On 6 July 2026, a document about two hundred words long went out from an office at 61 avenue de l'Observatoire, in the fourteenth arrondissement of Paris. It is set in a fixed-width font, on letterhead that carries a telephone number (+33 1 40 51 23 35) and an email address, and it is addressed "To authorities responsible for the measurement and distribution of time."

The operative part runs to three lines.

NO leap second will be introduced at the end of December 2026. The difference between Coordinated Universal Time UTC and the International Atomic Time TAI is: from 2017 January 1, 0h UTC, until further notice : UTC-TAI = -37 s

Underneath is a name, Christian Bizouard, and a title, Director, Earth Orientation Center of IERS. The document is called Bulletin C 72. One like it arrived in January, saying the same thing about June, and one will arrive in January 2027, saying something about the following June. This is how the world is told what time it is.

Thirty-seven seconds is the entire accumulated disagreement between two ways of counting. It has not changed since New Year's Day 2017.

The office is small. The IERS annual report for 2018, the most recent roster I could find, lists six people at the Earth Orientation Centre, Bizouard among them as its head. Whether all six are still there in 2026 I cannot say. Bizouard signs the bulletins.

The same report describes the working rhythm. The long-term series is delivered officially on Tuesdays and Fridays, updated every day, and the values are checked several times a week, "even during the week-end with the active participation of Jean-Yves Richard."

Two clocks that do not agree

The first way of counting is atomic. International Atomic Time is computed at the Bureau International des Poids et Mesures in Sèvres, on the western edge of Paris, from the readings of roughly 450 high-performance clocks kept in about 85 laboratories around the world. It is beautifully steady and completely indifferent to the planet.

The second way is astronomical. UT1 is the time you get from measuring how far the Earth has actually turned, which is done with radio telescopes observing quasars, with lasers bounced off satellites and off the Moon, and with the GPS constellation. The Earth is not a good clock. Tides drag on it, the atmosphere pushes it around seasonally, the liquid outer core exchanges angular momentum with the mantle on decadal timescales, and the whole thing is still relaxing from the last ice age.

Civil time is a treaty between the two. UTC ticks at the atomic rate and stays within an integer number of seconds of TAI, and the integer gets adjusted so that UTC never drifts more than 0.9 seconds from the Earth's actual rotation. When the gap approaches the limit, the Earth Orientation Centre in Paris announces a leap second, and everyone with a clock is expected to comply.

This has happened twenty-seven times since 1972. Every single one was an insertion: an extra second, written 23:59:60, wedged in at the end of June or the end of December while the planet caught up. The most recent was on 31 December 2016. There has not been a twenty-eighth.

A planet that started running early

Around 2020 the Earth began to spin faster. In the IERS record, which I downloaded and parsed rather than take on trust, the annual mean length of day dropped below its nominal value for the first time in 2020, by 0.0035 milliseconds. In 2022 the mean was 0.2487 milliseconds short. The shortest day ever recorded fell on 5 July 2024, at 1.6568 milliseconds under 86,400 seconds.

Milliseconds are not much. They accumulate. The gap between rotation time and atomic time, UT1 minus UTC, sat at minus 0.256 seconds on 5 June 2020 and climbed steadily from there. A rising gap of that kind, if it kept rising, would eventually reach plus 0.9 seconds, and the treaty would require the opposite operation to the one everybody knows. A second of civil time would be removed. The clock would read 23:59:58 and then, without ceremony, 00:00:00.

Nobody has ever done this. Operating system kernels, filesystems, database transaction logs, exchange matching engines and GPS receivers have all met an inserted second, some of them badly. None has ever met a subtracted one in production. The Unix time model, which underlies most of the software in the world, counts seconds since 1970 on the assumption that every day contains exactly 86,400 of them. It has no way of writing 23:59:60, which is why positive leap seconds are handled by repeating a value or stepping the clock, and it has never been asked to skip one.

The people who worry about this professionally have the incident reports.

The rehearsals

On 30 June 2012 a leap second was inserted, and a bug in the Linux kernel caused timers to fire repeatedly and pin CPUs at full load. The Amadeus Altea reservation platform went down for about an hour. Staff at Qantas and Virgin Australia checked passengers in by hand. Reddit, LinkedIn, Mozilla, Yelp and Foursquare all had a bad night. Jason Harvey, who was on call at Reddit, initially assumed the trouble was fallout from the large Amazon outage of the previous day.

