Skip to content

AI facts, numbers and statistics

The numbers behind AI, either derived in front of you from first principles or traced to the publisher who actually produced them.

Most AI statistics circulate without a source. A number is published once, gets quoted by a blog, gets quoted from that blog by a slide, and within a year it is common knowledge with no way back to whoever measured it. Several of the most repeated figures in the field turn out, when traced, to be a projection, a single unrepresentative experiment, or a sentence somebody wrote in 2013.

These pages take one of two positions on every figure. Either the arithmetic happens here, with every assumption written down, so that the number stays true when hardware and prices change and you can check it yourself in a spreadsheet. Or the publisher and the year are named in the sentence, so you can go and read what they actually did.

Where the honest answer is that no reliable public figure exists, that is what the page says, along with what it would take to produce one. There are more of those than the internet suggests.

How Many People Use AI? What the Numbers Can and Cannot Tell You

Who publishes AI usage figures, what each one's sampling frame supports, and a physical check you can run on any claimed user count.

9 min read

How Much Compute One Answer Takes: FLOPs Per Token, Derived

The 2 x parameters x tokens rule derived from first principles, worked through for a real request, and checked against published GPU specifications.

10 min read

How Far Model Prices Have Fallen, and How to Measure It Properly

The arithmetic for turning provider price pages into a comparable series, why per-token prices mislead, and how to rebuild the history yourself.

10 min read

How Big the Training Runs Are: Deriving Training FLOPs

The 6ND estimate for training compute, derived and then cross-checked two ways against figures the model papers report for themselves.

10 min read

How Much Text There Is to Train On, and How It Is Counted

The bytes-to-tokens conversion that every dataset figure depends on, the named public corpora that state their own sizes, and why the totals disagree.

10 min read

How Many AI Chips Exist: The Arithmetic and Its Error Bars

Nobody publishes a unit count of AI accelerators. Here are the three routes people use to estimate one, and how wide each one's error bars really are.

9 min read

How Long a Model Stays Current: Measuring Deprecation Properly

The five vendor pages that carry real retirement dates, the vocabulary problem that makes naive measurement wrong, and how to compute lifespan yourself.

9 min read

What AI Engineers Earn, and Which Salary Sources Are Worth Reading

The named salary sources, what each one's sampling and levelling can support, and the arithmetic that makes two total-compensation figures incomparable.

9 min read

How Many Companies Actually Ship AI: Reading Adoption Surveys

Why official business surveys and consultancy surveys disagree by an order of magnitude, derived from response bias, and which sources support which claims.

10 min read

The Benchmark Scores People Quote, and What Must Be Attached

The eight settings that must accompany a benchmark score for it to mean anything, with the statistics showing when a gap between two models is noise.

10 min read

How Much Carbon Inference Emits: The Calculation, as a Range

Energy per token derived from accelerator power and throughput, multiplied by grid intensity, presented as the two-order-of-magnitude range it honestly is.

11 min read

How Big a Model Is on Disk: Parameters Times Bytes Per Weight

Model file size derived at every precision, plus the KV cache arithmetic that explains why the download size is not the memory you need.

10 min read

How Many Tokens Are in a Book, a Codebase, a Year of Email

A reference table built from published word counts and a stated tokens-per-word ratio, with the script to reproduce every row for your own tokenizer.

10 min read

How Fast Models Generate: The Roofline and What It Cannot Tell You

The memory-bandwidth ceiling on generation speed, derived, plus the batch size at which it stops applying and why quoted figures rarely describe your request.

10 min read

How Many AI Papers Appear Each Year, and the Query That Counts Them

The exact API queries that return paper counts from arXiv, OpenAlex and Crossref, and the five counting decisions that change the answer by a factor of three.

9 min read

What the Big Labs Spend: Which Figures Are Filed and Which Are Not

The exact line items in public filings that carry AI spending, the API that retrieves them, and why nothing comparable exists for the private labs.

10 min read

How Many Languages Models Really Support: Claimed Against Evaluated

The named multilingual benchmarks and what each covers, plus the derived cost of the tokenizer tax that makes a supported language expensive anyway.

10 min read

How Often Models Hallucinate, According to the Studies

Why hallucination rates from different studies are not comparable, what each named benchmark actually measures, and how to compute a rate for your own task.

10 min read

How Much Electricity Training One Model Uses: Disclosures and Gaps

The published training-energy disclosures by name and year, the arithmetic that reproduces them from GPU-hours, and the models nobody has disclosed anything for.

10 min read

The AI Numbers Everyone Quotes That Nobody Can Source

Five famous AI figures traced back to their origin, what each original actually said, and the method for tracing the next one yourself.

11 min read

Other topics