A scale for the species
The AGI Scale
Like Richter for quakes, or Kardashev for energy: a civilization is typed by the calculation it can hold. The zero of this scale is not a laboratory. It is humanity’s entire pile of thought — every AI chip Earth has plugged in — as of September 2026.
The unit is installed AI compute, the physical stock that is on and usable, not a test score and not a job title. Type zero is the species as it stands now: a convention called C0. Each integer step is ten times that human pile, the same logarithmic move as Kardashev’s energy types. After AGI-4, this scale and the Kardashev energy scale become the same object: a planet, then a star, spent as thought.
Fig. 1 — the human pile
Humanity stands at the bottom. Each badge on the rail is how much the species’ pile must grow. Click a type: a dispatch opens — one civilizational line for that rung. Follow compute and power up: Earth’s stock becomes a planet, then a star.
A star, run as a clock and a furnace for thought. Earth, if it is remembered, is a rounding error of mass. This is Kardashev Type II: a sun that thinks.
The planet’s energy budget, captured. Weather is waste heat. Oceans are the cooling loop. Kardashev Type I is not an empire. It is Earth, thinking.
Steel, copper, dams, and deserts now exist to grow the pile. Civilization is no longer using compute. Civilization is compute.
Every watt the species generates — mills, metros, midnight — poured into calculation. Earth’s workday becomes one cognition.
A large nation switches its grid from light to mind. Houses dim. Something the size of a country begins to think.
This is not a company and not a model. This is the species: eight billion people, and a pile of silicon already drawing a city’s power, humming. The first time humanity can be typed.
Humanity’s first three climbs are ×10. Then ×1,000 to a planet’s energy. Then ×10 billion to a star. All power in watts at 2026 chip efficiency. Dates, if the 2022–26 slope survives the grid: AGI-1 late 2028 · AGI-2 ~2031 · AGI-3 ~2033–34.
Key
Selected type
- Multiple of now
- 1×
- Compute
- 6×1022 FLOP/s
- Power (W, 2026 efficiency)
- ~3×1010 W
What the species has plugged in, Sept 2026.
Select a rung, a symbol, or a comparison.
Fig. 2 The six types
| Type | G | Multiple of now | Compute | Power (W, 2026 efficiency) | Meaning |
|---|---|---|---|---|---|
| AGI-0 | 0 | 1× | 6×1022 FLOP/s | ~3×1010 W | What humanity has plugged in, Sept 2026 |
| AGI-1 | 1 | 10× | 6×1023 FLOP/s | ~3×1011 W | A large nation’s entire grid, spent on thought |
| AGI-2 | 2 | 100× | 6×1024 FLOP/s | ~3×1012 W | All electricity humanity generates today, spent on thought |
| AGI-3 | 3 | 1,000× | 6×1025 FLOP/s | ~3×1013 W | Industrial civilization rebuilt around compute |
| AGI-4 | 6 | 1,000,000× | 6×1028 FLOP/s | ~3×1016 W | Kardashev Type I: a planet’s energy budget as cognition |
| AGI-5 | 16 | 1016× | 6×1038 FLOP/s | ~3×1026 W | Kardashev Type II: a star as a computer |
Note How to read this
- Stock, not a demo. C is installed usable AI compute.
- One model is not a type. Astra, or any one model, is a program inside AGI-0.
- Efficiency is a footnote. If chips get more FLOP/J, the same watts buy a higher G; the unit does not change.
- Dates assume the current doubling. At a ~7-month doubling, G rises ~0.43/year: AGI-1 ~late 2028, AGI-2 ~2031, AGI-3 ~2033–34, if the 2022–26 slope survives the grid.
Def. How the type is written
Every real scale needs a zero and a unit. This one is a logarithm, because civilizations do not grow by addition — they grow by decades. G counts how many decades of the human pile have been assembled. C is the installed thought of the civilization being graded. C0 is us, in September 2026, before the grid, the planet, and the star.
= log10( / )
C0 = 3×107 H100-equivalents ≈ 6×1022 FLOP/s (FP8, Epoch-style H100e) ≈ 3×1010 W chip load / ~4×1010 W at the meter.
Reference epoch: September 2026. C0 is a stated convention for the human pile, not a sacred constant. Update C0 when the stock is remeasured. Power throughout is watts at 2026 efficiency.