Public Release - 2026-07-24

AI task cost is not a parameter count.

Artifact: active_compute_cost_sheet_v1. This cost sheet separates a model-size headline from the runtime resources used by a real task. Its accounting structure is real_cost = active_compute + routing + communication + activation_memory + idle_capacity.

total_parameters active_parameters model_storage peak_activation communication_wait task_throughput energy_per_task
Active Compute Cost Sheet v1 public evidence visual.

Accounting

A model-size headline is not a procurement ledger.

Total parameters describe one architectural scale. They do not state how many parameters are active for a task, how much memory peaks during execution, how long distributed workers wait, or how many accepted tasks the system completes per unit of energy and time.

Sparse activation reduces one term, not the entire bill.

Runtime

Sparse activation reduces one term, not the entire bill.

The cost sheet records real_cost = active_compute + routing + communication + activation_memory + idle_capacity. Routing and sparse experts can reduce active compute while communication, memory, replication, retry, and idle-capacity costs remain material at the task level.

A 35B-total, roughly 3B-active case is an example, not a universal price.

Boundary

A 35B-total, roughly 3B-active case is an example, not a universal price.

The public example is bounded to one mixture-of-experts model and a specific software and hardware environment. This sheet does not guarantee another model's performance or cost, predict electricity prices, or replace workload-specific measurement of throughput, retries, memory, and energy per task.

1

total_parameters

The full parameter count represented by the model architecture and stored weights.

2

active_parameters

The subset of parameters activated for the measured token, request, or task path.

3

model_storage

The storage footprint for weights, formats, replicas, and required runtime copies.

4

peak_activation

The maximum activation-memory demand observed under the stated workload and batch.

5

communication_wait

The measured delay from routing, synchronization, transfer, and distributed execution.

6

task_throughput

Completed tasks per unit time under a declared workload, quality threshold, and retry policy.

7

energy_per_task

Energy consumed per accepted task result, including retries and relevant idle capacity.