TOKENHOUSE
INFERENCE EXCHANGE
FUTURE INFERENCEMADE TRADABLE.
AI is the new cost line.Everyone wants it lower.
Spending on generative AI models will more than double in 2026, to $28bn.1
1Source: Gartner, worldwide AI spending forecast, September 2026.
Most users pay real-time rates.Much of the work can wait.
Only some workloads need real-time answers. However, most are billed as if they do. This flexibility has value.
The cheapest tokencomes from a full GPU.
A GPU costs the same whether it is busy or idle. Fill it, and every token gets cheaper.
Output token cost · $ per million tokens · log scale
- Kimi K3
- Qwen3.8 27B
- DeepSeek V4.1 Flash
- GLM 5.3 Flash
Output token cost · $ per million tokens · log scale
- Kimi K3
- Qwen3.8 27B
- DeepSeek V4.1 Flash
- GLM 5.3 Flash
Sell tomorrow'sinference today.
- 01Booking ahead unlocks the cheapest unit cost.
- 02GPUs differ. The output doesn't.
- 03Guaranteed revenue for suppliers.
The Inference Forward
Deliver Q tokens of model M during window W at price P.
- Model
- GLM-5.3-Flash
- Quantity
- 50M tokens
- Submit by
- 00:00 UTC
- Delivered by
- 00:00 UTC, next day
- Price
- $0.20 per M tokens
- Model
- GLM-5.3-Flash
- Quantity
- 5M tokens
- Window
- 14:00–15:00 UTC
- Rate
- Up to 100k tokens a minute
- Turnaround
- 30 seconds per request
- Price
- $0.40 per M tokens
The Inference Forward
Deliver Q tokens of model M during window W at price P.
- Model
- GLM-5.3-Flash
- Quantity
- 50M tokens
- Submit by
- 00:00 UTC
- Delivered by
- 00:00 UTC, next day
- Price
- $0.20 per M tokens
- Model
- GLM-5.3-Flash
- Quantity
- 5M tokens
- Window
- 14:00–15:00 UTC
- Rate
- Up to 100k tokens a minute
- Turnaround
- 30 seconds per request
- Price
- $0.40 per M tokens
One building block.For all of it.
Set the window and delivery type to fit any job. Build on it the instruments compute markets are missing: an index, futures and options.
The Inference Forward= one forward
Deliver Q tokens of model M during window W at price P.
The Inference Forward= one forward
Deliver Q tokens of model M during window W at price P.
InstrumentUse case
- One forward=An overnight Batch job
- One every hour, 9 to 5=An On-Demand peak-hours deal
- One a day for thirty days=A month of reserved capacity
StructureDerivatives
- Settled in cash, not tokens=A future
- Rolled forward, never expires=A perpetual
- A premium for the right, not the duty=An option
- The average of every trade=The Inference Rate
The Inference Forward= one forward
Deliver Q tokens of model M during window W at price P.
InstrumentUse case
- One forward=An overnight Batch job
- One every hour, 9 to 5=An On-Demand peak-hours deal
- One a day for thirty days=A month of reserved capacity
StructureDerivatives
- Settled in cash, not tokens=A future
- Rolled forward, never expires=A perpetual
- A premium for the right, not the duty=An option
- The average of every trade=The Inference Rate
Before the WTI index,physical delivery needed to be solved.
Cushing, Oklahoma · 35°59′N 96°46′W
Before the WTI index,physical delivery needed to be solved.
Cushing, OK, solved this for oil.Token House is solving this for inference.
Cushing, Oklahoma, is where America's crude pipelines meet. Because barrels change hands there, WTI became a benchmark. Token House is where inference capacity is delivered, so its price can form and be traded.
Cushing, Oklahoma · 35°59′N 96°46′W
Any provider.One exchange.
Token House is where inference capacity clears. Any provider can connect and start supplying inference. Any buyer can budget and trade excess workload capacity.
Providers · each sells in its own unit
- Inference marketplacesOpen-source models: Kimi, GLM, DeepSeek$/M tokens, per provider
- Frontier modelsModel APIs from the frontier labs$/M tokens, list price
- GPU clustersH100, RTX PRO and more$/GPU-hour
- DeAI networksDecentralised AI computeLeases and emissions
- Idle capacitySpare overnight hoursSpare hours
Token House
- 01StandardisedEvery source is benchmarked and written in one unit: tokens of a model, in a window.
- 02MatchedBids and asks cleared on one book.
- 03DispatchedWorkloads are routed to providers and delivery is verified.
Buyers · one unit
- AppsBackground and overnight jobs
- AgentsLong-running tasks
- EnterprisesBatch workloads
Each square = 1 standard unit · Q tokens · model M · window W
Providers · each sells in its own unit
- Inference marketplacesOpen-source models: Kimi, GLM, DeepSeek$/M tokens, per provider
- Frontier modelsModel APIs from the frontier labs$/M tokens, list price
- GPU clustersH100, RTX PRO and more$/GPU-hour
- DeAI networksDecentralised AI computeLeases and emissions
- Idle capacitySpare overnight hoursSpare hours
Token House
- 01StandardisedEvery source is benchmarked and written in one unit: tokens of a model, in a window.
- 02MatchedBids and asks cleared on one book.
- 03DispatchedWorkloads are routed to providers and delivery is verified.
Buyers · one unit
- AppsBackground and overnight jobs
- AgentsLong-running tasks
- EnterprisesBatch workloads
Each square = 1 standard unit · Q tokens · model M · window W
Businesses hedge oil,not drill hours.
Today's compute markets trade GPU-hours: an input whose value moves with every new chip and update, not the output buyers pay for.
The empty quadrant
- Standard terms
- One contract for every seller, so any provider can fill any order.
- Priced per token
- Priced in the unit buyers actually consume. Not a bet on hardware.
- Fungible
- Fragmented supply and demand meet on one book, not in one-off deals.
- Financial layer: indices, clearing, credit
- Venues and marketplaces
- Compute Desk × Nodal
- Silicon Data
- Liquid Compute
- Ornn
Token House
Standard contract for inference
- SF Compute
- Compute Exchange
- Stoa
- Touchmark
- Financial layer: indices, clearing, credit
- Venues and marketplaces
- Compute Desk × Nodal
- Silicon Data
- Liquid Compute
- Ornn
Token House
Standard contract for inference
- SF Compute
- Compute Exchange
- Stoa
- Touchmark