The countdown · when our computers could run each simulation
Each scenario in the sandbox needs a certain number of operations per second. Humanity’s computers add up to a total. This page puts the two on one chart. We measure up to now, then project three scenarios that level off as they hit physical limits. It’s an illustration built on measured data, not a forecast. The three scenarios are ours, and every year below means “if growth stayed at this rate”.
Solid lines are measured. Dashed lines are three scenarios. Each one grows at a steady rate while it’s far from its ceiling, then levels off as it gets close, like logistic growth. Each band shows the same scenario with growth a fifth lower and a fifth higher. That spread is arbitrary. It isn’t a probability. Where a dashed line crosses a finish line, that’s the year our computers could run it, if a civilisation threw its full capacity at the job. The scenarios that go past today’s grid also assume energy per operation keeps falling towards the thermodynamic floor. The best chips have barely improved since 2022, so that’s the shakiest assumption on the page.
Show the numbers
| Year | Fastest single machine | All of humanity’s computers | Fast | Steady | Stalled |
|---|---|---|---|---|---|
| 1946 | 5.0 × 10² FLOPS | ||||
| 1964 | 3.0 × 10⁶ FLOPS | ||||
| 1976 | 1.6 × 10⁸ FLOPS | ||||
| 1986 | 3.0 × 10¹⁴ FLOPS | ||||
| 1993 | 6.0 × 10¹⁰ FLOPS | 4.4 × 10¹⁵ FLOPS | |||
| 1997 | 1.3 × 10¹² FLOPS | ||||
| 2000 | 2.9 × 10¹⁷ FLOPS | ||||
| 2002 | 3.6 × 10¹³ FLOPS | ||||
| 2007 | 6.4 × 10¹⁸ FLOPS | ||||
| 2008 | 1.1 × 10¹⁵ FLOPS | ||||
| 2011 | 1.1 × 10¹⁶ FLOPS | ||||
| 2013 | 3.4 × 10¹⁶ FLOPS | ||||
| 2016 | 9.3 × 10¹⁶ FLOPS | ||||
| 2018 | 1.4 × 10¹⁷ FLOPS | ||||
| 2020 | 4.4 × 10¹⁷ FLOPS | ||||
| 2022 | 1.1 × 10¹⁸ FLOPS | ||||
| 2024 | 1.7 × 10¹⁸ FLOPS | ||||
| 2025 | 10²¹ FLOPS | 10²¹ FLOPS | 10²¹ FLOPS | 10²¹ FLOPS | |
| 2026 | 2.2 × 10¹⁸ FLOPS | ||||
| 2050 | 1.3 × 10²⁶ FLOPS | 1.8 × 10²⁴ FLOPS | 7.0 × 10²² FLOPS | ||
| 2075 | 1.6 × 10³¹ FLOPS | 3.3 × 10²⁷ FLOPS | 2.5 × 10²³ FLOPS | ||
| 2100 | 2.0 × 10³⁶ FLOPS | 6.0 × 10³⁰ FLOPS | 2.6 × 10²³ FLOPS | ||
| 2125 | 2.6 × 10⁴¹ FLOPS | 1.1 × 10³⁴ FLOPS | 2.6 × 10²³ FLOPS | ||
| 2150 | 2.6 × 10⁴⁶ FLOPS | 5.8 × 10³⁵ FLOPS | 2.6 × 10²³ FLOPS | ||
| 2175 | 1.3 × 10⁴⁷ FLOPS | 6.0 × 10³⁵ FLOPS | 2.6 × 10²³ FLOPS | ||
| 2200 | 1.3 × 10⁴⁷ FLOPS | 6.0 × 10³⁵ FLOPS | 2.6 × 10²³ FLOPS | ||
| 2225 | 1.3 × 10⁴⁷ FLOPS | 6.0 × 10³⁵ FLOPS | 2.6 × 10²³ FLOPS | ||
| 2250 | 1.3 × 10⁴⁷ FLOPS | 6.0 × 10³⁵ FLOPS | 2.6 × 10²³ FLOPS | ||
| 2275 | 1.3 × 10⁴⁷ FLOPS | 6.0 × 10³⁵ FLOPS | 2.6 × 10²³ FLOPS | ||
| 2300 | 1.3 × 10⁴⁷ FLOPS | 6.0 × 10³⁵ FLOPS | 2.6 × 10²³ FLOPS |
- Fast. 60% more computing every year, the pace of the 1990s, until a civilisation beyond Earth has used up the Sun’s output. wild guess
- Steady. 35% a year, about the last twenty years, until hardware hits the thermodynamic floor and a slice of Earth’s sunlight goes to computing. wild guess
- Stalled. 20% a year with no new energy supply. Computing grows until it uses the whole electricity grid at today’s efficiency. wild guess