The 2015 leap second passed quietly, mostly because everyone had spent three years preparing for it.

The instructive one came at the end of 2016. At midnight UTC on New Year's Day 2017, inside Cloudflare's custom DNS software, a number went negative when it should have been at worst zero. The code is unremarkable. It measured how long an upstream resolver took to answer by subtracting one timestamp from another, using Go's time.Now(), which offers no guarantee that the clock only moves forward. When the operating system stepped the clock backwards over the leap second, a resolver that had answered in three milliseconds appeared to have answered in slightly less than no time at all. Cloudflare smooths those measurements, so a handful of negative readings eventually dragged the smoothed average below zero, and the smoothed average was handed to a random number function that panics when given a negative argument.

About 0.2 per cent of DNS queries were affected at peak, and fewer than one per cent of HTTP requests to Cloudflare returned an error. Engineers were escalated to at 00:10 and had confirmed the issue by 00:34. The last of the impact ended at 06:45 UTC. Cloudflare's write-up gives the root cause in one sentence: "the belief that time cannot go backwards."

That was a positive leap second. The backwards jump was a side effect of how the clock was adjusted, and it took down a fraction of the DNS for a global network with 102 data centres. A negative leap second would produce the same reversal deliberately, everywhere, at a scheduled moment.

Google has, since 2008, declined to step its clocks at all, and instead smears the leap across a 24-hour window from noon to noon, running every second about 11.6 microseconds long. The frequency error is 11.6 parts per million, comfortably inside the thermal drift of an ordinary quartz oscillator. Amazon uses the same smear in AWS. Google's documentation has already thought about the reverse case: a negative leap second, it notes, "would be smeared by speeding up clocks over the 86,399 SI seconds from noon to noon."

A smeared clock is not showing UTC. For a day either side of the event, the largest cloud providers in the world are deliberately wrong by up to half a second, in slightly different ways from each other, and none of them are ticking SI seconds while they do it.

The ice

In March 2024, Duncan Carr Agnew of the Scripps Institution of Oceanography published a four-page paper in Nature with a title that reads like a shrug: "A global timekeeping problem postponed by global warming."

Agnew's argument goes through the Earth's interior. Satellite gravity measurements show that meltwater from Greenland and Antarctica has been redistributing mass from the poles towards the equator, which makes the planet very slightly wider at the waist and therefore slower, the way a spinning skater slows by extending her arms. Strip that effect out of the observed rotation, and what remains is a clean signal: since 1972 the liquid core has been slowing at a constant rate, and the rest of the planet has been speeding up to compensate.

Extrapolate the core trend forward, and the paper concludes that UTC as currently defined "will require a negative discontinuity by 2029." Then the line that got it onto the news bulletins: "If polar ice melting had not recently accelerated, this problem would occur 3 years earlier." Three years earlier than 2029 is 2026. The ice sheets bought the world's timekeepers roughly the interval we are living in.

The paper models trends and projects them. It does not announce a date.

A committee, and a decision to make the problem impossible

On 13 March 2025, between 14:00 and 16:00 UTC, the BIPM's Consultative Committee for Time and Frequency and the IERS convened a joint webinar: seven short presentations, two hours, one question, covering the observations, the historical geophysics, the core, and the stochastic prediction, aimed at arriving at a single probability.

The number they arrived at was 30 per cent: roughly a three-in-ten chance of needing a negative leap second before 2035, rising steeply after that. The committee then ran the arithmetic backwards. To hold the risk down to something like 5 per cent, given how badly Earth rotation resists prediction, the leap second would have to be gone by 2027 or 2028.

The task group weighing the replacement, co-chaired by Judah Levine at NIST in Boulder and Patrizia Tavella, who heads the BIPM's time department, worked from a table of candidate tolerances. One second, the status quo, meaning an adjustment roughly every eighteen months. Sixty seconds, about a century. Two hundred and fifty-six seconds, justified in the presentation as a way to "maximize use of bytes in computing," about five hundred years. Three thousand six hundred seconds, four thousand years. Infinite.

Which is what is now on the table. In November 2022, the General Conference on Weights and Measures passed Resolution 4, deciding that the 0.9-second tolerance would be widened "in, or before, 2035," and instructing its committees to bring back a specific proposal in 2026. Among the reasons the resolution gave for acting at all was that recent observations of the Earth's rotation "indicate the possible need for the first negative leap second whose insertion has never been foreseen or tested."

