Introduction

Watt turns curtailed renewable energy into compute capacity you can own.

The energy that never exists

A solar farm on a clear, mild afternoon can produce more than the grid can take. Transmission is finite, demand is elsewhere, and the physics do not negotiate. When that happens the grid operator sends a curtailment instruction. Output is dialled down. In many markets the operator is compensated for the shortfall, paid, in effect, to stop generating.

That energy is not stored, sold or wasted in the usual sense. It is never produced at all. There is no inventory, no spill, nothing to recover later. It is a quantity that exists only as a counterfactual on a settlement statement.

Curtailment is not a defect of renewables. It is the predictable consequence of building generation faster than transmission, and it grows with every megawatt of new solar and wind added ahead of the wires. Estimates of global curtailed volume run into the tens of terawatt-hours annually, but published figures vary widely by methodology and market, treat any single number with suspicion, including ours.

The constraint on the other side

At the same time, the limiting input for AI has moved. For several years the scarce resource was silicon. It is now increasingly interconnection: the ability to get a large, firm, reasonably priced power connection to a site, on a timeline measured in quarters rather than years. Data-centre power demand is widely projected to grow sharply through 2030 (source: Goldman Sachs Research, to verify), and grid connection queues in several markets are now measured in years.

So there are two problems facing each other. Energy that cannot reach a load. A load that cannot reach energy.

What Watt does

Watt places modular data centres behind the meter at generation sites. Behind the meter means the compute sits on the generator's side of the grid connection point, so it can consume output that the grid has refused. The surplus is absorbed at the moment of generation, the only moment it exists.

Compute is one of the few loads that suits this. It is interruptible, it does not need to be near a city, and it can be shipped to a field in a container. Moving compute to the energy is cheaper and faster than moving energy to the compute, because a container ships in weeks and a transmission line takes a decade.

What you can own

Watt makes that absorbed energy into a transferable claim.

Stake $WATT. Staking registers a claim on the network's surplus. It is a position in physical capacity, not a subscription to an emissions schedule.
Earn Energy Points. At the close of each epoch the network attests how much surplus it actually absorbed. That measured quantity is then split pro-rata across everything that was staked while the energy was being produced. Points are denominated in kWh.
Mint compute vouchers. Burn points to mint ERC1155 vouchers in real units: an H100-GPU-hour, a GB-month of storage, a CPU-hour.
Redeem or sell. Run a job and the voucher burns on delivery. Or list it on the marketplace and let someone else run the job.

The conversion between the two sides is arithmetic, not narrative. An NVIDIA H100 SXM5 draws roughly 700 W. At a power usage effectiveness of 1.2, one GPU-hour of that accelerator sits behind about 0.84 kWh of delivered energy. A voucher is that quantity of energy, already committed to a machine.

Where this stands today

Watt is not deployed. The contracts described in these docs are a design under construction, not a running system. No site, partner, capacity figure, launch date or token address is announced here, because none is settled. Where a number would normally appear and we do not have one, it is written TBD and it means undetermined.

These pages describe how the mechanism is intended to work. Read them as specification, not as a record of achievement.

The protocol targets Robinhood Chain, an Arbitrum Orbit L2, chainId 4663 (testnet 46630). The token is planned to launch as a fixed-supply ERC20.

Continue to How it works for the loop end to end.