| Finish line | Computing needed | Fast | Steady | Stalled |
|---|---|---|---|---|
| Brain in a Vat A real brain in base reality. Only its world is computed, and everyone else is scripted. | 8.2 × 1015 | already | already | already |
| Solipsism One real mind. The world is only as detailed as that mind can test. | 1018 | already | already | already |
| The Afterlife A world with no suffering, computed only when someone’s there. | 1018 | already | already | already |
| The 2050 Server Everyone on Earth, with only as many real minds as 2050’s computers can run. | 1.8 × 1024 | 2039–2044 | 2046–2055 | never |
| Everyone alive, as minds Eight billion minds, working at the functional level, and no simulated environment. | 8.2 × 1027 | 2054–2066 | 2070–2089 | never |
| The Matrix Present-day Earth with billions of fully computed minds, drawn in only where someone is looking. | 8.2 × 1031 | 2071–2089 | 2097–2127 | never |
| Earth, atom by atom, since 1900 Earth atom by atom since 1900, as a starting point for changing one dial at a time. | 1.3 × 1068 | never | never | never |
| Local Universe The Milky Way at atomic resolution, from the Big Bang. | 7.0 × 1085 | never | never | never |
| Universe Lite The observable universe at molecular resolution, with half of it left out. | 5.8 × 1091 | never | never | never |
On one star’s power, with ordinary logic that erases a bit per operation at room temperature, we never get enough computing to run a full world. Reversible computing and a colder heat sink would lower that floor. We don’t count them here. Past the Sun’s line, the only way to get enough is to turn matter itself into a computer near Lloyd’s bound. That bound is a ceiling for matter as hot as a plasma, not for anything like rock. For the operations alone, Earth atom by atom needs at least 2.5 × 1017 kg. That’s 24 million times less than the Earth. Storing its state takes at least 5 × 1052 bits, and that needs far more matter (over 1018 kg even at the black-hole limit). Bostrom’s argument assumes posthuman civilisations that can build computers like that. That’s why it’s about them and not us.
Drivers of growth
Show the numbers
| Year | Fastest single machine | All computers on Earth |
|---|---|---|
| 1946 | 5.0 × 10² FLOPS | |
| 1964 | 3.0 × 10⁶ FLOPS | |
| 1976 | 1.6 × 10⁸ FLOPS | |
| 1986 | 3.0 × 10¹⁴ FLOPS | |
| 1993 | 6.0 × 10¹⁰ FLOPS | 4.4 × 10¹⁵ FLOPS |
| 1997 | 1.3 × 10¹² FLOPS | |
| 2000 | 2.9 × 10¹⁷ FLOPS | |
| 2002 | 3.6 × 10¹³ FLOPS | |
| 2007 | 6.4 × 10¹⁸ FLOPS | |
| 2008 | 1.1 × 10¹⁵ FLOPS | |
| 2011 | 1.1 × 10¹⁶ FLOPS | |
| 2013 | 3.4 × 10¹⁶ FLOPS | |
| 2016 | 9.3 × 10¹⁶ FLOPS | |
| 2018 | 1.4 × 10¹⁷ FLOPS | |
| 2020 | 4.4 × 10¹⁷ FLOPS | |
| 2022 | 1.1 × 10¹⁸ FLOPS | |
| 2024 | 1.7 × 10¹⁸ FLOPS | |
| 2025 | 10²¹ FLOPS | |
| 2026 | 2.2 × 10¹⁸ FLOPS |
Seventy-nine years, fifteen orders of magnitude: about a thousandfold every fifteen years. measured Source: the twice-yearly TOP500 list (measured on the Linpack benchmark) from 1993, latest June 2026; peak figures from machine histories before that, and the 1946 one is approximate.
