ROI Calculator · one certified result

An energy estimate
you can check.

Built on one pre-registered result on a confidential H100: the clock setting GPU Optimizer recommends, 1185 MHz, used 29.3% less GPU energy per token, with 3.3% less throughput. The result is a signed certificate. Your fleet and prices are assumptions.

NVML · H100 · Intel TDX + GPU CC · Qwen2.5-7B · batch 8 · 8+8 repetitions · Polygon-anchored

sa-85e66e75164848d4a02f23d4ee3839f9 · SUPPORTED

Calculate my ROI → See the certificate ↓

Live ROI

Your fleet, against one certified result

The energy per token and the throughput are fixed: they come from the certificate. Everything in amber is an assumption you control. The result holds for this model and setup; yours may differ.

Inputs

Measured data stays fixed · your fleet and prices are assumptions

GPU count2,000

Share of time decoding70%

Assumptions — not certified

Standard planning figures, not measurements. Change them freely.

Electricity€0.08/kWh

Joint plan: Serial Alice + GPU Optimizer€2,500/mo

Grid CO₂0.20 kg/kWh

Without GPU Optimizer · default clocks

203.9µWh/token

Daily energy—

Daily cost—

Annual cost—

sa-85e66e75164848d4a02f23d4ee3839f9

With GPU Optimizer · 1185 MHz

144.2µWh/token

Daily energy—

Daily cost—

Annual cost—

sa-85e66e75164848d4a02f23d4ee3839f9

−29.3%

GPU energy per token, for 3.3% less throughput. GPU Optimizer finds the saving; Serial Alice measures and signs it.

Estimated annual energy savingestimate —

Joint subscriptionassumption —

CO₂ avoided, derivedgrid assumed —

Break-even electricity price —

Break-even fleet size —

Estimated net savingGPU electricity − joint subscription —

The Evidence

One certificate.
Registered before the run.

The hypothesis and thresholds were registered before the runs, then 8 repetitions of each setting were measured and the verdict was recomputed from them. Click Verify to check it.

Qwen2.5-7B-Instruct · bf16 · batch 8 · 128 new tokens · H100 · Intel TDX + GPU CC · NVML

Configuration	µWh / token	Visual	Certificate

✓ Open the certificate. Confirm the number yourself. The verify endpoint returns the signed measurement, with no login and no account.

Full disclosure

What is measured, and what is assumed

A calculator that hides its assumptions is a sales tool. This one shows them.

✓ Measured — fixed, signed, verifiable

203.9 µWh/token at default clocks, 144.2 at 1185 MHz: 8 repetitions each
Throughput 263.8 → 255.0 tokens/s (−3.3%)
Pre-registered outcome, verdict SUPPORTED — cert sa-85e66e75164848d4a02f23d4ee3839f9
Limit: hardware_measured, weakest repetition self_reported. The TEE claim was refused on every repetition, because the host's TCB is OutOfDate
Limit: GPU energy only, one model and one setup, not a fleet measurement
The test applied the setting GPU Optimizer recommends through Serial Alice's executor; GPU Optimizer's own software did not run. GPU Optimizer is a related company of Serial Alice
Verification: free, public, no login

~ Assumed — your inputs, editable above

Fleet size and share of time decoding
Electricity price (default €0.08/kWh)
Grid CO₂ intensity (default 0.20 kg/kWh)
Joint plan price (default €2,500/mo, price to be confirmed)
That your workload behaves like this one. Cooling and CPU are not included