How every number here is made.

01

A machine runs the model itself

A gram is someone’s laptop, desktop or rented box running 50GRAMx. It downloads a build of a model from its publisher (Hugging Face, Ollama) and runs real prompts on its own engine — Ollama, llama.cpp, vLLM, or a speech or vision runtime. Nothing is estimated from a model card.

02

It measures what a user would feel

Tokens per second of generated output, the longest prompt it answered correctly (a context ladder), how much memory the model held, and the checks it passed: tool calls, valid JSON, code, reasoning, images, languages. Each build is measured per machine; a laptop and a GPU server are never averaged.

03

It signs the result with its own key

Every gram has an Ed25519 key; its id (did:epn:…) is that public key. The gram signs the exact bytes of each result. The site publishes those bytes and the signature, so your browser can check any number without trusting us — press “verify this yourself” next to a run.

04

It benchmarks its own parts

The same gram signs benchmarks of its CPU (model, clock, cores, single- and all-core throughput), memory (size, bandwidth, latency), card (name, memory, fp16 TFLOP/s, watts) and disk (read, write, IOPS). That is the machine sheet behind every result.

05

Price is one posted list, applied to the signed machine

Every machine is metered at the same rates. A million tokens costs what the machine’s CPUs, memory, card and energy cost for the time it takes to write them at the signed speed. The breakdown is shown on every run.

06

Everything is open data

The aggregate is rebuilt from the signed results every few minutes and committed to a public repository, so every version of every number has a history.

The price list

• CPU: $0.036 per thread-hour
• Memory: $0.003 per GB-hour
• GPU: $0.36 per card-hour
• Energy: $0.1 per kWh

USD per million tokens = metered µUSD per second ÷ tokens per second. Source: epn-daemon config.DefaultPriceFloors: what metering charges every gram.

The provider standings

Points from signed work only: 1 per signed run, 3 per build where the provider holds the network’s fastest result on that kind of machine, 10 per model they were the first machine anywhere to measure. A claimed person’s machines count together. Position change is against the standings a week earlier.