent years pricing, executing, and measuring latency inside one. The live price table is on the overview →

How inference trades: read it as a corporate-bond market

Bonds are the closest market for how inference is transacted: dealer-quoted, heterogeneous, priced by asking several dealers, and until TRACE in 2002, with no public record of what actually traded. Spreads were wide because nobody could see them. For what a token is, electricity is the better guide: produced on demand, consumed on delivery, impossible to store.

Corporate bondsInference
The instrument, what you are buying
Issuer, the companyThe model and its weights
The issue, one specific CUSIPThe exact served configuration you buy from one host
No bond equivalent, a commodity grade, like sulfur in crude or protein in wheatQuantization (fp16 / fp8 / int4), the same artifact at a measurably different spec, not a discount
Block size, how much clears at onceContext window
Settlement convention (T+1 vs T+2)Service tier, real-time vs batch, contracted up front
The market, who quotes, and how you buy
Dealer / market makerThe inference provider, quotes continuously, holds inventory, earns a spread
Dealer inventoryGPU capacity
Asking several dealers for a price (RFQ)What a router does, every request
On-the-run vs off-the-runNewest model vs last generation, same curve, ~60× faster
The fill, what you actually got, and only measurement shows it
Default riskError rate, the fill that never arrives
Time to fill, slippageRealized time to first token, and its variance
Ratings agenciesModel evals, they grade the issuer, never the fill
TRACE, the post-trade tapeWhat we're building: a tape for inference

Inference has ratings agencies for the issuer and no tape for the fill.

The dividing line is the point: everything above “the fill” is contracted before you buy, and you can read it off a price page. Everything below it only exists after the request runs, and nobody publishes it. That is the whole argument for a tape. Where bonds run out, and it matters: a bond's deliverable is a fixed cash flow, so only the odds of receiving it vary, credit ratings grade the issuer's ability to pay, never the quality of what arrives. A token's deliverable itself varies. That's a commodity-grade problem, and an unsolved one: CME publishes exactly what “No. 2 Yellow Corn” means before anyone can trade it, while three of five providers quoting the same price for a model don't disclose the grade they ship. And a bond can be resold where a token cannot, which removes one specific mechanism: nobody can buy cheap tokens and sell them dear, so there is no arbitrageur forcing quotes together. (Not permanently: a contract on future tokens is transferable even though the token isn’t. Electricity forwards trade for exactly that reason, but writing one requires a defined deliverable first.) What remains is the other convergence mechanism (buyers switching venue) and that one is blocked by information, not by physics: you can only move your traffic to the better venue if someone publishes which venue was better. That is the argument for a tape, in one sentence. TRACE needed neither. It added no bid side and no matching engine, published what actually traded, and spreads compressed anyway.

Kottos AI is building exchange-grade infrastructure for LLM inference: an open-source microsecond gateway today, real-time intelligence that establishes best price-performance next, and (as GPU compute financializes) price discovery for inference capacity itself. The open core is how you start; the routing intelligence is where it compounds.

Inference map

One model. A dozen providers. Spread across the country.

The same open-weights model (Llama, DeepSeek, Qwen) is served from datacenters coast to coast. Distance is latency: the Bay Area and Virginia are ~60 ms RTT apart on today's fiber, before the model emits a single token. Routing has to know the map.

US inference datacenter map

Metro hubs are approximate; provider placement is illustrative of where open-weights inference clusters, not a claim about any one provider's exact facility. Distances are great-circle; the ~60 ms RTT is a real-world fiber estimate, not a Kottos measurement (cf. AWS N. California↔N. Virginia ≈ 58 ms). The straight-line speed-of-light floor is ~26 ms. Node color is static today: live demand coloring (green→amber→red from real measured load) ships with the hosted gateway's tape.