Every processor, summed. The 2011 census has not been repeated, so the last stretch is a straight line to a rough anchor. It counts general-purpose instructions, as the census did. AI chips alone reached about 4 × 10^21 low-precision operations a second by late 2024; those are a cheaper unit, and adding them would raise the total about fivefold and bring each steady-growth crossing forward by about five years. rough Source: Hilbert & López, “The World’s Technological Capacity to Store, Communicate, and Compute Information”, Science 332 (2011), counting one instruction as one operation; the 2025 point is the site’s own order-of-magnitude anchor.
Show the numbers
| Year | Energy per operation, best machine |
|---|---|
| 1946 | 3.0 × 10² J |
| 1964 | 10⁻² J |
| 1976 | 7.0 × 10⁻⁴ J |
| 2008 | 1.9 × 10⁻⁹ J |
| 2012 | 4.0 × 10⁻¹⁰ J |
| 2016 | 1.1 × 10⁻¹⁰ J |
| 2020 | 3.8 × 10⁻¹¹ J |
| 2022 | 1.5 × 10⁻¹¹ J |
| 2024 | 1.4 × 10⁻¹¹ J |
| 2025 | 1.4 × 10⁻¹¹ J |
Joules spent per operation fell by a factor of twenty trillion. The red line is the floor physics allows at room temperature; the best chips are still about five billion times above it. measured Source: the Green500 list of the most energy-efficient supercomputers from 2008 (November lists, to 73.3 billion operations per watt in 2025); approximate power and speed from machine histories before that. Koomey et al. (2011) measured a doubling every 1.57 years from 1946 to 2009.
Show the numbers
| Year | Transistors on one chip |
|---|---|
| 1971 | 2.3 × 10³ |
| 1978 | 2.9 × 10⁴ |
| 1985 | 2.8 × 10⁵ |
| 1989 | 1.2 × 10⁶ |
| 1993 | 3.1 × 10⁶ |
| 2000 | 4.2 × 10⁷ |
| 2010 | 3.0 × 10⁹ |
| 2016 | 1.5 × 10¹⁰ |
| 2020 | 5.4 × 10¹⁰ |
| 2022 | 8.0 × 10¹⁰ |
| 2024 | 2.1 × 10¹¹ |
Moore’s 1965 doubling, still roughly on schedule, now by joining two dies in one package. Chips made from a whole wafer reached four trillion in 2024; they are left off because they are not like for like. measured Source: manufacturers’ published counts for the largest chip or package of each year.
Show the numbers
| Year | Largest AI training run |
|---|---|
| 2012 | 6.0 × 10¹⁷ operations |
| 2015 | 4.0 × 10²⁰ operations |
| 2017 | 8.0 × 10²⁰ operations |
| 2020 | 3.0 × 10²³ operations |
| 2022 | 3.0 × 10²⁴ operations |
| 2023 | 2.0 × 10²⁵ operations |
| 2024 | 4.0 × 10²⁵ operations |
| 2025 | 4.0 × 10²⁶ operations |
| 2026 | 10²⁷ operations |
Total operations spent on the biggest training run of each year, growing four to five times a year since 2010. The grey lines are the same total counted in years of one human brain at the functional rate: the 2025 record is about a dozen brain-years. rough Source: Epoch AI’s data on frontier AI models (estimates, rounded to one figure; 2026 is preliminary); growth rate from Sevilla and Roldán (2024).
