From bytes to terabytes: how much data does AI use?
A byte is 8 bits, enough to store one English letter. Each step up is 1,000 times bigger: kilobyte, megabyte, gigabyte, terabyte. A tiny teaching dataset can fit in about 120 bytes. A well-known set of 60,000 small photos is a 163-megabyte download. One monthly public web crawl holds about 396 terabytes.
Last reviewed: 20 October 2026 9 min read
What is a byte?
A byte is a group of 8 bits, and a bit is a single 0 or 1. One bit can only say yes or no. Eight bits together can count from 0 to 255, which is enough to stand for one typed letter, digit or symbol.
That is why plain English text is easy to measure. In UTF-8, a common way of storing text, each basic English letter, digit, space and punctuation mark takes exactly 1 byte. Other characters, such as é or 日, take 2 to 4 bytes.
So the sentence "The toast is done." is 18 bytes: 14 letters, 3 spaces and 1 period.
How do data sizes go up, from bytes to terabytes?
Each named size is 1,000 times the one before it. The names come from the metric prefixes kilo, mega, giga and so on, the same prefixes as in kilometer and kilogram.
| Size | How many bytes | Written as a power of 10 | Something that size (worked out) |
|---|---|---|---|
| Bit | one 0 or 1 (an eighth of a byte) | a single yes or no | |
| Byte (B) | 1 | 10⁰ | one English letter |
| Kilobyte (KB) | 1,000 | 10³ | 3 KB: one small 32 × 32 color picture, the size used in a well-known image dataset |
| Megabyte (MB) | 1,000,000 | 10⁶ | 36 MB: one 12-megapixel photo before it is compressed |
| Gigabyte (GB) | 1,000,000,000 | 10⁹ | about 27 of those uncompressed photos |
| Terabyte (TB) | 1,000,000,000,000 | 10¹² | about 27,700 of those photos |
| Petabyte (PB) | 10¹⁵ bytes | 10¹⁵ | a thousand terabytes |
| Exabyte (EB) | 10¹⁸ bytes | 10¹⁸ | a million terabytes |
| Zettabyte (ZB) | 10²¹ bytes | 10²¹ | a billion terabytes |
The ladder keeps going. Yottabyte is 10²⁴ bytes. In 2022, the General Conference on Weights and Measures added ronna (10²⁷) and quetta (10³⁰), so a quettabyte is now the largest named size.
How big is a sentence, a photo or a dataset?
Here is each size worked out step by step, so you can check the math yourself.
A sentence: 18 bytes. Count the characters in "The toast is done." There are 18, and each one is 1 byte.
A 12-example dataset: about 120 bytes. Picture a small table a child might make about toast: twelve slices, each with a darkness score from 1 to 9 and a label. Saved as plain text, one slice per line, it looks like this:
darkness,label 2,not yet 7,done ...
The heading line is 15 bytes. Each "not yet" line is 10 bytes and each "done" line is 7, counting the invisible end-of-line character. Six of each comes to 15 + 60 + 42 = 117 bytes. A dataset can teach a real idea and still be smaller than this paragraph.
A small color picture: 3,072 bytes. In the RGB color scheme, each pixel is three numbers from 0 to 255, one each for red, green and blue. Each number fits in 1 byte. A 32 × 32 picture has 1,024 pixels, so it needs 1,024 × 3 = 3,072 bytes, about 3 KB.
A phone-sized photo: 36 megabytes, before compression. A 12-megapixel photo has 12,000,000 pixels. At 3 bytes each, that is 36,000,000 bytes, or 36 MB. Photo files are usually much smaller than that, because they are compressed: the file stores a shorter description that rebuilds a very close copy of the picture.
A well-known image dataset: 163 megabytes. CIFAR-10, a well-known image dataset, has 60,000 color pictures, each 32 × 32 pixels, sorted into 10 classes such as bird, cat and truck. Its raw pixels come to 60,000 × 3,072 = 184,320,000 bytes, about 184 MB. Its compressed download is 163 MB.
A large public text collection: about 396 terabytes. Common Crawl, a nonprofit, publishes an open archive of web pages. Its August 2026 crawl holds 2.14 billion web pages, or 360 tebibytes of uncompressed content. A tebibyte is a little bigger than a terabyte (see the next question but one), so that is about 396 terabytes.
How much data does AI use?
There is no single answer. It depends on what the system does and how it was built.
- A classroom machine that sorts toast can learn from a dozen examples: about 120 bytes.
- A picture classifier in a university course might train on 50,000 small pictures and test on 10,000 more, as CIFAR-10 is split. That is a few hundred megabytes.
- Large text systems are trained on far more text than any person could read. A single month of public web pages, as above, is hundreds of terabytes before anyone removes duplicates or cleans it up.
Text systems also count their input in tokens, small chunks of text, rather than in bytes. A token is often a word or part of a word. See What is a token in AI?.
How do you compare two data sizes?
Put both sizes in the same unit, then divide.
- Convert. Write both in bytes, or both in megabytes. 3 KB is 0.003 MB.
