Discover · Stage 2 of 6
Token Economics Sprint
An independent model of what your AI workloads will cost to run (per token, on rented cloud GPUs or on hardware you own) and the volume at which the cheapest option changes. Built by the team before capacity is ordered, and handed over to you at the end.
- Duration2 weeks
- Pricing through your partners or on a call.
Start here if
You have a named, owned use case and are about to order capacity for it.
The Sprint takes a use case a Discovery Sprint has already named and given an owner, and produces the cost case a Proof-of-Value is then run against.
The question it answers
For your actual and forecast AI workloads: which way of paying for AI compute costs least, and at what volume does that change? The Sprint compares three options for inference, training and tuning runs (whichever apply to your workloads), computed from your own figures.
Staff, software and support, network, storage and space are counted for every option. The answer comes back as a range and conditions: where the break-even points fall, the range around them, and what would move them.
Metered
Paying a provider per token through an API.
Rented
Paying for cloud GPU instances by the hour: on-demand, reserved or committed.
Owned
Hardware you buy or lease, run on-premises or in colocation.
What you get
Eight deliverables. The report and the working model carry the answer; the rest make it usable by the people who have to act on it.
Workload inventory
Every workload, its type and its owner.
Your current AI bill
Metered, rented and owned, itemized as it stands today.
A written report
The range and conditions, a method appendix, and every number sourced.
The working model
Runnable by your finance team: change an input and see the answer change.
Infrastructure sizing
Accelerators, nodes, and network and storage bandwidth, supplier-neutral.
Scenario cards
Two to four one-page illustrations of the sizing in practice. Real quotes can sit beside them.
Executive readout
AI cost trends and the choice between metered, rented and owned.
Infrastructure readout
The detailed cost comparison for this workload, for your infrastructure team.
How the six phases run
Each phase ends with a specific deliverable in hand.
- Workload inventory. AI usage sorted into the four types that use compute differently (conversational, retrieval-grounded, multi-step agentic, and training or tuning runs), each with a named owner.
- Baseline. What you spend today on metered, rented and owned capacity, itemized against measured throughput, concurrency and latency targets. Where usage is not measured yet, the report says so.
- Adoption forecast. Demand per use case and user group over 12 to 36 months, calibrated against any pilot data. Without a pilot, the forecast is labeled as an assumption.
- Unit economics. Cost per million tokens for inference and per run for training and tuning, for each option, with power, cooling and realistic utilization counted.
- Break-even and sensitivity. The break-even points between the three options, tested by Monte Carlo simulation across the two assumptions that move them most: how fast adoption ramps and how fast token prices keep falling.
- Infrastructure translation. Forecast demand and latency targets converted into accelerators, nodes, and network and storage bandwidth, then shown as two to four scenario cards.
What the model needs
Eleven inputs, asked for by name in the first week: volume and concurrency, where the model sits in the workflow, model class, training and tuning runs, utilization, power, retention and data movement, current bills and contracts, hardware cost basis, hardware already in use, and operating costs. The report shows where each one came from.
Where an input does not arrive, the cell stays blank, the report names the blank, and the sensitivity analysis shows how much the answer depends on it. A named blank is something you can act on.
What a read-out looks like
Two recent read-outs, labelled by sector. Client names and figures are withheld: each card shows the question asked and the shape of the answer.
Global technology company · 2026
The question: which way of paying for AI compute costs least for its planned inference workloads, and at what volume that changes.
The answer: metered, rented and owned compared for each workload, the volume at which the cheapest option changes, and break-even ranges with the conditions that move them. The working model was handed over to its finance team.
Enterprise technology provider · 2026
The question: what its forecast AI workloads will cost to run over the next 12 to 36 months, and which option to plan capacity around.
The answer: metered, rented and owned compared on the same basis, break-even ranges and the conditions that move them (how fast adoption ramps, how fast token prices fall), and sizing stated in accelerators, nodes and bandwidth. The working model was handed over to its team.
How the model stays independent
We don't resell the hardware, cloud or licenses we recommend.
Keep buying tokens is an allowed answer
Continuing to buy metered or rented capacity is a result this model can return, and we report it when the numbers support it.
Supplier-neutral sizing
Sizing is stated in accelerators, nodes and bandwidth, so it can sit beside any quote. Your current-state baseline uses your own contracts.
Every figure shows its source
Your meter or invoice, a published rate with the date it was read, or our estimate, marked as ours.
The same method however it is funded
The result does not depend on who sells the infrastructure: the method, sources and supplier-neutral sizing are the same however the engagement is funded.
Good to know
What it does not cover
The Sprint sizes the capacity the forecast needs; choosing a supplier, accelerator and configuration is your decision, informed by it. Platform design (networking, storage architecture, scheduling and the operating model) is a separate engagement, and a build is tested in a Proof-of-Value.
Your data stays in your environment
Only aggregated extracts leave it (usage and volume figures, bills and rate cards, hardware and power figures), shared the way you approve and deleted or returned after the engagement. The Sprint asks for no prompts, model outputs or personal data.
The model is licensed to you
You keep your copy and your inputs, to run and change for internal use. The model itself remains BridgeTek's intellectual property.
Forecasts are estimates
Every range, break-even point, scenario card and sizing is an estimate based on stated inputs and assumptions. Actual results will differ.
Who delivers it
One of the team's senior data scientists, drawn from the twelve people who staff every stage. Their prior work includes GPU kernel and inference optimization, production retrieval and agentic systems, demand forecasting for a state regulator, and the observability platform behind a GPU cloud. Meet the team →
- avg experience per person
- 15+ years
- publications
- 39
- citations
- 837
Talk it through with the team
Book a consultation with the team.