What a Machine Learning Project Really Costs (2D)

Project Size

Approach

Time horizon 1 year

Live Readout

Total cost

Hidden share

Largest category

—

TCO vs. year 1

1.00×

Cost Breakdown

How it works

Each of the six categories draws from a fixed year-1 budget split (data, compute, talent, tooling, integration, maintenance) scaled to the total budget for the chosen project size. Building in-house multiplies each category by its own build-multiplier; buying a vendor platform swaps in a buy-multiplier instead — some categories (tooling) get more expensive to buy, others (talent) get cheaper:

base_i = total(size) × split_i × (build ? buildMul_i : buyMul_i)
extraYears = years − 1
withYears_i = base_i × (1 + recur_i × extraYears × 0.65)
hidden_i = includeHidden ? withYears_i × hiddenFrac_i : 0
cost_i = withYears_i + hidden_i

"Recur" is the fraction of a category's cost that repeats every additional year — compute and maintenance recur heavily, one-time data prep and integration barely do. That compounding is the Total Cost of Ownership (TCO) effect shown in the year-by-year line on the right; the "hidden" cap on each bar is the amount teams typically underestimate (labeling rework, drift monitoring, compliance review, delay).

Bar chart — one bar per category; the translucent amber cap is the hidden-cost layer.
TCO line — total spend recomputed at each horizon from 1 to 5 years at the current size/approach/hidden settings, with the current year marked.

Cost categories Hidden / underestimated cost TCO over time