Bits versus bytes — why your internet downloads eight times slower

Mark paid for a "100 Mb/s" connection. The ad promised speed, the technician ran the fibre, the neighbour was envious. Then Mark clicked "download" on a new game and the browser showed him a cruel number: 12 MB/s. A hundred promised, twelve delivered. Where did the other 88 go?

Nowhere. Mark got exactly what he paid for. The advertisement and the browser simply speak two languages that sound almost identical and differ by a factor of eight. This is not fraud — it is units. The whole story of data rates, from the screech of a telephone modem to the fibre under the pavement, is a story about how easily a bit gets confused with a byte, and a symbol with a bit.

Eight bits in a byte

Start with the foundation. A bit (abbreviated with a lowercase b) is the smallest portion of digital information — a zero or a one. A byte (uppercase B) is a bundle of eight bits. One byte is eight bits, not by marketing convention but by definition, written down in the IEC 80000-13 standard and in IEEE 1541.

And there lies the whole misunderstanding. Internet providers measure speed in bits per second, while your computer shows download progress in bytes per second. Since a byte holds eight bits:

1 MB/s (megabyte per second) = 8 Mbit/s (megabits per second).

So a 100 Mbit/s line has a theoretical download ceiling of 100 ÷ 8 = 12.5 MB/s. Mark's browser did not lie — it counts in bytes and divided the advertised number by eight. A rule of thumb for life: to turn megabits into megabytes, divide by 8. 300 Mbit/s is ~37.5 MB/s; 1 Gbit/s is ~125 MB/s.

Watch the capitalisation: Mb is a megabit, MB is a megabyte. One letter, an eightfold difference — the single most common source of confusion in all of computing. Our data rate converter takes both units literally: 1 MB/s in it is exactly 8,000,000 bit/s.

Why providers speak in megabits

If it is so misleading, why not simply quote megabytes? Two reasons.

First — telecom tradition. Networks have always been counted in bits per second. The roots of the convention reach back to telegraphy and early modems, where the basic portion was a signal carrying a single bit. When Ethernet, Wi-Fi, cellular networks and fibre arrived, all of them inherited the same unit. The IEEE 802.3 standards, the ITU-T recommendations, the datasheets of networking gear — everything counts in bits. It is a coherent language for the physical layer: a cable or a transmitter spits out bits, not bytes. The byte belongs to a layer above, to memory and files.

Second — the number is eight times bigger and looks better. "100 Mb/s" sounds far more impressive than "12.5 MB/s", although it means precisely the same thing. Marketing does not pass up an opportunity like that. It is worth stressing, though, that no unit is being manipulated here. The manipulation would be writing MB where Mb is meant — and such "typos" do turn up in advertising.

Networks count in decimal

Here a second, subtler trap waits. In network speeds the prefixes are decimal, exactly as in the SI system:

• 1 kbit/s = 1,000 bit/s
• 1 Mbit/s = 1,000,000 bit/s (10⁶)
• 1 Gbit/s = 1,000,000,000 bit/s (10⁹)

This matters, because in computer memory the same prefixes traditionally meant powers of two — a "kilobyte" was often understood as 1,024 bytes. To tidy up the chaos, the International Electrotechnical Commission (IEC) introduced separate binary prefixes in December 1998: kibi (Ki = 1,024), mebi (Mi = 1,048,576), gibi (Gi = 1,073,741,824), today codified in IEC 80000-13. NIST states plainly that SI prefixes refer strictly to powers of ten and should not be used for powers of two.

In data rates, however, only the decimal convention applies. A gigabit Ethernet card really does clock a thousand million bits per second, not 2³⁰. That is precisely why a "1 TB" drive shows up as ~931 GB in the operating system — and why line speed has nothing to do with that phenomenon: one concerns capacity, the other pace. We untangle it separately, in why a 1 TB drive shows 931 GB and megabytes versus mebibytes.