- Divide. How many 3 KB pictures fit in 163 MB? 163,000 KB ÷ 3 KB is about 54,000. The real dataset holds 60,000, because compression makes each one smaller than 3 KB on average.
- Count the zeros. Each step up the ladder adds three zeros. A terabyte (12 zeros) is a billion times a kilobyte (3 zeros), because 12 − 3 = 9 zeros, and 10⁹ is a billion.
Going further, ages 11 and up. Powers of 10 turn multiplying into adding. 10³ × 10⁶ = 10⁹, because 3 + 6 = 9. That is why scientists draw data sizes on a scale where each step is ten times the last: a byte and a terabyte fit on the same page.
Going further, ages 14 and up. Two sizes can differ by a factor of a billion and still both be called "data". When you read a claim about how much data a system used, ask which unit, counted how (bytes, tokens, pictures), before or after cleaning, and before or after compression. The same dataset can honestly be described with numbers that differ several times over.
Is a kilobyte 1,000 bytes or 1,024 bytes?
In the metric system a kilobyte is exactly 1,000 bytes. Computers work in powers of 2, though, and for years people also used "kilo" to mean 1,024 (which is 2¹⁰).
To tell the two apart, the International Electrotechnical Commission approved separate names for the powers of 2 in December 1998.
| Metric (powers of 10) | Bytes | Binary (powers of 2) | Bytes |
|---|---|---|---|
| kilobyte (KB) | 1,000 | kibibyte (KiB) | 1,024 |
| megabyte (MB) | 1,000,000 | mebibyte (MiB) | 1,048,576 |
| gigabyte (GB) | 1,000,000,000 | gibibyte (GiB) | 1,073,741,824 |
| terabyte (TB) | 1,000,000,000,000 | tebibyte (TiB) | 1,099,511,627,776 |
Storage makers count in metric units. If a computer counts the same drive in powers of 2, the number it shows is smaller: about 7% smaller for gigabytes and about 9% smaller for terabytes. No space is missing; only the unit changed.
What does data size tell you about an AI model, and what doesn't it?
Size tells you:
- How much storage the data needs, and roughly how long it takes to download or copy.
- Whether the data can fit on one computer or needs many.
- A first hint at how much the system has seen.
Size does not tell you:
- Whether the examples are correct, or whether their labels are right.
- Whether the data covers every kind of case the system will meet. A huge dataset can still leave out a whole group of people or situations.
- Whether the examples are repeated. A thousand copies of one example add bytes, not new information.
- Whether the system is good at its job. That only shows when you test it on fresh cases it never saw while learning.
For more on this, see What is a dataset? and What is training data?.
Try it at home: weigh a sentence (no device needed)
You need paper, a pencil and about ten minutes.
- Each person writes one sentence about their day.
- Guess first: how many bytes is it?
- Now count every letter, digit, space and punctuation mark. Each one is 1 byte.
- Who was closest? Whose sentence was "heaviest"?
- Work out how many of your sentences would fill 1 kilobyte (1,000 bytes), then 1 gigabyte.
What to talk about afterwards: a sentence is tiny, so why do AI systems need so much text? And would a thousand copies of the same sentence teach a machine anything new?
Where does Mechna Academy fit?
Mechna Academy is a city of machines for curious children aged 8 to 18, opening first to ages 8 to 10. In every lab, a child guesses how a machine decides, opens it to check, changes one thing and tests it on a case it has never seen.
Questions
A gigabyte is 1,000,000,000 bytes, a billion, in the metric units that storage makers use. A gibibyte, the binary unit, is 1,073,741,824 bytes, which is 2 to the power of 30. Some software uses the binary count but still writes "GB", so check which one is meant.
A petabyte, which is 1,000 terabytes. After that come the exabyte, zettabyte, yottabyte, ronnabyte and quettabyte, each 1,000 times bigger than the one before. Ronna and quetta were added to the metric system in 2022.
One basic English letter, digit, space or punctuation mark is 1 byte in UTF-8, a common text format. Letters with accents and characters from many other writing systems take 2 to 4 bytes.
It ranges from a few hundred bytes to hundreds of terabytes. A classroom dataset of twelve labeled examples can be about 120 bytes, a well-known set of 60,000 small pictures is a 163-megabyte download, and one month of a public web crawl is about 396 terabytes.
Not on its own. Data also has to be correct, cover the kinds of cases the system will meet and avoid endless repeats. The only way to know whether a system works is to test it on fresh cases it never saw while learning.
The box counts in metric units, steps of 1,000. If your computer counts in steps of 1,024, the same space shows as a smaller number, about 9% smaller for a terabyte. Nothing is missing; the unit is different.
Mechna Academy is a city of machines for curious children aged 8 to 18, opening first to ages 8 to 10.
Mechna Academy is a city of machines for curious children aged 8 to 18, opening first to ages 8 to 10.
For grown-ups: join the waitlistPlaces are limited. The waitlist is for parents and guardians 18 and older.