Interactive calculator

Exploring AI: Data Napkin Math

Interactive napkin math, back-of-the-envelope estimates, and ballpark figures for training-data value, compensation, and distribution questions. Start with a concrete public benchmark, then inspect the assumptions and share the exact state you used.

The full workspace brings the calculator, inputs library, inspector, trust filters, and shareable URLs into one place.

4 curated scenarios
71 editable inputs
16 cited sources and trackers

Open full calculator Open inputs library Read sources and asks

Section 1

Featured inputs

Start with plain numbers from public sources, grouped by the kind of question you want to ask. This home page surfaces a few anchor inputs; the full editable library lives in the calculator workspace.

Open full inputs library

I'm curious about money I'm curious about training sizes I'm curious about people or audience size I'm curious about data mix

I'm curious about money

Money, payouts, and prices

Revenue anchors, deal values, labour rates, and inference pricing.

Official Reviewed 2026-03-10

Annualized revenue (OpenAI)

billions of dollars

Reported annualized revenue run rate for OpenAI.

Official Reviewed 2026-03-10

Payment made to Reddit by Google

millions of dollars

Reported yearly value of the Google-Reddit data licensing deal.

I'm curious about training sizes

Training sizes and benchmark totals

Token counts, example counts, benchmark sizes, and other scale assumptions.

Official Reviewed 2026-03-10

Total pre-training tokens (Llama 3)

15,000

billions of tokens

Total tokens used to pre-training a model

Official Reviewed 2026-03-10

Total pretraining tokens (OLMo 3 7B)

5,930

billions of tokens

Approximate number of tokens used in the released OLMo 3 7B pretraining mix.

Official Reviewed 2026-03-10

Average words per token

0.75

words per token

A rule-of-thumb conversion between English words and tokens.

Official Reviewed 2026-03-10

Total public questions (Humanity's Last Exam)

2,500

questions

The number of public benchmark questions in Humanity's Last Exam.

Official Reviewed 2026-03-10

Post-training SFT examples (Tulu 3)

939,344

examples

Number of supervised fine-tuning examples in the public Tulu 3 SFT mixture.

I'm curious about people or audience size

People, groups, and distribution targets

Population, audience, workforce, and contributor counts used in per-person math.

Official Reviewed 2026-03-10

Reddit daily active uniques

121

millions of daily active users

Average daily active unique users on Reddit.

Official Reviewed 2026-03-10

Number of people on Earth

8.09

billions of people

Projected world population.

I'm curious about data mix

Data mix and dataset structure

Composition shares, source slices, document size, and related dataset structure.

Official Reviewed 2026-03-10

Pretraining mix share from web text (Dolma v1.6)

percent of tokens

Approximate share of Dolma v1.6 tokens that come from web crawls.

Need the full set? Open the library to search by scenario, filter to official or primary sources, and inspect every cited assumption in one place.

Section 2

Calculator preview

Try the live scenarios right here, then open the full workspace when you want the entire inputs library, source filters, inspector details, and shareable state.

Open full calculator

Best path for first-time readers: pick a starter scenario, tweak one or two assumptions, then inspect the source notes or switch to the inputs library for deeper work.

Open workspace Browse official inputs

Scenarios

Change the assumptions directly on each card. Inline comparison menus let you swap in related public benchmarks without leaving the page, and every input card links to its cited source.

Distributing money

Build Your Own Scenario

Build a custom formula from the shared input library when the curated scenarios do not quite match the question you want to ask.

Custom result

Check formula

dollars

Enter a formula to start exploring.

Keep this advanced builder tucked away until the curated scenarios stop being enough.

Open the builder when you want a custom formula, markdown scaffold, and shareable draft.

Section 3

Read, cite, and follow the numbers

Keep the calculator close to the evidence: working notes, highlighted papers, live indicators, and concrete asks all sit in one supporting section instead of competing with the calculator.

Open discussions GitHub repository

Essay in progress

Open calculator Discuss it

Commissioning a new frontier dataset versus current AI revenue

This working note traces one unresolved tension in the calculator: replacement-cost estimates for commissioned LLM-scale data climb quickly, while current public frontier-AI revenue anchors remain much smaller.

Commissioning estimate 1.01 trillion dollars

OpenAI annualized revenue 10.00 billion dollars

Anthropic annualized revenue 7.00 billion dollars

Combined revenue anchor 17.00 billion dollars

At the current defaults, the commissioning estimate is roughly 59.6x the combined OpenAI-plus-Anthropic revenue anchor. That mismatch is the core tension this placeholder post should unpack.

Reading list

Highlighted papers and trackers

These are the sources doing real work for the numbers on this site.

Tracker

Read source Data deals tracker

A Sustainable AI Economy Needs Data Deals That Work for Generators

Ruoxi Jia et al., NeurIPS 2025 position paper

Position paper plus public tracker on AI data deals, useful for grounding bargaining-power and transparency questions in a concrete set of disclosed agreements.

Position

Read source

The Most Expensive Part of an LLM Should be its Training Data

Urvashi Kandpal and Colin Raffel, ICML 2025 position paper

Strong anchor for the claim that scarce, high-quality data should be treated as the costly input in the AI stack rather than an effectively free byproduct.

Policy

Read source

Foundation Models and Fair Use

Peter Henderson et al., 2023

Early and still-useful argument that fair use is not automatic for model training, especially when developers can foresee substitution, market harm, or memorization.

Policy

Read source

Win-win: How to Remove Copyright Obstacles to AI Training While Ensuring Author Remuneration

Martin Senftleben, 2024

Clear legal-policy proposal for paying authors without making ex ante licensing the only path for model training.

Empirical

Read source

AI and the Dynamic Supply of Training Data

Alexander Peukert et al., 2025

Important economics paper arguing that training-data rules feed back into future creator incentives and therefore into the long-run supply of high-quality data.

Empirical

Read source

Consent in Crisis: The Rapid Decline of the AI Data Commons

Shayne Longpre et al., 2024

Best empirical anchor I found for the claim that the open web is becoming less permissive for AI training, with large-scale evidence from Common Crawl and robots restrictions.

Watch list

Indicators and asks

The calculator stays useful only if disclosures and public reporting improve around it.

Indicators we want to watch

AI company earnings calls, SEC filings, and updated revenue run rates.
Publisher and platform licensing deals, especially renewals with disclosed terms.
Open-model releases that publish training mix, corpus size, or post-training composition.
Benchmark growth and contributor-compensation norms for expert evaluation work.
Inference price movement across flagship and mini model tiers.

Our asks

Publish deal values and renewal terms in ways researchers and journalists can cite.
Break out AI-specific revenue from total company revenue whenever possible.
Disclose pretraining size, post-training size, and mix composition as first-class metrics.
Make contributor compensation for benchmarks and labeling more legible.
Treat data access, licensing, and opt-out infrastructure as measurable public-interest indicators.