Local or cloud AI? Count the cost per finished document
[PLACEHOLDER: re-check the token list prices and add the source link on the day this note is published. Replace the hardware line with a vendor quote and the power and upkeep lines with real figures.]
Whether local or cloud AI is cheaper depends on cost per approved document, not token price alone. This worked example labels every input and shows the volume where local hardware breaks even.
Per-seat and per-token prices don't answer the question on their own. What matters is the cost of each approved output. Here is one worked example, with every input labelled as an assumption.
The assumptions
| Assumption | Value | Basis |
|---|---|---|
| One document package | 250,000 input + 25,000 output tokens | Several drafting and checking passes over notes, templates and reference material |
| One tailored training module | 300,000 input + 40,000 output tokens | Same approach, larger source module |
| Monthly volume | 20 packages + 10 modules = 30 documents | Illustrative pilot volume |
| Cloud list price, mid-tier model | $2 input / $10 output per million tokens | A major model vendor's public price list, checked 26 Sep 2026 |
| Cloud list price, frontier model | $4 input / $20 output per million tokens | Same source |
| Local hardware | $15,000 over 36 months = $417 a month | Illustrative placeholder, not a quote |
| Power | 100 kWh at $0.40 = $40 a month | Assumption |
| Upkeep | 4 hours at $100 = $400 a month | Assumption; use your actual IT cost |
| Local fixed cost | About $857 a month | Sum of the three lines above |
The result
Metered cloud costs about $0.75 per package and $1.00 per module on the mid-tier model, or about $1.50 and $2.00 on the frontier model. At 30 documents a month that is roughly $25 a month (mid-tier) or $50 (frontier), against about $857 a month fixed for local.
Break-even is the fixed monthly cost divided by the cloud cost per document:
- about 1,030 documents a month against a mid-tier cloud model;
- about 510 against a frontier model;
- about 100 if each document runs ten times the tokens, as heavier agentic checking can.
| Volume | 15 docs a month | 30 docs a month (base) | 45 docs a month |
|---|---|---|---|
| Metered, mid-tier (monthly) | $12.50 | $25 | $37.50 |
| Metered, frontier (monthly) | $25 | $50 | $75 |
| Local, fixed cost per document | $57 | $29 | $19 |
The honest reading, both ways
Below break-even, metered cloud is cheaper on token cost, often by a wide margin. The case for the local lane at low volume is control, not token price: sensitive data stays on your network, a known hardware cost replaces an open-ended bill, vendor price changes don't reach you, and spare capacity allows heavier checking at no extra charge. Above break-even, local also wins on cost.
What usually decides the economics is people's time. Token cost per draft runs from cents to low dollars; an hour of a senior reviewer's time is worth far more. So the first phase of an engagement measures reviewer minutes per draft, the share of drafts that need only light edits, drafts per hour on the candidate hardware, and cost per approved output, before anything is bought.
All figures are estimates or public list prices on the date stated. None is a quote or a guarantee. More on FoxGrade. To change the assumptions yourself, try the cost per approved document demo.
Sources
- [PLACEHOLDER: link to the model vendor price list used, as checked on the publication date]