Hardware is only part of the trend. Better algorithms cut the operations a task needs every year, so the scenarios above may be conservative. Source: Hernandez & Brown (2020); Ho et al. (2024).
Show the numbers
| Year | Data-centre electricity | Data-centre electricity, projected |
|---|---|---|
| 2010 | 194 TWh | |
| 2015 | 200 TWh | |
| 2018 | 205 TWh | |
| 2020 | 270 TWh | |
| 2024 | 415 TWh | |
| 2025 | 485 TWh | 485 TWh |
| 2030 | 945 TWh |
Terawatt-hours a year. Flat for a decade while chips got more efficient, then bent upward by AI; the dashed stretch is the agency’s base case. The whole world used about 29,000 terawatt-hours in 2024. rough Source: Masanet et al., Science 367 (2020), for 2010 to 2018; the International Energy Agency, “Energy and AI” (2025), for 2020 onward and the 2030 projection.
Show the numbers
| Year | Mapped, every synapse | Simulated, simplified |
|---|---|---|
| 1986 | 3.0 × 10² neurons | |
| 2013 | 1.7 × 10⁹ neurons | |
| 2023 | 3.0 × 10³ neurons | |
| 2024 | 1.4 × 10⁵ neurons | 8.6 × 10¹⁰ neurons |
Mapped means every neuron and synapse traced from real tissue: a worm, a larva, a fly. Simulation is a separate measure: 1.7 billion model neurons in 2013, and in 2024 a run with as many units as a human brain, each one simplified to a point with guessed wiring. The top line is one human brain. Human and mouse cubic millimetres (2024, 2025) are mapped too, but they are fragments, not whole brains. measured Source: White et al. (1986), the worm; Winding et al. (2023), the fly larva; Dorkenwald et al. (2024), the adult fly. RIKEN and Jülich (2013) for the K computer run; Fudan University (2024) for the human-scale run.
Machine-generated content
Nobody has measured this for the world as a whole. There’s one real time series, from a single music service counting its own uploads. There are also a few single counts, each for one kind of item in one place. Treat them as samples.
Show the numbers
| Year | Fully machine-made share of new music uploads |
|---|---|
| Jan 2025 | 10% |
| Jun 2025 | 18% |
| Sep 2025 | 28% |
| Nov 2025 | 34% |
| Jan 2026 | 39% |
| Apr 2026 | 44% |
| Jul 2026 | 50% |
Percent of each day’s new tracks. It passed half in July 2026. The often-quoted “90% of online content will be synthetic by 2026” has no traceable source and is not used here. rough Source: a major music-streaming service’s own detector, from its announcements (January 2025 to July 2026).
Interest
We use two stand-ins for interest. One is visits to the encyclopedia article for each idea (English, human traffic only). The other is new scholarly papers that use the phrase. Each chart has its own scale. Hollow dots mark years that aren’t finished. The paper index itself grew fast in 2025 and 2026, so part of the late rise is an artefact of counting, not growth in the field.
Show the numbers
| Year | Simulation hypothesis · visits |
|---|---|
| 2016 | 338,998 |
| 2017 | 300,904 |
| 2018 | 362,320 |
| 2019 | 453,578 |
| 2020 | 449,018 |
| 2021 | 479,322 |
| 2022 | 318,153 |
| 2023 | 338,869 |
| 2024 | 360,061 |
| 2025 | 347,213 |
| 2026 | 165,640 |
Encyclopedia visits a year. 2026 runs to 2026-08.
Show the numbers
| Year | Generative AI · visits |
|---|---|
| 2023 | 561,844 |
| 2024 | 949,100 |
| 2025 | 939,238 |
| 2026 | 294,748 |
Encyclopedia visits a year. 2026 runs to 2026-08.
