Development cost explainer
From elite researchers and advanced computing to a working intelligence platform.
The development journey
The stages below group the same money over the life of the build. They overlap in practice: hiring continues while training runs, and security work starts before testing ends.
Months 12–26
$40M–$60M
40% of the illustrative scenario
The main model is trained on a large cluster over several months, and the run is often repeated after a failure. This is the largest line in the budget: computing time, fast networking between machines, and the electricity behind both.
Overlapping schedule
Month 1 to month 30
The bars overlap on purpose: hiring continues while training runs, and security work begins before testing ends.
Where does the money go
The stages above describe the money over time. These five lines describe the same money by type. Nothing is added twice: a salary appears in a stage and in the people line, not in both budgets.
40% · Sits inside research, team growth, agents and deployment
The people who can build this are few, so they are paid accordingly and have to be retained for years rather than months. Fully loaded cost includes salary, equity, benefits and the support around each person.
A planning assumption, configured below. It is not a claim about average AI salaries.
37% · Sits inside early experiments and large-scale training
Large-scale training needs thousands of GPUs working as one machine, with fast networking between them and capacity that is available when the schedule needs it. Capacity is paid for by the hour, whether or not every card is busy.
Calculated from the four inputs below, with the same arithmetic used in the budget.
15% · Sits inside collection, licensing and experimental iteration
Data is collected, cleaned, processed, licensed and stored, and simulations are built to test behaviour before training is paid for. Much of the cost is repetition, because most experiments end in a decision to try something else.
4% · Sits inside agents, safety evaluation and deployment
Serving the model, monitoring it in operation, running evaluation, hardening it against attack, integrating it with other systems and deploying it. These costs arrive late in the build and continue afterwards as operating costs.
4% · Held once for the whole programme
A single reserve for the assumptions that will be wrong somewhere: a training run that fails, a price that moves, a hire that takes longer than planned. It is not distributed across the stages, so it is shown once.
The two assumptions you can change are below. Everything else on this page stays fixed, so the arithmetic can be checked line by line.
Personnel: configure the assumption
Total personnel cost
$49.0M
$21.8M
per year, all in
Fully loaded means salary, equity, benefits and the support around each person. This is a planning assumption, not a claim that $1.5M is an average AI salary.
Compute: the GPU arithmetic
Computed cost
$45.5M
4,000 × $1.63 × 7,000 = $45.5M
One useful GPU hour: $2.71
Reserved capacity is paid for whether the cluster is busy or waiting, so utilization changes the cost of a useful hour, not the bill. The comparison price is discounted by 35% for the commitment.
Illustrative inputs. The defaults land inside the $40M to $60M training stage.
Internal research & information disclaimer
Illustrative internal modelling, not a quotation. Every figure is a planning assumption built on published unit costs where available; none is a supplier price, an audited programme or an approved budget.
Read the full research disclaimerEvidence behind the estimates
Two published sources inform the scenario. Everything else on this page is a modelled assumption, and it is labelled as one.
Stanford HAI · AI Index Report 2025
Estimates of training compute cost for three published models, as compiled in the AI Index 2025.
Open the AI Index 2025 reportEpoch AI · How Much Does It Cost to Train Frontier AI Models? · 2024
The share of training cost by category, across the models Epoch studied. The ranges overlap because each model differed.
Open the Epoch AI analysisWhy our number is not their number
Those percentages describe the training runs studied in that analysis, not a universal budget for a financial world model. Our scenario includes more than training: data work, agents, security and deployment, each counted once. It sits in the same order of magnitude as the published training estimates, which is why the comparison is useful, and it remains a planning assumption all the same.
Compare the alternatives
The figures below answer different questions. One buys ownership over years, one buys access now, and one widens the ambition.
$100M–$150M
Illustrative development scenario
$150M–$300M
Illustrative expanded scenario
$20M
Project assumption, three years
A licence is not ownership
Three years of access and ownership of the resulting intellectual property are not the same purchase. The licence starts sooner and commits less capital, but at the end of the term the platform is still someone else's, including the roadmap and the terms.
A fourth route worth costing
Integrating existing commercial or open-weight models may cost much less than all three routes above, depending on the capabilities the platform actually needs. It is also a different product: less control over the model itself, and more dependence on another provider's roadmap and terms.
How to read these numbers