Glossary
Exabyte
An exabyte (EB) is a unit of digital information equal to 10¹⁸ bytes: one thousand petabytes or one million terabytes. Its binary counterpart, the exbibyte (EiB), equals 2⁶⁰ bytes and is about 15% larger.
Why the exabyte matters for storage planning
Exabytes once described the internet as a whole, or the combined holdings of the largest distributed storage platforms. Today they describe single organisations: research institutes, media archives, telecom operators, cloud and service providers, and AI programmes that keep every training set, every model version and every checkpoint. At that scale, issues that are rounding errors at a few petabytes (unit conversion, metadata, drive failures, the time it takes to move data) start to set the architecture.
Decimal and binary units
Storage capacity is quoted with two sets of prefixes. SI prefixes (kilo, mega, giga, tera, peta, exa) are powers of 1,000. Binary prefixes (kibi, mebi, gibi, tebi, pebi, exbi) are powers of 1,024, and NIST lists exbi (Ei) as (2¹⁰)⁶ alongside exa as (10³)⁶. The gap compounds with every step:
| Decimal unit | Binary unit | Binary larger by |
|---|---|---|
| terabyte (TB, 10¹²) | tebibyte (TiB, 2⁴⁰) | 9.95% |
| petabyte (PB, 10¹⁵) | pebibyte (PiB, 2⁵⁰) | 12.59% |
| exabyte (EB, 10¹⁸) | exbibyte (EiB, 2⁶⁰) | 15.29% |
Drive vendors sell decimal capacity, while many operating systems and tools report binary units under decimal-looking labels. At one exabyte the difference is about 153 petabytes, more than the whole of many large storage systems.
Physical scale of an exabyte
Holding one exabyte on 20 TB hard drives takes 10¹⁸ ÷ (20 × 10¹²) = 50,000 drives before any redundancy. Protection multiplies that figure:
| Protection scheme | Raw capacity per EB stored | 20 TB drives required |
|---|---|---|
| None | 1.00 EB | 50,000 |
| Erasure coding, 9 data + 3 parity | 12 ÷ 9 ≈ 1.33 EB | ≈ 66,667 |
| Three-way replication | 3.00 EB | 150,000 |
The gap between the last two rows, more than 83,000 drives with their servers, power and floor space, is why erasure coding dominates at this scale. In servers of 60 drives, the erasure-coded layout still fills about 1,112 servers.
Objects, metadata and data movement
Object count depends on average size. At 100 MB per object an exabyte holds 10 billion objects; at 1 MB it holds a trillion. Each object carries metadata (key, size, timestamps, fragment locations, user attributes), and at 1 KB per object a trillion objects carry a petabyte of metadata, an index that needs distributed storage of its own. Enumeration shows the same pressure: an S3 listing returns at most 1,000 keys per request, so listing a trillion keys takes a billion requests, about 28 hours at 10,000 list requests per second.
Growth arrives late. An estate of 100 PB growing 30% a year reaches an exabyte in about 8.8 years (ln 10 ÷ ln 1.3), and the final three years add more than half of the total, so most of the capacity in an exabyte estate is bought on the newest hardware generations. The petabyte storage entry covers the scale one step below.
Moving data is slow at this scale. A 100 Gb/s link carries at most 12.5 GB/s, so one exabyte takes 10¹⁸ ÷ (12.5 × 10⁹) = 8 × 10⁷ seconds, about 926 days, at full line rate with no overhead. That arithmetic underlies data gravity, and transfers out of public clouds add per-gigabyte egress fees on top.
What exabyte scale means for storage architecture
For a team whose holdings are heading toward an exabyte, the binding constraints move away from raw capacity and toward the behaviour of the system over many years.
- Protection overhead becomes a budget line. Choosing 9 + 3 erasure coding over three copies at one exabyte is the difference between roughly 67,000 and 150,000 drives.
- Failure is continuous. At an annual failure rate of 1%, a fleet of 66,667 drives loses about 667 drives a year, close to two a day, so rebuilds run constantly and their performance cost is part of normal operation.
- Hardware refresh never finishes. Over a multi-year life, drive and server generations change, so the platform runs mixed generations and migrates data in the background without downtime.
- First placement tends to be final. With an exabyte taking years to move over one fast link, the site, jurisdiction and cloud where data lands first is effectively where it stays, which ties capacity planning to sovereignty decisions.
- Object count can be the limit reached first. Small average object sizes push metadata and listing load up faster than capacity.
Scality RING at exabyte scale
Scality RING has about 6 exabytes under management across its customer base, and a single RING scales to 300 billion objects; Scality describes it as providing multi-petabyte to exabyte capacity in one logical namespace. Storing one exabyte as 300 billion objects implies an average object of 10¹⁸ ÷ (3 × 10¹¹) ≈ 3.3 MB, and erasure coding schemes in RING are defined per storage class.