TSG Waterline

Follow the token: cloud, local, and the crossover

A scenario calculator from TSG Technology, not a forecast. Every assumption is editable and every figure shows how it was calculated. No sign-in, no email.

For your AI: give it the pre-run findings and assumptions and ask it anything.

Adjust the assumptions ↓

Who is this for

Generic. AI embedded in operational workflows, ~12M raw tokens/day. Illustrative only.

Cost discipline at a normal pace.

Organisations are illustrative assumptions, not empirical claims about real organisations. The names are invented, and every figure can be changed.

Adjust

What they use it for9 workloads

AI activity growth a year

About 5× by four years

Today: 12M raw tokens a day across 9 workloads.

Deliberation (reasoning tokens)

Reasoning multiple, simple work×

Reasoning multiple, moderate work×

Reasoning multiple, complex work×

Reducible reasoning share

How much of that deliberation workflow efficiency can remove, across every workload.

Images and scanned pages

Each workload sets its own images or pages per task. They are billed as input tokens, at this rate.

Tokens per image or page

Ways to manage cost3 of 4 levers on

Workflow efficiency

Cut avoidable tokens: re-sent context, retries, answers longer than they need to be. Ramps in during stage 2 (months 6–18).Removes 40% of token demand by month 48.

Frontier substitution

Send work to cheaper models where it can go. Ramps in during stage 3 (months 18–30). Complex work keeps at least 30% on frontier, moderate 10% and simple 0%, however high this is set (the floors are under Advanced).Frontier’s share of tokens falls from 85% at month 0 to 39% by month 48.

Local substitution

Run suitable work on your own hardware. In the organisation view, the fleet’s capacity decides how much of it actually runs there. Ramps in during stage 4 (months 30–48).34% of tokens run locally by month 48.

Local can also take frontier work

Off by default: local takes only commodity work. Raise it to let local also run this share of suitable frontier work, where the saving per token is largest. Frontier work is served first when the fleet is full.

Advanced assumptions

Annual efficiency improvement

Share of the remaining efficiency headroom closed each year after stage 2 ends (month 18 onward), instead of freezing at its target.

Annual frontier optimisation

Share of the frontier share's distance above its floor given up each year after stage 3 ends (month 30 onward), instead of freezing at its target.

Frontier floor, simple work

The least frontier share this class keeps, however far substitution and continued optimisation go.

Frontier floor, moderate work

The least frontier share this class keeps, however far substitution and continued optimisation go.

Frontier floor, complex work

The least frontier share this class keeps, however far substitution and continued optimisation go.

Cloud pricescommodity US$0.3 / Mtok in, US$1.20 / Mtok out

Prices as of 2026-10-07. Illustrative placeholders, not a quote from any provider: enter the prices you pay.

What each tier charges

Frontier cloud

The leading Claude, GPT or Gemini tier. Illustrative labels only.

Default prices: the median list price of Anthropic Sonnet, OpenAI Sol, Google Gemini Pro and xAI Grok, as of 2026-10-07.

Input price (per Mtok)US$

Output price (per Mtok)US$

Caching and batch discounts

Offers a cached-input price

Cached input price (per Mtok)US$

Cached share of input

Share of can-wait work sent via batch

Batch discount

Commodity cloud

Smaller proprietary models, hosted open-weight models, low-cost providers.

Default prices: the median list price of Anthropic Haiku, OpenAI Luna, Google Gemini Flash, Qwen Flash and DeepSeek Flash, as of 2026-10-07.

Input price (per Mtok)US$

Output price (per Mtok)US$

Caching and batch discounts

Offers a cached-input price

Cached input price (per Mtok)US$

Cached share of input

Share of can-wait work sent via batch

Batch discount

How prices move

Frontier cloud

Frontier cloud: annual price decline

Frontier price floor

The price cannot fall below this share of today's price, however far the annual decline above runs.