That proposal exists. It is Draft Resolution C, "On the technical actions needed to ensure the continuity of UTC," version 5, dated 13 July 2026, and it decides two things: that continuous UTC becomes effective on 20 May 2027, and that the maximum permitted value of the gap between rotation and civil time becomes 3,600 seconds. One hour. At current rates that postpones the next adjustment by several centuries.

A negative leap second, the draft says, "is considered to pose a high risk of causing anomalies and disruption to critical infrastructures that are largely unprepared," and preparing for one would cost the affected industries something "similar to their preparations for the millennium bug." The proliferation of incompatible smears is listed as a defect in its own right. So is the risk that exasperated users abandon UTC for GPS time, "seriously compromising the concept of international standardization in metrology." The recitals are unusually candid for a metrology document.

The vote takes place at the Palais des Congrès de Versailles, 10 rue de la Chancellerie, from Tuesday 13 to Thursday 15 October 2026, chaired by Françoise Combes, President of the Académie des sciences. Under rules written into the Metre Convention of 1875, each member state has one vote. The implementation date of 20 May happens to be World Metrology Day.

The last time the question reached a vote elsewhere, at the ITU in 2015, the alignment was reported at the time as the United States and China wanting the leap second gone, the United Kingdom and Russia wanting civil time kept tied to the sun. Russia's GLONASS satellites, unlike GPS, carry leap seconds in their broadcast signal. That was eleven years and one CGPM resolution ago, and the delegations have not been the same since.

The Earth's own opinion

While the committees were fixing dates, the planet was doing something the models did not call for.

The gap between rotation and atomic time peaked at plus 0.09477 seconds on 17 October 2025 and has been falling since. On 26 August 2025 it stood at plus 0.081 seconds. On the same date this year it read plus 0.007, a fall of 0.074 seconds in twelve months. The most recent reading available, for 27 August 2026, put the gap at plus 0.006 seconds. Mean length of day for 2026 so far runs 0.28 milliseconds long, the first positive annual figure since 2019. The IERS Bulletin A of 27 August 2026 projects the gap continuing downward, to roughly minus 0.095 seconds by next August, though that projection is a straight line with a seasonal correction bolted on, and the service says as much.

Set against the threshold, the five-year scare moved the number by about a third of a second, out of the 0.9 required. As of the last observation, the world is 0.894 seconds away from needing to subtract anything at all.

None of this refutes Agnew, who was modelling decades. It does mean that the committee assembling in Versailles in October will vote to abolish an operation that the Earth has, for the moment, stopped demanding, on the grounds that if it ever does demand it the answer will be an event nobody has rehearsed.

If Draft Resolution C passes, the leap second that ended 2016 was the last one. What gets retired at Versailles is a habit: asking the planet what time it is, and amending the paperwork to match.

A note on sources

Bulletin C 72 (6 July 2026) and Bulletin A, Vol. XXXIX No. 035 (27 August 2026) are published by the IERS Earth Orientation Centre at the Paris Observatory. All length-of-day and UT1−UTC figures quoted here are my own computations from two primary IERS data files, the EOP 20 C04 long-term series and finals2000A.all, retrieved on 31 August 2026. Resolution 4 of the 27th CGPM (2022), Draft Resolution C for the 28th CGPM (version 5, 13 July 2026), the convocation and key dates for the Versailles meeting, and the CCTF Strategy 2025–2035 containing the 30 per cent estimate are all published by the BIPM. Duncan Carr Agnew's paper appeared as Nature 628, 333–336 (27 March 2024). The Cloudflare account and all figures from it are from the company's own post-mortem, "How and why the leap second affected Cloudflare DNS." Leap-smear details are from Google's Public NTP documentation. The candidate tolerance values and the membership of the CCTF task group on continuous UTC come from Coleman, "Towards Continuous Universal Time and the Future of the Leap Second," presented to the Civil GPS Service Interface Committee in September 2024. The 2015 alignment of delegations at the ITU is as summarised by the Brookings Institution on 2 December 2015; it is a think-tank account rather than a delegation record, and I have not found a primary source naming France as a proponent, so the piece does not claim one. The 2012 outages were reported at the time by The Register and Wired, and the kernel fault was discussed on the linux-kernel mailing list on 1 July 2012. The staff composition of the Earth Orientation Centre is drawn from the IERS Annual Report for 2018 and may since have changed; Christian Bizouard's tenure is confirmed by his signature on the current bulletin.

Draft Resolution C is a draft. It can be amended or defeated on the floor at Versailles.