For most of history a second was a slice of the spinning Earth. Since 1967 it has been a count: the number of times a caesium atom flickers between two states. The clocks built to hold that count are now so good that the planet itself is the unreliable part.
9,192,631,770Oscillations = one second
The caesium-133 hyperfine transition, counted out. Since the 13th CGPM in 1967, that number is the definition of the SI second — no stars, no pendulum, no Earth. BIPM SI Brochure
01 · The spectrum of precision
Fifteen orders of magnitude, in one strip
Clock quality is measured as fractional frequency uncertainty — the share of itself a clock might be wrong by. A wristwatch sits near one part in a million. The best clock ever built sits near one part in a billion billion.
Fractional frequency uncertainty · log scalecoarse ← → fine
10⁻⁴10⁻⁷10⁻¹⁰10⁻¹³10⁻¹⁶10⁻¹⁹
1×10⁻⁶
Quartz wristwatch illustrative
1.1×10⁻¹³
NBS-6, 1975
5×10⁻¹⁵
NIST-7, 1993
2.2×10⁻¹⁶
NIST-F4 caesium fountain, 2025
8.1×10⁻¹⁹
JILA strontium optical clock, 2024
02 · The ladder
How long before it is wrong by one second
Invert the uncertainty and you get an intuitive number: the time a clock would need to run before it drifted a full second away from perfect time. Bars are on a log scale — every gridline is a factor of one thousand.
Years to accumulate one second of error · log scaleviolet = caesium · red = quartz / optical
Quartz watch illustrative
2 days
NBS-6 1975
300,000 yr
NIST-7 1993
6.3 million yr
NIST-F1 1999
20 million yr
NIST-F4 2025
140 million yr
JILA strontium 2024
39 billion yr
Values for NBS-6, NIST-F1 and NIST-F2 are NIST's own phrasing; the rest are computed as 1 ÷ uncertainty. 39 billion years is roughly three times the age of the universe.
03 · The climb
Five orders of magnitude in fifty years
Caesium fountains improved steadily until they ran into their own physics. Then optical clocks — ticking hundreds of thousands of times faster, in visible light rather than microwaves — jumped a further factor of a thousand in a decade.
Best evaluated systematic uncertainty, by year · lower is better■ caesium standards ■ optical clocks
Why the gap matters. Optical standards have reached systematic uncertainties roughly one hundred times lower than the caesium fountains that legally define the second. The definition is now the weakest link in the chain.
So the second gets redefined. The CGPM approved a roadmap in 2022 with a baseline target of a new, optical definition in 2030 — the first change to the second since 1967, worth an immediate 10–100× in accuracy.
04 · The unreliable planet
Earth stopped keeping up — then overshot
Civil time (UTC) is atomic time nudged to stay within 0.9 seconds of the turning Earth. Tides slow the planet, so leap seconds were added: 27 of them since 1972. Then, around 2020, Earth started spinning faster, and the additions stopped.
Leap seconds inserted into UTC, by decade27 total · last one 31 Dec 2016
9
1972–79
6
1980s
7
1990s
2
2000s
3
2010s
0
2020s
5 July 2024 was the shortest day since precise records began in the 1970s: 1.66 milliseconds under 86,400 seconds. July and August 2025 delivered another cluster of near-record short days. Suspected cause: the liquid outer core slowing, which speeds the crust up.
A negative leap second is coming. Duncan Agnew's 2024 Nature analysis projects UTC will need to subtract a second around 2029 — a minute of 59 seconds, never done before. Melting polar ice, by moving mass toward the equator, has already delayed it by about three years. Separately, the CGPM voted in 2022 to stop leap seconds altogether by 2035.
05 · What it holds up
A nanosecond is thirty centimetres
Satellite navigation is not a positioning system with clocks in it; it is a clock network you read positions out of. Light travels 29.98 cm in a nanosecond, so every nanosecond of timing error is a foot of location error — and everything downstream inherits it.
38 µs
per day
How much faster GPS satellite clocks run than ground clocks, once relativity is accounted for. Uncorrected, positions would drift about 10 km every day.
$1B
per day
Estimated US economic cost of a GPS outage. GPS has generated about $1.4 trillion in private-sector benefits since the 1980s (RTI for NIST, 2019).
100 µs
of UTC
The maximum clock divergence Europe's MiFID II rules allow for high-frequency trading timestamps, with 1 µs granularity — time as financial law.
Nobody owns the time. UTC is not read off one master clock. It is a weighted paper average of roughly 450 atomic clocks in about 85 institutes, reconciled monthly by the BIPM and published after the fact in a bulletin called Circular T.
The chain is thin. Fewer than 20 caesium fountains worldwide serve as primary frequency standards — the small set of instruments that anchor the SI second for every phone, power grid, data centre and stock exchange on Earth.