Show the numbers
| Year | Digital twins · visits |
|---|---|
| 2017 | 51,601 |
| 2018 | 163,210 |
| 2019 | 224,792 |
| 2020 | 241,048 |
| 2021 | 270,726 |
| 2022 | 257,797 |
| 2023 | 204,803 |
| 2024 | 276,219 |
| 2025 | 176,899 |
| 2026 | 83,070 |
Encyclopedia visits a year. 2026 runs to 2026-08.
Show the numbers
| Year | Simulation hypothesis · papers |
|---|---|
| 2003 | 2 |
| 2004 | 2 |
| 2005 | 3 |
| 2006 | 6 |
| 2007 | 8 |
| 2008 | 4 |
| 2009 | 5 |
| 2010 | 5 |
| 2011 | 10 |
| 2012 | 6 |
| 2013 | 11 |
| 2014 | 7 |
| 2015 | 9 |
| 2016 | 12 |
| 2017 | 12 |
| 2018 | 14 |
| 2019 | 22 |
| 2020 | 19 |
| 2021 | 27 |
| 2022 | 17 |
| 2023 | 29 |
| 2024 | 35 |
| 2025 | 92 |
| 2026 | 291 |
Papers a year using “simulation hypothesis” in the title or abstract.
Show the numbers
| Year | Digital twins · papers |
|---|---|
| 2010 | 4 |
| 2011 | 3 |
| 2012 | 10 |
| 2013 | 5 |
| 2014 | 11 |
| 2015 | 13 |
| 2016 | 43 |
| 2017 | 129 |
| 2018 | 493 |
| 2019 | 1,301 |
| 2020 | 2,371 |
| 2021 | 4,228 |
| 2022 | 6,334 |
| 2023 | 9,513 |
| 2024 | 13,036 |
| 2025 | 22,180 |
| 2026 | 32,175 |
Papers a year using “digital twin” in the title or abstract.
Show the numbers
| Year | World models · papers |
|---|---|
| 2000 | 98 |
| 2001 | 157 |
| 2002 | 288 |
| 2003 | 214 |
| 2004 | 186 |
| 2005 | 227 |
| 2006 | 219 |
| 2007 | 204 |
| 2008 | 206 |
| 2009 | 246 |
| 2010 | 266 |
| 2011 | 264 |
| 2012 | 268 |
| 2013 | 257 |
| 2014 | 265 |
| 2015 | 283 |
| 2016 | 233 |
| 2017 | 266 |
| 2018 | 286 |
| 2019 | 299 |
| 2020 | 357 |
| 2021 | 357 |
| 2022 | 374 |
| 2023 | 533 |
| 2024 | 803 |
| 2025 | 1,744 |
| 2026 | 5,352 |
Papers a year using “world model” in the title or abstract.
Show the numbers
| Year | AI-generated · papers |
|---|---|
| 2010 | 1 |
| 2011 | 3 |
| 2012 | 1 |
| 2013 | 2 |
| 2014 | 1 |
| 2015 | 6 |
| 2016 | 8 |
| 2017 | 6 |
| 2018 | 24 |
| 2019 | 75 |
| 2020 | 140 |
| 2021 | 208 |
| 2022 | 295 |
| 2023 | 1,752 |
| 2024 | 4,867 |
| 2025 | 13,032 |
| 2026 | 22,260 |
Papers a year using “AI-generated” in the title or abstract.
Show the numbers
| Year | Synthetic data · papers |
|---|---|
| 2000 | 250 |
| 2001 | 283 |
| 2002 | 537 |
| 2003 | 524 |
| 2004 | 573 |
| 2005 | 637 |
| 2006 | 850 |
| 2007 | 875 |
| 2008 | 945 |
| 2009 | 1,028 |
| 2010 | 1,111 |
| 2011 | 1,253 |
| 2012 | 1,300 |
| 2013 | 1,422 |
| 2014 | 1,558 |
| 2015 | 1,620 |
| 2016 | 1,818 |
| 2017 | 2,001 |
| 2018 | 2,272 |
| 2019 | 2,772 |
| 2020 | 3,214 |
| 2021 | 3,913 |
| 2022 | 4,379 |
| 2023 | 5,402 |
| 2024 | 7,243 |
| 2025 | 10,796 |
| 2026 | 12,775 |
Papers a year using “synthetic data” in the title or abstract.
