3.1× price spread in one day
On 14 October 2026 the German day-ahead electricity price ranged from €97/MWh at 13:00 to €296/MWh at 19:00. Cooling overhead rises with heat and load on top of that.
Pricing optimization · early access
Know what each GPU-hour costs you, and price above it.
Cyclops combines usage from your schedulers with hourly grid prices, cooling, staff, vendor contracts and hardware depreciation. You see the cost to serve of every GPU-hour, token and contract, and you get price floors and tariffs that protect your margin.
Cost to serve$2.51
- 1-year contract, Mar 2026
- $2.35
- On-demand median, Oct 2026
- $4.20
- Hardware share of cost
- 49 %
- Energy and cooling share
- 11 %
Illustrative: 70 % utilisation, 5-year depreciation
Why margins are hard to see
GPU prices change by the year. Power prices change by the hour.
Most operators set GPU and token prices from market rates and a rough cost estimate. The real cost of each hour depends on power prices, cooling load, utilisation and contracts that sit in different systems.
GPU rental prices move fast
H100 on-demand prices fell from about $8–10 an hour in 2024 to $2.85–3.50 in mid-2025, then rose to a median of $4.20 by October 2026. Contract prices moved too.
Costs sit in separate systems
Leases, support contracts, payroll and depreciation live in ERP systems and spreadsheets, apart from usage data. Few operators can say what a single GPU-hour costs them.
In our example model, a typical 1-year H100 contract at $2.35 per GPU-hour sits below the full cost to serve of $2.51. Without hourly cost data, contracts like this are hard to spot.
Sources: Bundesnetzagentur | SMARD.de via Energy-Charts (CC BY 4.0); IntuitionLabs and SemiAnalysis (contract prices); Uvik GPU price index (on-demand median).
What Cyclops calculates
Cost, price and margin for every unit you sell
Cyclops already meters every GPU-hour, job and token it bills. Pricing optimization adds the cost side, hour by hour, in the same ledger.
Cost to serve
Hourly cost per GPU-hour, token, job and tenant, built from energy, cooling, facility, staff, vendor and hardware costs.
Price floors per product
A minimum price for each GPU type, token tier or service, set to the margin you choose, so no unit is sold below cost.
Time-of-use tariffs
Cheaper off-peak GPU-hours and tokens when power is cheap, and peak prices that cover peak costs.
Margin per contract
Profit by tenant, contract, project and model, from the same numbers that produce the invoice.
Shift flexible jobs
Recommendations to move batch and training jobs to hours with lower power prices.
What-if scenarios
Test a power purchase agreement, a hardware refresh, a new tariff or a staffing change before you commit.
How it works
From meters and contracts to a price you can defend
Cyclops reads the systems you already run. Nothing changes in your schedulers or your billing process.
- Connect. Read usage, energy, price and cost data from the sources listed here.
- Allocate. Assign every cost to the hour, the hardware and the tenant that caused it.
- Price. Calculate cost to serve, price floors and tariffs at the margin you set.
- Bill and review. Use the prices in Cyclops billing and track margin per contract every month.
Inputs Cyclops reads
- Usage from Slurm, Kubernetes, OpenStack and LLM gateways
- Energy and cooling meters through BMS, DCIM, Modbus and Redfish
- Day-ahead and intraday power prices, and your PPA terms
- Vendor contracts: leases, support, licences and network
- Staff costs per hall and shift
- Asset register: hardware cost and depreciation
Example
What one H100-class GPU-hour costs
An illustrative model with typical assumptions. In a real deployment, every input comes from your own meters, contracts and ledgers.
| Cost component | US$ per GPU-hour | Assumption |
|---|---|---|
| Hardware depreciation | 1.22 | US$300k 8-GPU server, 5 years, 70 % utilisation |
| Facility and colocation | 0.51 | US$200 per kW-month, 1.3 kW per GPU |
| Vendors and software | 0.43 | 7 % of hardware cost a year |
| IT energy | 0.21 | 1.0 kW at €0.18/kWh, the average for the week of 13–19 Oct 2026 |
| Staff | 0.07 | 2 FTE per MW at US$160k a year |
| Cooling and power overhead | 0.07 | PUE 1.35 |
| Cost to serve | 2.51 | Price floor for a 20 % margin: $2.98 to $3.36 over 14 Oct 2026 |
€1 = US$1.15. Price floor = (fixed costs + energy cost for that hour) ÷ 0.8. About 12 % of the cost to serve moves with the hourly power price; the rest is fixed cost, which is why utilisation and contract terms matter as much as energy.
One platform
Bill it, report it, price it
Pricing optimization runs on the same ledger as Cyclops billing and ESG reporting. Usage, energy and cost are counted once and used three times.
Datacenter monetization
Meter, rate and invoice every GPU-hour, job and token, including reseller settlement.
ESG reporting
PUE, WUE, EED and CSRD data, with energy attributed to each tenant.
Pricing optimization
Cost to serve, price floors and tariffs for every unit you sell. You are here.
Early access
Help us build pricing optimization on real cost data
Pricing optimization is in development. We are working with a small number of GPU cloud, HPC and colocation operators. Early-access partners get a cost-to-serve review of one site and a say in the roadmap.