Where the rest of the bandwidth goes

Even after converting correctly to bytes, Mark will not see a clean 12.5 MB/s. He will see slightly less — and that too is normal. Nominal throughput is not the same as usable throughput. Along the way, overhead eats part of the pipe.

Every portion of data travelling through a network is wrapped in headers. In a typical Ethernet frame with a payload (MTU) of 1,500 bytes, the IP and TCP headers take 20 bytes each — leaving 1,460 bytes for user data, an efficiency of about 97.3%. Add the frame's own header and checksum, the preamble, and the mandatory interframe gap, and the arithmetic runs like this: with full-size frames a single frame occupies 1,538 bytes on the wire, so gigabit Ethernet pushes 81,274 frames per second. Multiplied by 1,460 bytes of data, that gives 118.66 MB/s, or about 949 Mbit/s of usable data instead of the nominal gigabit. Roughly 5% was swallowed by the protocols themselves.

On top of that comes line coding — the way bits are physically written into the signal so that the receiver can read them and keep its clock in step. That costs too:

• 100BASE-TX (Fast Ethernet) uses 4b/5b coding: four data bits travel as five signal bits.
• 1000BASE-X (gigabit over fibre) and USB 3.0 use 8b/10b: every byte becomes a ten-bit symbol. In both cases a fifth of the transmitted stream is overhead — to carry 1 Gbit/s of data, the cable actually runs 1.25 Gbit/s of raw signal. (Gigabit over twisted pair, 1000BASE-T, takes a different route: PAM-5 coding across four pairs at once.)
• 10-gigabit 10GBASE-R uses the thriftier 64b/66b — a mere two bits of overhead per sixty-four, about 3%.

And then reality intrudes: congestion at peak hours, a weakening Wi-Fi signal, a slow server at the other end. Together these push real throughput to somewhere between 70% and 95% of the nominal figure.

Baud is not bit — the most famous mix-up in telecoms

Baud (symbol: Bd) is the number of symbols — changes in the state of the signal — per second. Bits per second is the number of bits per second. They are not the same thing; the two units coincide only when a single symbol carries exactly one bit.

The name honours the French engineer Émile Baudot, author of the five-bit telegraph code patented in 1874; the unit itself was defined by the CCITT (today's ITU-T) in November 1926, replacing an earlier and far vaguer measure in words per minute.

The whole trick of modern modems lies in packing many bits into one symbol. Take concrete numbers from the ITU-T standards:

• V.32 (9,600 bit/s) ran at a symbol rate of 2,400 baud, carrying an effective 4 data bits per symbol (2,400 × 4 = 9,600).
• V.32bis (14,400 bit/s) ran at the same 2,400 baud, but with 6 data bits per symbol (2,400 × 6 = 14,400). Its constellation held 128 points, because a seventh bit served the trellis error-correcting code.

Notice: two modems, the "9600" and the "14400", ran at an identical symbol rate. Only the number of bits per symbol differed. Calling something a "9600 baud modem" was therefore a mistake — it was a 9,600 bit/s modem, but only 2,400 baud. Where did the error come from? From the earliest modems, such as 300 bit/s, where one symbol genuinely carried one bit and a baud equalled a bit per second. The habit stayed; the technology moved on.

The pinnacle of the art was the V.34 modem: at symbol rates reaching 3,429 baud it crammed close to 9.8 data bits into a symbol, drawing on QAM constellations of up to 1,664 points, and reached 33,600 bit/s — not far from Shannon's theoretical limit for a telephone line. The standard specified symbol rates of 2,400, 3,000 and 3,200 as mandatory, with 2,743, 2,800 and 3,429 optional.

A short history of the chase for speed

Dial-up. ITU-T V.90 (1998) offered up to 56 kbit/s downstream and 33.6 kbit/s upstream — the asymmetry was deliberate, since we download more than we upload. Its successor, V.92, raised the upstream to 48 kbit/s. The ceiling came from simple arithmetic: 8,000 samples per second × 7 bits = 56,000. In the United States, FCC rules capping signal power trimmed the real maximum to ~53.3 kbit/s, and a poor line would yield 40–50 kbit/s.