06 · The next second
Clocks precise enough to feel a step up
General relativity says a clock higher in a gravity well ticks faster. That effect is now bigger than the noise floor of the best clocks — which turns a clock into an altimeter.
Measured across one centimetre. A JILA team resolved the gravitational redshift inside a single millimetre-scale cloud of strontium atoms: a fractional gradient of −12.4×10⁻¹⁹ per centimetre, against a predicted −10.9×10⁻¹⁹ (Bothwell et al., Nature, 2022). Raise the clock a hair and time speeds up, measurably.
What that buys. Relativistic geodesy — mapping elevation and mass underground by comparing clock rates rather than surveying — plus searches for drifting fundamental constants and dark-matter signatures that would show up as tiny, correlated wobbles across a global clock network.
Why it matters
Time is infrastructure
It is a shared public utility, quietly maintained. National metrology labs give away the second; the entire digital economy free-rides on about 450 machines and a monthly bulletin.
The definition is about to change under everyone's feet. A 2030 optical redefinition is the first rewrite of the second in over 60 years — a good moment to notice that these constants are decisions, made by people, in rooms, on a schedule.
Software assumes time is monotonic. It isn't. Leap seconds have taken down airlines, exchanges and cloud providers. A negative one, projected for around 2029, has never been tested against any of it.
If you build systems: check your leap-second handling now, timestamp in TAI or monotonic clocks where you can, and don't assume a minute has 60 seconds.
Sources & notes
BIPM, The International System of Units (SI), 9th ed. (2019) — the second defined by ΔνCs = 9,192,631,770 Hz; definition adopted at the 13th CGPM, 1967.
BIPM, FAQ: redefinition of the second — optical standards reach 10⁻¹⁸ or better, roughly two orders of magnitude below caesium; redefinition would give an immediate 10–100× gain.
Dimarcq et al., "Roadmap towards the redefinition of the second," Metrologia 61 (2024); CGPM (2022) Resolution 5 — baseline redefinition at the 29th CGPM, 2030.
NIST, A Brief History of Atomic Clocks at NIST — NBS-6 (1975): one second in 300,000 years; NIST-7 (1993): 5×10⁻¹⁵; NIST-F1 (1999): 1.7×10⁻¹⁵, ~one second in 20 million years. NIST-F2 (2014): ~one second in 300 million years.
Elvin et al., "Accuracy evaluation of primary frequency standard NIST-F4," Metrologia (2025) — type-B uncertainty 2.2×10⁻¹⁶; operational since April 2025; fewer than 20 caesium fountains worldwide.
Aeppli, Kim, Warfield, Safronova, Ye, "Clock with 8×10⁻¹⁹ Systematic Uncertainty," Phys. Rev. Lett. 133, 023401 (2024) — JILA/NIST strontium optical lattice clock, 8.1×10⁻¹⁹.
Bothwell et al., "Resolving the gravitational redshift across a millimetre-scale atomic sample," Nature 602, 420 (2022) — measured gradient −12.4×10⁻¹⁹/cm vs −10.9×10⁻¹⁹/cm expected.
Agnew, D., "A global timekeeping problem postponed by global warming," Nature 628 (2024) — negative leap second projected ~2029, delayed roughly three years by polar ice melt.
IERS leap-second history: 27 positive insertions since 1972, most recently 31 December 2016. Decade counts tallied from the IERS bulletin C record.
CGPM (2022) Resolution 4 — leap-second insertions in UTC to cease by or before 2035.
IERS length-of-day data, reported by timeanddate.com (2024–2025): 5 July 2024 ran 1.66 ms short — the shortest day since precise measurement began in the 1970s; further near-record short days in July–August 2025 (2025 figures preliminary).
RTI International for NIST, Economic Benefits of the Global Positioning System (2019) — ~$1.4 trillion in US private-sector benefits since the 1980s; ~$1 billion/day cost of an outage.
GPS relativity: satellite clocks gain ≈38 µs/day (≈45 µs gravitational, −7 µs kinematic); uncorrected drift ≈10 km/day. Ohio State University Astronomy, "Real-World Relativity: The GPS Navigation System."
ESMA, MiFID II RTS 25 (in force January 2018) — high-frequency trading business clocks within 100 µs of UTC, 1 µs timestamp granularity.
BIPM, Circular T — UTC computed from ~450 atomic clocks in ~85 institutes worldwide.
Illustrative Quartz wristwatch at ~1×10⁻⁶ (≈1 second per 12 days) is a typical consumer figure used for scale, not a measured standard; its "2 days" ladder bar is derived from a common ±15 s/month spec. "Years to lose one second" values other than NIST's own phrasing are computed as 1 ÷ uncertainty.