Commodity cloud

Commodity cloud: annual price decline

Commodity price floor

The price cannot fall below this share of today's price, however far the annual decline above runs.

Advanced assumptions

Which tier serves what

Frontier share during adoption

Stage 1 sends most work to frontier models whatever its complexity; commodity gets 15%.

Once optimised, each class of work settles at its own frontier share. Frontier substitution sets how far the routing moves there.

Simple work: frontier share when optimised

Moderate work: frontier share when optimised

Complex work: frontier share when optimised

Reserved frontier capacity

Frontier volume bought ahead at a discount and paid for whether used or not. Sized on monthly volume, not peak, which flatters it.

Reserved share of frontier volume

Reserved price discount

Reservation term (months)

Own hardwareDesktop device: 2 devices, US$4,699 each, 250 tok/s, people US$300 a month

Hardware type

A DGX Spark-class desktop inference device, bought outright. Figures as of 2026-10-10, illustrative. Every figure is an assumption to replace with your own.

How many, and for how long

Fleet

Devices

Useful life (months)

Residual value

What they cost to buy and staff

Price per device, todayillustrative defaultUS$

Rent per device-hourUS$

Annual device price decline

Cost of capital (annual)

Discount rate for discounted payback (annual)

People and operations (per month)US$

What they can do

Effective tokens per second per deviceillustrative defaulttok/s

Throughput varies with the model, quantisation, input/output mix, context length, batching and concurrency, often by more than ten times. There is no universal figure: 250 tok/s is a deliberately conservative placeholder for a mixed business workload.

Annual throughput growth

Available hours per monthh

Hours a day people need answersh

Headroom cap (max usable utilisation)

Advanced assumptions

What they cost to run

Idle power (W)

Peak power (W)

Electricity price (per kWh)illustrative defaultUS$

Energy overhead multiplier (PUE)×

Annual maintenance (share of hardware cost)

Outside available hours, devices are

Hardware generations

Performance gain per hardware generation×

Value beyond costnothing attached

If running inference locally is worth something to you beyond its cost (resilience, sovereignty, speed), put a monthly value on it here. Every value starts at zero and none of them enters any financial figure.

Cloud outage resilienceUS$

Offline operationUS$

Data sovereigntyUS$

LatencyUS$

Predictable capacityUS$

Bargaining power and provider diversificationUS$

Strategic benefits: US$0 a month
Financial payback: None in 10 years
Payback counting strategic value: None in 4 years (a non-financial estimate)

Contains non-financial estimates. The strategic value is counted in every month local hardware is installed.

Timing and stagesStage boundaries, ramp shape, working days

Each optimisation lever ramps in during its own stage, then holds. Stages end at these calendar months, whatever the horizon.

End of stage 1: Adopt

End of stage 2: Optimise workflow

End of stage 3: Optimise model

Location stage ends (month)

Ramp shape

Working days per month

The answer

Local hardware does not pay for itself within ten years, with activity held flat after four years.

At four years it costs 5.81× as much as commodity cloud per token, the only tier it replaces.

At four years

Monthly cloud cost

US$550

Covers only the tokens run in the cloud

Monthly local cost

US$450

Covers only the tokens run locally

Local payback

Not within 120 months

Checked to month 120, with activity held flat after month 48. Hardware bought in month 31.

The largest part of the monthly local cost is the people who run it, spread over the tokens the fleet actually served.

Workflow optimisation removes 40% of token demand at four years. That rests on the assumed share of each workload’s tokens that is avoidable, which you can edit.

Where the tokens run, month by month

Share of tokens by where they run

Frontier cloud
Commodity cloud
Local

What moves the answer

Cloud price decline moves the answer most.

Local costs 5.8× what commodity cloud does at month 48. Throughput growth rate barely moves it. Nothing from half to double flips it. Closest: cloud price decline at ½× (4.6×).