Show the numbers
| Year | Whole brain emulation · papers |
|---|---|
| 2008 | 1 |
| 2010 | 1 |
| 2011 | 1 |
| 2012 | 7 |
| 2013 | 5 |
| 2014 | 3 |
| 2015 | 6 |
| 2016 | 2 |
| 2017 | 7 |
| 2018 | 6 |
| 2019 | 4 |
| 2020 | 3 |
| 2021 | 3 |
| 2022 | 7 |
| 2023 | 2 |
| 2024 | 7 |
| 2025 | 13 |
| 2026 | 47 |
Papers a year using “whole brain emulation” in the title or abstract.
Belief
No repeated measurement. There is no survey of the public, or of any profession, that has asked whether we live in a simulation in two different years. So there is no trend to draw, and this page does not draw one. measured
Philosophers, 2020. The largest survey of professional philosophers had no question about simulation. The nearest is about the outside world: 79.5% chose non-sceptical realism, that it exists and we can know it. A well-made simulation would get the same answer. measured
Unsourced figures. A “22% of adults” figure and older claims about a television network’s poll are widely repeated; neither could be traced to a real poll, so neither is used here. The interest charts above stand in as proxies: public lookups of the idea and scholarly publication on it. wild guess
- A scenario is the logistic curve written in log10: growth r a year, levelling off towards a ceiling C. From 1021 in 2025, the level in year t is C − log10(1 + (10C−21 − 1) · (1 + r)−(t−2025)).
- Each ceiling is a power supply times an efficiency: Earth’s grid (about 3.5 × 1012 W) at today’s best 7 × 1010 FLOPS per watt; 1% of the sunlight reaching Earth, or the Sun’s whole output, at Landauer’s limit (1 / 2.87 × 10−21 operations per joule); Earth’s mass at Lloyd’s bound.
- Finish lines are the sandbox presets, measured by the same engine as the rest of the site. The method lists every constant.
- An operation, a FLOP and an erased bit are counted as one unit, as elsewhere on this site. This overstates the ceilings by a factor of about 103 to 104.
- TOP500 and Green500 lists (top500.org), June 1993 to June 2026.
- Hilbert & López, Science 332 (2011): the world’s general-purpose computing, 1986 to 2007.
- Koomey, Berard, Sanchez & Wong, IEEE Annals of the History of Computing (2011): computations per kilowatt-hour.
- Epoch AI, data on frontier AI models; Sevilla et al. (2022), “Compute Trends Across Three Eras of Machine Learning”; Sevilla & Roldán (2024) on 4 to 5 times a year.
- Hernandez & Brown (2020): 44 times less compute for the same image-recognition result, 2012 to 2019. Ho et al. (2024): language models need half the compute for the same result about every 8 months.
- Masanet et al., Science 367 (2020); International Energy Agency, “Energy and AI” (2025).
- White et al. (1986); Winding et al. (2023); Dorkenwald et al. (2024); Shapson-Coe et al. (2024); the MICrONS consortium (2025).
- Bourget & Chalmers, the 2020 PhilPapers survey.
- Interest data fetched 2026-09-28 from Wikimedia pageviews (English Wikipedia, human traffic) and OpenAlex (title and abstract phrase search).
Link format
The whole state of this page lives in its link, so a link can be written by hand, by a script or by an agent. Parameters for /countdown:
g: annual growth in computing, in percent. integer5to100c: physical ceiling the growth levels off at.grid,sunlight,sun,earth-mass
Example: https://simulation-hypothesis.com/countdown?g=35&c=sunlight. The same for every tool, and the site’s open data, at /llms.txt.