Compute economics tool

When Does AI Compute Spend Exceed Payroll?

Nvidia's Bryan Catanzaro said his applied deep learning team now spends more on compute than on the employees using it. Use this calculator to test when that crossover happens for your team, and whether the productivity uplift pays back the compute bill.

Annual payroll
—
Annual compute
—
Compute vs payroll
—
Uplift value / yr
—
ROI on compute
—
Break-even
—
Net annual value (uplift minus compute)Your current compute spendBreak-even line

How the math works

Annual payroll = headcount times fully loaded salary. Annual compute = 12 times monthly compute. Uplift value = headcount times revenue per employee times productivity uplift. ROI = (uplift value minus annual compute) divided by annual compute. Break-even is the number of months of uplift needed to cover a year of compute spend.

Verified fact: Bryan Catanzaro, Nvidia's VP of applied deep learning, publicly stated that his team's compute costs now far exceed employee costs (reported by Forbes, via Unusual Whales).

Illustrative estimates: the default values here — a 40 person team, $280k fully loaded salary, $1.2M/month compute, 25% uplift, $900k revenue per employee — are plausible enterprise AI figures, not disclosed Nvidia numbers.

Results are educational only and not financial advice.

Compute spend, modeled uplift and payback units

Read the explanation

The worked example sets forty people and two hundred eighty thousand dollars fully loaded annual salary. Their payroll is eleven point two million dollars per year. The monthly compute input is one point two million, which multiplies by twelve to fourteen point four million per year. Bars use thirty pixels per annual million dollars. Compute divided by payroll is about one point two nine, so the interface says compute exceeds payroll. These are plausible authored example inputs, not disclosed company figures. Annualization assumes the monthly bill remains constant and does not prove the spend is justified. With nine hundred thousand dollars annual revenue per employee and twenty-five-percent productivity uplift, the source multiplies forty by nine hundred thousand by point twenty-five to obtain nine million dollars modeled annual value. Bars compare this value and fourteen point four million annual compute at thirty pixels per million. Net value is negative five point four million. Dividing that net by annual compute yields negative thirty-seven point five percent, displayed rounded to negative thirty-eight percent by the source numeric formatting. Revenue uplift is an assumption, not measured causal profit, cash flow or realizable labor savings. The model does not subtract review, deployment or financing costs. The source divides annual compute, fourteen point four million, by modeled monthly uplift, nine million divided by twelve. This gives nineteen point two months, and its display branch converts that to one point six years. A different comparison using the recurring monthly compute input would give one point six months, but neither ratio is a cash-flow payback model because no distinct upfront investment is entered. Bars show source nineteen point two and alternative one point six months on the same twenty-five-pixels-per-month scale. The chart sweeps monthly compute and subtracts twelve times that bill from modeled annual uplift; it is sensitivity arithmetic, not a forecast. If modeled uplift is zero the source says never, while zero compute forces the displayed ROI to zero.

Link copied
Super generates helpful tools and automates fact-checking across the internet proactively. If you enjoyed this tool, build your own with Super and share it with a friend.