Factor IO / Local AI savings

Put routine work on one server.
Keep premium AI for the hard parts.

Start with a complete example. See what stays local, what still goes to the API, and whether the hardware pays for itself. No sizing interview required.

Pricing this single-server example…

The setup is preconfigured. Cited prices load automatically; no demand entry is needed.

Read code locally

One tower workstation, one RTX PRO 6000 96 GB GPU, Qwen3 8B and vLLM. Local summaries support premium reasoning.

Example tasks per monthCompare over months

Assumptions, limitations and sources

AI is optional and opens on its own page. Nothing is sent to a model unless you press Send or Ask. Privacy

Start here

A worked example is already filled in. Choose a workload, then adjust the assumptions for your client.

Worked example · not a client quote

View results ↓

How much demand

How many people, how often they come back, and the horizon the comparison runs over.

People using it

Sessions each, per day

Working days / month

Compare over (months)

Busiest moment (%) Percentage of people typing at once. 5 = 1 in 20.

Reading speed (tok/s) Below ~20 an answer feels slow. This also sets how many requests one GPU can take at once.

What the work is

The mix of traffic, which is what decides how many tokens a session costs. Percentages must add to exactly 100 — a mix that doesn't is refused, never quietly rescaled.

Chat (%)

RAG (%)

Graph RAG (%)

Agentic (%)

Tokens per turn

Workload	Turns	In	Out	Cached
Chat
RAG
Graph RAG
Agentic

Which model you'd run

This is what decides how many GPUs you need. A big model, a long context or full attention can change the answer by more than 10×.

Model Pick the closest match. Every field below fills in from it and stays editable.

Size (billion params) What must sit in memory.

Active (billion) What gets read per token. Same as size unless it's a mixture-of-experts.

Context (tokens) How much history each request carries.

Weight precision Smaller = less memory, more speed.

Serving stack vLLM or SGLang if you're unsure. This only sets how much of the card's bandwidth you actually get.

Serving mode Interactive keeps every answer above your reading speed. Batch packs in more requests and accepts slower ones.

Architecture detail

KV cache precision Halving this roughly doubles how many requests fit at the same context.

Max concurrent requests Leave blank unless your runtime caps it. Memory and speed already bound this.

Attention layers The preset's own layer structure, editable. This is what decides the cache cost per token: a sliding window stops growing past its span, GDN layers never grow at all, and MLA stores one compressed latent instead of keys and values. A hybrid model keeps one block per group.

KV bytes per token override Expert shortcut. Leave blank and the layer structure above is used. Fill it in and that structure is bypassed entirely — including window-based retention, which a single flat number cannot express.

What you'd compare against

The card you'd buy, the provider you'd rent from, and the API you'd pay instead — the three columns in the result.

Accelerator (owned & self-hosted) The card you'd buy. Its memory decides what fits; its bandwidth decides how fast.

GPUs

Measured tok/s per GPU Your own benchmark. Overrides the model above.

Rent from

Rented accelerator

Utilization (%) How busy you keep a rented card. 70 = idle 30% of the time and still paying.

Price feed

API model

What you assume about money

Electricity, staff, depreciation and the licence — the assumptions the owned option is costed against, outside the tokens themselves.

Server you'd buy Sets the one-time hardware cost from the fleet size above. You buy whole servers, so a part-full node still costs a full node — the waste is shown rather than averaged away.

Price point Sources publish a range, not a price. Some publish a typical only.

Hardware, one-time (฿) Leave blank to use the derived figure. Type your own quote to override it.

Platform software subscription The AI platform you licence on top of the hardware. Vendors count these differently — per GPU, per GB of GPU memory, per seat — and the meter is what decides your bill.

Price per unit (฿) Per whatever the meter counts. Most vendors here publish no list price, so this is your quote.

Term

Running, per month (฿) Leave blank to derive the electricity bill for every GPU installed in the selected nodes. Rack/colocation, network/egress, staff, and support/maintenance are not derived; include them in an overriding value here.

Facility PUE Assumption. 1.0 is IT equipment alone; cooling and delivery losses raise it.

Electricity (฿/kWh) Assumption. Use your blended facility tariff.

Non-GPU node overhead (%) Assumption as a percentage of GPU draw for CPUs, memory, storage and fans.

Service level & routing

What happens when local capacity runs out, and what else you would be billed for.

Monthly token budget

Required p95 tok/s A speed 95% of requests must beat. Left empty, no feasibility claim is made.

Price as of The instant prices are read at, declared rather than taken from the clock, so the same inputs always give the same quote.

Routing policy Which option handles a request when both could.

% self-hosted Used only by the fixed split.

Failover share Share of traffic that spills to the fallback. 0 = none.

Failover rate × How much dearer that spilled traffic is. 2 = twice the price.

Licensing (฿/mo)

Consulting (฿/mo)

Model API, one-time (฿) Integration and migration work you'd pay once to go the API route. Charged to that option alone, so it shortens the payback rather than being invisible.

Rented GPU, one-time (฿) Onboarding, egress setup, reserved-capacity commitments.

Implementation (฿) One-time, like the hardware. Charged once to every option, unlike the two above, so it lands in each option's one-time and horizon total and cancels out of the payback difference.

Rates live only in this page — nothing is committed, published, or folded into unit prices.

Updates as you type.

AI inference cost comparison

Explore a worked example. Compare the modelled costs over your planning period, then review the assumptions before sharing.

Loading the worked example and cited prices…

Cumulative modelled cost

Waiting for the comparison…

Infrastructure, applicable platform subscriptions and entered one-time costs. Consulting and enterprise-licensing fees are excluded. A crossing marks payback, not guaranteed savings.

Capacity and workload details

Cost breakdown and evidence

Sensitivity analysis