GPU-hours now have reference prices. Finished AI work does not. The SIT Index prices a constant amount of capability, whichever lab sells it, and links that price to the GPU-hours needed to produce it. That gives a common unit for buying agent work and a way to measure who earns the margin between hardware and output.
| Index | Prices | Unit | Holds quality constant |
|---|---|---|---|
| Ornn OCPI · Silicon Data | GPU rental | $/GPU-hr | n/a |
| Ornn OTPI | Tokens, one index per lab | $/M tok | No |
| Silicon Data SDLLMTK | Tokens, usage-weighted market | $/M tok | No: falls when buyers switch to cheaper models |
| SIT Index | Tokens at fixed capability, two tiers | $/M SIT | Yes |
| SIT Throughput | Tokens produced per GPU-hour | tok/s/GPU | Yes, per model and speed |
Usage-weighted indices answer "what did the market pay on average". SIT answers "what does the same work cost now", which is the number a budget, a hedge or a margin calculation needs.
Loading fixings…
Each finding names who it matters to and where the evidence is. All update with the assumptions.
Ornn and Silicon Data already publish token prices. They measure what the market paid on average. SIT measures what the same capability costs. The levels can look alike while meaning different things.
Same capability, very different prices. Pick your tier, volume and model to see what the work costs at the index and at the cheapest model that qualifies.
Dot size is index weight; your model is outlined. Solid line: index. Dashed: cheapest to deliver. Log scale.
| Model | Lab | Score | In $/M | Out $/M | Blended $/M | Tokens per SIT | $/M SIT | Weight |
|---|
Score: Artificial Analysis Intelligence Index v5. e: input price estimated at one fifth of output. Tokens per SIT: how many of that model's tokens one SIT is worth.
Revenue per GPU-hour (token price × tokens produced) minus GPU rent. Positive means renting GPUs to sell tokens pays. The other leg, owner margin, is rent minus the cost of owning the GPU.
GPU prices higher and token prices lower means the crunch is compressing.Don Wilson, DRW. Gavin Baker replied that the spread is (token price × token volume) − rent, and Wilson agreed. This section uses Baker's form.
Four lanes: each tier sold at its index price and at its cheapest-to-deliver price. The cheapest price is where competition sets the floor, so it is the lane that shows whether the squeeze is real.
Change in crunch, $ per GPU-hour, from one move at a time. Use it to see which risk to hedge or lend against.
The conversion rate between GPU markets and token markets, like a heat rate in power. It depends on the model, the GPU and how fast each user is served, so a SIT unit has to name a speed class.
Serving each user faster means smaller batches and fewer tokens per GPU: Kimi K3 on a B200 produces about 83% less at 150 tokens/s per user than at 30. Fast and slow tokens are different goods, like peak and off-peak power.
SemiAnalysis InferenceX, late October 2026. Agentic-coding traces for Kimi K3 and MiniMax M3; the older 8K-in/1K-out test for gpt-oss-120b. Counts input and output tokens together. Owner cost is SemiAnalysis's hyperscaler cost of ownership per GPU-hour. Kimi K3 stands in for the frontier because closed labs don't publish throughput.
One unit, two settlement routes. Builders take delivery as tokens; traders settle in cash against the same daily fixing. Because both use one number, arbitrage keeps them aligned, as in power markets.
Every daily fixing, with three scenarios projected from the latest one using Xing's price model. Overlay an existing index to see where SIT and the market average part ways.
One row per day. The note says what changed. Raw data: /api/fixings, or /api/fixings?date=YYYY-MM-DD for one day with every constituent.
| Date | SITB | SITF | SITC | SITB cheapest | SITF cheapest | SDLLMTK | Note |
|---|
Every number above depends on these. The defaults model an agent workload on rented hardware with measured throughput.