BV Compute — why the AI economy needs the same primitives as the energy economy
Every AI-native company has the same operational problem: compute is half their P&L, the spot market is volatile, the forward market is missing, and the IP layer (model weights) has no clean licensing primitive.
In every commodity market we already know how to model — oil, gas, electricity, agricultural — the four primitives below get tradable within the first few years of any meaningful production:
- Unit of capacity (a barrel of oil, a MWh of electricity)
- Unit of access / IP (a refinery franchise, a generation license)
- Unit of consumption (a delivered gallon, a metered kWh)
- Unit of outcome (a finished refined product, a delivered ton)
In compute, the equivalents are obvious once you write them down:
| Primitive | Equivalent | Our token |
|---|---|---|
| Capacity | GPU-hour | BVGPU |
| Access / IP | Model weights | BVMDL |
| Consumption | Inference call | BVINF |
| Outcome | Compute result | BVOUT |
Why this matters
Today, an AI lab buying compute is in the same position as an electric utility before the 1990s deregulation. They commit to long-term capacity contracts (Azure / AWS reserved instances) with no secondary market. Unused capacity goes to waste; over-commitment strands cash; price-discovery happens behind one or two operators.
A BVGPU-style commodity primitive lets the market function:
- Forward curves for inference cost, helping treasurers fund R&D against known unit economics
- Resale of unused capacity without operator approval
- Hedging via BV Derivatives — both cash-settled and physical-delivery
- Composability — a model creator (BVMDL) can wrap their model into a BVINF stream with deterministic per-call revenue share
What we built
Each BV Compute token tokenizes a verified unit:
- BVGPU: one verified hour on qualifying hardware (H100 / A100
/ equivalent), attested by signed
nvidia-smisnapshots and uptime ledger entries from the provider. Redeemable for cash via BV Derivatives or for physical delivery (SSH credentials issued at redemption). - BVMDL: a license to N tokens-of-inference against a registered weight set, with per-call royalty back to the original creator.
- BVINF: one paid, TEE-attested inference call against a registered model.
- BVOUT: a delivered outcome — finished training run, batch prediction set, completed simulation — escrowed in MXD until the oracle attests the spec was met.
Settled in MXD (Q3 2026 launch), gated by AccreditedRegistry for the larger-volume instruments, and integrated with the existing BV Authorized Agents rail so an AI lab's treasury can authorize an agent to roll BVGPU forwards without giving up controls.
Status
All four tokens are on testnet. We're integrating with three qualifying compute providers ahead of mainnet. Capacity providers who want to be on the inaugural list: talk to us.