£183.04/month
Token usage plus plan fee, less applicable usage credit.
InnoWeb VenturesPrivate LLM deployment for UK SMEs, shaped around your information, workflow and users. Compare local AI for business with cloud and hybrid options before choosing a system.
Start with a CRM email agent: 100 emails a day across 22 working days. Each email uses four model calls to classify it, extract CRM details, draft a reply and check the proposed action. A person approves outbound messages. Adjust this example to your workflow. These are editable working estimates, not live supplier prices or an InnoWeb quotation. All amounts are in pounds and exclude VAT.
The local scenario is a capable, quantised model around 30B parameters for a focused workflow—not a claim that it matches every frontier model. Compare it with large proprietary cloud models using the same calls and token budget. Parameter count alone does not determine API price, task quality or suitability.
Default proprietary rates: £3.20 input and £16.00 output per million tokens; premium frontier: £6.40 input and £32.00 output. Based on published Claude Opus 5.5 ($4/$20) and Fable 5.2 ($8/$40) standard rates, using an illustrative £0.80 per US dollar, not a live exchange rate. No caching or batch discount is assumed. Cheaper cloud models exist; these are specifically large proprietary model comparisons. Both profiles start with no monthly fee or usage credit.
The default uses a suitable workstation you already own, so its incremental cost is electricity with no token bill. Paid support is set to £0—not included or free of effort; enter an allowance if required. A new £2,500 workstation spread over 36 months is also shown separately. Electricity is £0.28/kWh. Active power is 300W and idle power 60W; the machine is powered for 8 hours on each operating day. These are planning assumptions, not measured benchmarks or service prices.
Token usage plus plan fee, less applicable usage credit.
A higher-priced frontier profile; quality must be tested on your workflow.
Selected hardware allocation, electricity and any support allowance you enter.
Local cost, two views: existing suitable hardware: £9.25/month; new hardware spread over 36 months: £78.70/month, plus £2,500.00 upfront purchase. The purchase is allocated in the monthly figure, not charged twice. Support, replacement and setup need separate budgeting if not entered.
| Cost component | Proprietary cloud | Premium frontier cloud | Local |
|---|---|---|---|
| Token usage | £183.04 | £366.08 | No token charge assumed |
| Monthly plan fee / local support allowance | £0.00 | £0.00 | £0.00 |
| Usage credit applied | −£0.00 | −£0.00 | Not applicable |
| Hardware purchase allocation | Not included | Not included | £0.00 |
| Local electricity | Not included | Not included | £9.25 |
| Estimated total / one-machine budget | £183.04 | £366.08 | £9.25 |
Local processing: 93.7 estimated hours against 176 available powered hours. Electricity: 33 kWh. Throughput is a placeholder estimate; this tool does not model concurrency, queueing or task-quality differences.
Daily requests × operating days gives monthly tasks. Tasks × model calls gives monthly calls. Cloud cost is monthly fee + max(0, input token cost + output token cost − included usage credit). Hardware allocation is purchase price ÷ months, or zero when existing hardware is selected.
Local processing hours are input tokens ÷ input speed plus output tokens ÷ output speed, divided by 3,600. Electricity combines active and remaining idle hours within the powered schedule. Idle draw is included even at zero usage; switching off the machine outside that schedule is assumed.
Excluded: VAT, exchange-rate movements, development and setup, integration, cloud application hosting, storage, OCR, retrieval services, tools, licensing charges, cooling, backup and disaster recovery unless you include them in your own allowances. Caching, batch discounts, retries and provider limits are not automatically modelled. Add repeated calls and billed reasoning tokens to the relevant inputs.
Existing-hardware mode excludes purchase cost but not electricity or support. It does not account for depreciation, replacement or the opportunity cost of using that equipment. A quantised model of around 30B parameters needs hardware and workload testing; model size alone does not establish hardware suitability or quality.
Assumptions set on 21 November 2026. The cloud figures are rounded planning estimates, not a quotation or claim about a specific model. Source: Anthropic’s public API pricing. GBP defaults use the stated illustrative conversion. Check supplier prices and your exchange rate before budgeting. No 500B parameter claim is made for proprietary models whose size is undisclosed.
Your entries stay in this browser page and are not submitted or emailed.
We can review task quality, privacy, hardware capacity and operating costs before you choose local, cloud or hybrid AI.
Discuss your LLM requirements →We agree a programme after reviewing your workflow, existing hardware, integration needs and privacy requirements.
Duration: Scoped to your requirements
A workflow owner, representative non-sensitive test tasks, expected usage and any requirements for where information can be processed.
We will align the right platform entry point, deployment sequence, and governance checkpoints.
Book a Workflow ReviewAnswer 12 questions about data, processes, team skills, governance and privacy. Get your readiness band and a practical next step, with no obligation.
Take the free AI readiness assessment →