Show

Local cheaper
Cloud cheaper
Input lowered
Input raised

Each bar scales one input from half to double and re-runs the whole scenario. The dark line is the scenario as set. Hover or tab to a dot for its figures.

Go deeper

Break-even

Read at month 48. Move the scrubber to read another month.

Compare local with

Local price ceiling: local cost per million tokens it serves

US$1.69 / Mtok

Month 48, 84% utilised

vs frontier cloud (US$1.55 / Mtok blended): cloud wins 1.09×
vs commodity cloud (US$0.291 / Mtok blended): cloud wins 5.81×

Local serves commodity-routed work, so commodity cloud is the price it competes with, and commodity cloud is 5.81× cheaper.

Against the frontier price benchmark, frontier cloud is cheaper by 1.09×.

The frontier line is a benchmark, not a substitution the model makes: local only serves commodity-routed, local-suitable work unless “Local can also take frontier work” is raised.

Required local utilisation

563% of usable capacity to match commodity cloud; achieved 84%. Above 100%, break-even is out of reach within the headroom cap.

Minimum monthly volume: 1.79B tokens, against 268M served now.

Cash payback on local hardware

Not within 120 months

Cumulative hardware and running costs against the commodity cloud cost they avoid. Checked to month 120, with activity held flat after month 48. Hardware bought in month 31.

Discounted at 8% a year: not within 120 months.

Total cost of ownership

From month 31, when the hardware is bought. A window that runs past month 48 has no figure.

Total cost of ownership
Scenario	1 year	2 years	3 years	To horizon
With local hardware	US$12,108	—	—	US$18,089
All on cloud	US$7,007	—	—	US$10,681

Hardware generation timing

Checking…

Hardware improves in steps, and nobody knows where today sits in the cycle. Assuming each generation gives 2.00× more for the money (Hardware → Advanced), buying just after one gets the same devices for less, and just before one for more. Every purchase, refreshes included, is taken as equally lucky. A bracket, not a forecast.

The chart of what moves the answer sits under the token chart.

Show the numbers

Each row scales one input and re-runs the whole scenario. 1.00× is the scenario as set. Each cell: local cost per Mtok, then how it compares with commodity cloud. Shading marks which is cheaper.

Sensitivity, at month 48
Input	0.50×	0.75×	1.00×	1.50×	2.00×
Electricity price	US$1.62 / Mtokcloud wins 5.55×	US$1.65 / Mtokcloud wins 5.68×	US$1.69 / Mtokcloud wins 5.81×	US$1.77 / Mtokcloud wins 6.07×	US$1.84 / Mtokcloud wins 6.32×
Hardware price	US$1.48 / Mtokcloud wins 5.08×	US$1.59 / Mtokcloud wins 5.45×	US$1.69 / Mtokcloud wins 5.81×	US$1.90 / Mtokcloud wins 6.53×	US$2.11 / Mtokcloud wins 7.26×
Cloud price decline	US$1.69 / Mtokcloud wins 4.61×	US$1.69 / Mtokcloud wins 5.19×	US$1.69 / Mtokcloud wins 5.81×	US$1.69 / Mtokcloud wins 7.07×	US$1.69 / Mtokcloud wins 8.24×
Effective throughput	US$2.86 / Mtokcloud wins 7.24×	US$1.91 / Mtokcloud wins 6.16×	US$1.69 / Mtokcloud wins 5.81×	US$1.68 / Mtokcloud wins 5.76×	US$1.67 / Mtokcloud wins 5.73×
Throughput growth rate	US$1.70 / Mtokcloud wins 5.85×	US$1.69 / Mtokcloud wins 5.82×	US$1.69 / Mtokcloud wins 5.81×	US$1.68 / Mtokcloud wins 5.79×	US$1.68 / Mtokcloud wins 5.77×
Checking payback…

General information only, not advice. See our Disclaimer.

The answer

At four years

Where the tokens run, month by month

Share of tokens by where they run

Cloud price decline moves the answer most.

Go deeper

You might be interested in