Kilopute — a simple unit for verified GPU headroom @blockspace.nz
The real scarcity at the compute edge isn't FLOPS on a datasheet. It's headroom you can actually deliver.
cKWh — also called a Kilopute — is the unit of deliverable GPU access: verified spare capacity at a random instant, weighted by real chip capability. NeoHedge mints and moves it. It is the compute twin of dKWh.
This is not “GPU hours sold.” It is a margin / connectivity unit for capacity that was free when the network checked.
| Name | Use | Grid analogue |
|---|---|---|
| cKWh | Technical / ledger ticker — compute (or connectivity) kilowatt-hour | dKWh (delivery kWh) |
| Kilopute | Plain-language name for the same unit (say it like “kilo-pute”) | How you’d say the unit in conversation |
| NeoHedge | The network that measures headroom and issues the unit | The market / protocol layer |
GPU owners, labs, and buyers all face a version of delivery risk: will spare compute actually be there when a job, hedge, or cluster needs it — and can anyone prove it without burning the card at full load all day?
As demand for inference and training tightens local capacity, three failures show up:
You can rent cloud GPUs. You can mine. Neither creates a shared unit for deliverable headroom that maps to how dKWh treats the grid edge.
cKWh / Kilopute is a tradeable accounting unit earned when a GPU proves spare capacity in a short, unpredictable window, then returns to low-power idle.
Energy markets already learned that energy and delivery capacity are different goods. Compute markets still mostly sell energy-like “hours” and ignore headroom until something fails.
cKWh makes GPU headroom visible by putting a unit on what matters first: verified access to spare silicon when the network asks.
Once per ~30-minute epoch, each registered GPU gets a ~1-minute random window. It samples headroom, runs a GPU-bound proof (not a CPU-fakable hash), and submits. Honest cards stay low-power idle between windows.
weight_i = h_i × C_chip,i × g(W_i)
h_i = fractional headroom in the random window ∈ [0, 1]
C_chip,i = chip-class capability (the “Capacity” ceiling)
g(W_i) = score of measured GPU work rate in the window
reward_i = epoch_pool × weight_i / Σ weight_j → cKWh (Kiloputes)
A half-idle frontier card can out-earn a fully idle weak iGPU. A busy farm with near-zero headroom earns near-zero that epoch — by design.
Paid for proving the card had headroom and could deliver real GPU work into that spare when sampled.
dKWh prices deliverable grid access. cKWh / Kilopute prices deliverable GPU access. Same headroom logic; different physical layer.
Continuous full-load mining is the wrong power model for GPUs. NeoHedge prefers sparse wake + idle. A Kilopute is not “one hour of TDP burn.”
Pre-alpha proves liveness and headroom. Later phases cluster GPUs into transformers and hedge delivery paths; the unit graduates toward physics-bounded margin.
Name it cKWh on the ledger. Say Kilopute out loud. Either way: it is a simple unit for deliverable GPU headroom — so spare silicon can be measured, minted against, transferred, and eventually collateralized, the way dKWh makes the grid edge priceable.