Fleet optimizers are often described as globally optimal: no other profile assignment earns more within the constraints. For the problem they solve, that's usually true. This page is about what happens when the problem gets bigger.
1. One miner, several profiles
Pick the most profitable profile at the current power price and hashprice. Trivial.
2. A fleet under a power ceiling
Machines now compete for kilowatts, and a merit order by profit per kW across the fleet is the natural answer. With instant, free changes and small enough power steps, it's an excellent one.
3. A real contract
Contracted quantity, an excess rate above it, a credit for unused power below it, usage adders on top. Still solvable one interval at a time, but the price of the next kilowatt now depends on what the rest of the fleet draws. It's an output of the allocation, not an input.
4. Time
Changes aren't instant or free. Some profile changes restart the hashboards, and if a miner takes twelve minutes to wake, drawing power before it reaches full hashrate, most of a 15-minute cheap interval is spent below full output. European day-ahead prices have come in 15-minute intervals since October 2025.
An optimizer that considers only the current interval either chases moves and pays the transition each time, or a switching penalty holds it still and it misses the moves that would have paid. A penalty can price wear, but it can't stand in for the future: the same state at the same current price can call for opposite decisions, depending on the intervals that follow. Where those prices are published a day ahead, they're already in the input. An optimizer that considers only the current interval discards them.
5. Time and space together
If the site may add only 0.2 MW per minute, getting from 2 MW to 6 MW by the time prices fall at 14:00 means starting at 13:40 while power is still expensive, or giving up the start of the cheap window:
time price start early start late 13:40 high 2 MW 2 MW 13:45 high 3 MW 2 MW 13:50 high 4 MW 2 MW 13:55 high 5 MW 2 MW 14:00 low 6 MW 2 MW 14:05 low 6 MW 3 MW 14:10 low 6 MW 4 MW 14:15 low 6 MW 5 MW 14:20 low 6 MW 6 MW
The table shows the site total. Which is better depends on the prices on both sides, and every step of the ramp also has to fit each transformer and feeder. An optimizer that considers only the current interval, exact or not, can't even pose that trade.
Why it doesn't decompose
| Nothing carried between intervals | State carried between intervals | |
|---|---|---|
| Nothing shared between miners | Each miner, each interval, on its own | Each miner over time, on its own |
| Shared limits and contract costs | Each interval, on its own | A real site: neither decomposition applies |
For a 10,000-miner fleet at five-minute steps that is 2.88 million device-steps per day, each with its own mode choice and carried transition state. An optimum for one interval says nothing about the optimum across the sequence. The real question is which approximations keep the couplings that matter, within the deadline.
Hashman
Hashman is a Rust codebase built around this problem: an engineering implementation, not a certified optimal solver or a turnkey service. It includes:
- Scheduler. Per-device power curves, power and ramp limits at every level of the site hierarchy, contract economics, transition states carried between steps, and coarse-to-fine lookahead from 15-minute to 5-minute steps.
- On-site agent. Receives a timeline of desired states rather than commands, and reconciles each miner toward it. Adapters for stock Antminer, Braiins OS, LuxOS, Vnish and Whatsminer firmware.
- Contract language. Spot indices, calendar rules, contracted capacity, excess and unused-power prices and surcharges, written as short expressions and evaluated at every step.
- Grid signals. Real-time and scheduled regulation messages from a reserve-market aggregator enter the same timeline at higher priority and expire on their own.
- Data. Power and hashrate per mode learned from each device's own telemetry, and historical replay against Nord Pool spot and mFRR data.
Available
Hashman is offered as technology, not as a running service: source code and rights, by acquisition or exclusive licence, with a technical handover. If any of the layers above is one you've had to approximate, I'd be glad to talk it through with an engineer.
Elias Kunnas. Contact details at kunnas.com.