Main finding
Independent development is a frontier research programme, and the spend is concentrated rather than even: senior research staff first, GPU time second, then repeat experiments, data preparation, infrastructure and deployment. On published unit costs a programme of this class lands between $100M and $300M over 18 to 30 months, with most of it committed before anything works.
Where the money goes
Illustrative $200M programme
Select any row to open its calculation
An illustrative split of the planning range, built on published unit costs. It is not a supplier quotation and it is not an audited programme.
How the three planning figures reconcile
- $100M–$150M describes a focused development scenario: a narrower applied system, a smaller team and fewer training runs.
- $150M–$300M describes a wider frontier research programme. The $200M allocation above is one illustrative composition inside that programme, not the budget of the focused case.
- The $20M three-year licence is a separate internal premise. It should be compared against total cost of ownership, including integration and operating expense, before any saving is claimed.
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. Full disclaimer
What supports it
- Staffing is the largest line: in our planning split it takes 40 to 50 percent of the total, across 60 to 120 people in research, infrastructure, data and evaluation.
- Compensation is two different markets. Public surveys place typical senior machine-learning engineering roles in the low hundreds of thousands of dollars a year, while a handful of frontier researchers command packages an order of magnitude higher; the $1.5M figure used here is that exceptional scenario, not an average salary.
- Compute takes 25 to 35 percent: published frontier training runs sit in the tens to low hundreds of millions of dollars before staff and data (Epoch AI), and cloud providers publish per-GPU list prices that make the unit cost auditable.
- Experimentation takes 10 to 15 percent, because most runs fail: training cycles of eight to sixteen weeks are repeated after negative results and safety evaluation.
- Data and deployment take the remaining 10 to 20 percent: financial and blockchain data must be cleaned, licensed and kept current, and deployment includes serving cost, monitoring and security review.
- The same range is bounded by the compute record: sustained growth in training compute across model generations is documented in the Stanford AI Index.
What this does not claim
- These are planning estimates built on published unit costs, not a supplier quotation or an audited programme.
- The range describes ambitious frontier development. A narrower applied system costs a fraction of it and should be priced separately.
- A $1.5M package is an exceptional hiring scenario. Treating it as a typical salary would overstate the staff line by multiples.
Planning assumptions used
- Sixty to one hundred and twenty people at fully loaded cost, split across research, infrastructure, data and evaluation.
- Training runs of eight to sixteen weeks, with repeat runs after failed experiments and safety evaluation.
- Figures are internal planning assumptions and remain unconfirmed.
Open before signature
- Which capability would we actually have to own rather than license?
- What is the measured cost of a narrower applied system on our own data?
- How sensitive is the range to GPU prices, staff retention and repeat training runs?
Sources
Figures marked illustrative are planning assumptions. This note separates published evidence from project assumptions and does not present a supplier quotation, a valuation or a forecast of returns.