ISDN. The digital line offered two 64 kbit/s channels, which could be bonded into 128 kbit/s.

ADSL. The G.992.1 (G.dmt) standard gave up to 8 Mbit/s down and about 1 Mbit/s up. The later ADSL2+ (G.992.5) raised the bar to 24 Mbit/s — provided you lived close enough to the exchange, since speed fell off with distance.

Cable. DOCSIS 1.0 (1997) delivered ~40 Mbit/s down; DOCSIS 3.0 (2006) reached 1 Gbit/s through channel bonding; DOCSIS 3.1 (2013) hit 10 Gbit/s down; DOCSIS 4.0 adds 6 Gbit/s upstream.

Fibre. GPON (ITU-T G.984) delivers 2.488 Gbit/s down and 1.244 Gbit/s up, shared among the subscribers on one splitter. The newer XGS-PON (G.9807.1, 2016) provides a symmetric 10 Gbit/s both ways.

Cellular. Here the spread between theory and practice is widest: 2G with EDGE reached ~384 kbit/s, 3G promised 2 Mbit/s (in practice usually ~384 kbit/s, peaking at 42 Mbit/s with HSPA+), LTE offered 150 Mbit/s in its basic category, and 5G is quoted in gigabits. Reality is more modest: in Ookla's report for the second half of 2025, the fastest operator on the American market, T-Mobile US, posted a median 5G download of 309.41 Mbit/s, with Verizon at 214 Mbit/s. A long way from the gigabit promises of the laboratory.

Ethernet. The backbone of wired networking grew in leaps: 10 Mbit/s originally, 100 Mbit/s (IEEE 802.3u, 1995), 1 Gbit/s over twisted pair (802.3ab, 1999), 10 Gbit/s (802.3ae, 2002) — and today the standards reach 800 Gbit/s.

What the provider advertises, and what you actually download

Fast internet is two different things

One final distinction, the one the adverts keep quiet about. Bandwidth is how much data fits through the pipe in a second. Latency is the time a single portion needs to travel from A to B. Two different quantities, measured in different units.

Bandwidth is like the width of a motorway; latency is the travel time of one car. For a video call or an online game, low latency matters more than a huge pipe. There is even a phenomenon called bufferbloat: over-inflated buffers in routers can send latency shooting from a dozen milliseconds to several hundred the moment the link comes under load — and more bandwidth does not fix it, sometimes making it worse. So "the fastest internet" does not automatically mean "the most responsive".

The takeaway: how to read a provider's offer

Next time you see "600 Mb/s", perform one operation in your head: divide by 8. You get 75 MB/s — and that, minus a few percent of overhead, is roughly what your browser will show. Remember the capitalisation (b is a bit, B is a byte), remember that networks count in decimal, and remember that real speed always sits a little below the nominal figure.

Mark was not cheated. He got a hundred megabits, he downloads twelve megabytes, and the one follows from the other. It is not a fault and not a scam — it is physics, protocols and a century of telecom tradition packed into one small letter.

Further reading

• IEC 80000-13, Quantities and units — Part 13: Information science and technology — the standard defining the bit, the byte and the binary prefixes.
• NIST, Definitions of the SI units: The binary prefixes (physics.nist.gov/cuu/Units/binary.html) — the official explanation of decimal versus binary prefixes.
• ITU-T, recommendations V.34, V.90 and V.92 — the telephone modem standards, with their constellations and symbol rates (itu.int).
• ITU-T, recommendations G.992.1 (ADSL) and G.9807.1 (XGS-PON) — the copper and fibre standards.
• IEEE 802.3 — the Ethernet family of standards, standards.ieee.org.
• CableLabs, DOCSIS 4.0 Technology (cablelabs.com) — the cable network specifications.
• Ookla, Speedtest Connectivity Report (H2 2025) — median mobile network speeds.
• bufferbloat.net — the project documenting excessive buffering and its effect on latency.