Follow the cheapest electrons.
Once a GPU is bought, the biggest ongoing cost of running it is electricity - bought at a price that changes every hour and differs across Europe at the same second. Slipstream watches every grid, moves flexible work to wherever power is cheapest and cleanest, and prices compute to match.
A closed loop, re-run every few minutes
The same principle hyperscalers use for carbon-aware scheduling, pointed at price as well as carbon.
- Step 1
Sense the grid
Live day-ahead price and carbon intensity stream in from every site's bidding zone.
12 zones · hourly - Step 2
Rank and route
Each site is scored on €/MWh, carbon, the job's service level and the cost of moving it there.
price × carbon × SLA - Step 3
Shift, live
Flexible workloads migrate to the winning site. Latency-pinned work stays exactly where it is.
re-scored every ~5 min
Then back to step 1: the whole network is re-scored every few minutes.
Source: Google, “Carbon-Intelligent Computing” (arXiv:2106.11750, 2021) - precedent for shifting flexible work across time and location.
Pricing that tracks the grid
Once work follows cheap power, the saving is real - so it shows up in the hourly price instead of disappearing into a margin.
The pricing function
When the grid runs at its 30-day average, a Dynamic GPU costs exactly its published base. Cheaper power lowers the price, dearer power raises it - smoothly, and never by more than half.
- Grid now
- --
- of average
- Dynamic
- --
- vs base
- Bound
- ±50%
- either way
- Price adjustmentHow far a Dynamic rate moves from its base
- This hourWhere today's grid puts the price now
- Ceiling and floorThe curve flattens before ±50%
- Cheaper or dearerLeft of 1×, power costs less than usual
Today, on the Dutch grid
The shaded curve is today's real day-ahead price. The bright line is what a €2.00/hr Dynamic GPU costs each hour against it; the dashed line is the same base sold at a flat rate.
- Now
- --
- per GPU-hour
- Cheapest
- --
- next published hour
- Static
- €2.00
- flat rate
- Grid priceNL day-ahead €/MWh, on its own scale
- Dynamic GPUFollows the grid within ±50%
- Static GPUThe same base at a flat rate
- Cheapest aheadLowest Dynamic hour still to come
Under the hood
Five layers on open infrastructure. Grid signals rise from telemetry to a price; hover a layer to see what it does. This is the R&D architecture the scheduler is being built on - the pricing layer is live on this site today.
Pricing
The only layer a buyer sees: one honest hourly price. Dynamic rates track the grid within ±50% of the published base; static rates stay flat.
- Takes in
- Scheduler placement + live €/MWh
- Hands up
- €/GPU-hour on this site
Honest, not throttled.
Cheaper means cheaper. A lower price is the grid getting cheaper, never a quieter cut somewhere else. It never means:
- a smaller model
- a slower queue
- a downgraded card
Cheaper
Every cheap-power window gets captured and passed through in the hourly price.
Up to half off in the cheapest hoursTransparent
One price tied to the real cost behind it - no silent rerouting, no quiet throttling.
Same card, same queue, same modelCleaner
Curtailed wind and solar get put to work instead of being switched off and thrown away.
Surplus renewables, put to work
Check it against the live market, or see which of your jobs can follow cheap power.