Eighteen years of silicon

Each band is one slice of the world's compute (zettaFLOP/s, peak dense BF16), 2008 → present — the top edge is the headline. Bands are stacked slowest-growing at the bottom, so the saturating consumer slices form the floor and the AI-accelerator bands visibly erupt through the top after 2023. Click any component below to drill in. History before the 2026 anchor is reconstructed from installed-base and shipment series — see the methodology.

Where the compute is

Compute does not sit where wealth sits. Datacenter silicon is extraordinarily concentrated — the United States holds roughly seventy percent of the world's AI accelerators — while personal compute follows people, which puts China first on phones. The map blends both keys by their real weight in the headline. Hover any country; click to drill in.

View: total compute · per person

Top 20 of 215 countries. Datacenter compute is apportioned from published AI-compute shares, personal compute by population weighted for device intensity — both documented in the methodology. Full list: JSON →.

Three kinds of silicon

The balance between them has inverted in about four years. Purpose-built AI accelerators were a rounding error in 2020. They are now the largest block of computing power on the planet — and the year-over-year figures below are the whole story: one of these three is compounding, the other two are on a replacement cycle.

• AI accelerators 31.40 ZFLOP/s

  39.2% of the total · +171% in a year · Datacenter silicon built for tensors — NVIDIA, TPUs, Trainium, Instinct and the rest.

• Personal & consumer 47.88 ZFLOP/s

  59.8% of the total · +16% in a year · Phones, gaming GPUs, PCs and consoles — the compute people own but almost never use.

• General-purpose & embedded 790.46 EFLOP/s

  1.0% of the total · +22% in a year · Server CPUs and the chips bolted to cars, cameras and robots.

What's counted

• NVIDIA datacenter GPUs → 17.06 ZFLOP/s

  21.3% of the total · 17.2M H100e · +141% in a year

• Google TPUs → 6.96 ZFLOP/s

  8.7% of the total · 7.0M H100e · +216% in a year

• AWS Trainium & Inferentia → 3.28 ZFLOP/s

  4.1% of the total · 3.3M H100e · +257% in a year

• AMD Instinct → 1.63 ZFLOP/s

  2.0% of the total · 1.7M H100e · +181% in a year

• Other AI silicon → 2.47 ZFLOP/s

  3.1% of the total · 2.5M H100e · +214% in a year

• Smartphones → 25.71 ZFLOP/s

  32.1% of the total · 26.0M H100e · +19% in a year

See all 11 constituents — and the five things we show but never add →

How big is that?

A zettaFLOP is a sextillion operations a second, which is not a quantity anyone has intuition for. Set against Earth's 80.08 ZFLOP/s:

• One NVIDIA H100 989.40 TFLOP/s

  Earth is 80,934,245× bigger. The unit the whole industry counts in. 989 teraFLOP/s of dense BF16, 700 watts.

• A flagship smartphone 5.00 TFLOP/s

  Earth is 16,015,268,329× bigger. GPU plus neural engine in something you carry. It would have topped the TOP500 in 2001.

• The fastest supercomputer 2.20 EFLOP/s

  Earth is 36,431× bigger. LineShine, 2.198 exaFLOP/s on HPL — the fastest machine ever ranked, at double precision.

• Every TOP500 machine combined 15.00 EFLOP/s

  Earth is 5,338× bigger. All five hundred ranked supercomputers on Earth, added together, at FP64.

• ASCI Red, 1997 1.30 TFLOP/s

  Earth is 61,597,185,881× bigger. The first computer to break one teraFLOP/s. It filled 149 square metres and drew 850 kW.

• The Apollo Guidance Computer 14,000 FLOP/s

  Earth is 5,719,738,688,909,673,472× bigger. The machine that landed people on the Moon: roughly fourteen thousand operations a second.

The whole world, once

Training a frontier model takes a fixed number of operations. If every processor on Earth stopped what it was doing and worked on one job, here is how long each would take — at nameplate peak, and at the delivered rate the world's silicon actually sustains.

Wildly hypothetical — you cannot shard a training run across seven billion phones, and the interconnect to try would not exist. It is here for scale: the compute that took OpenAI months to assemble is, in aggregate, about 31.1 minutes of the planet's output.

What we refuse to count

The fastest way to get this number badly wrong is to add up everything that switches. Five things are shown here and deliberately excluded from the headline — together they come to 10.69 YFLOP/s, about 133x the figure at the top of this page.

• Marketing headroom → 119.43 ZFLOP/s

  1.5x the headline — and not part of it.

• Announced, not built → 54.10 ZFLOP/s

  0.7x the headline — and not part of it.

• Bitcoin mining ASICs → 2.27 YFLOP/s

  28.4x the headline — and not part of it.

• TOP500 supercomputers → 19.07 EFLOP/s

  0.0x the headline — and not part of it.

• Human brains → 8.24 YFLOP/s

  103x the headline — and not part of it.

Bitcoin's mining fleet is the largest computer humanity has ever built and cannot execute a single floating-point operation. Eight billion human brains still out-compute every machine on Earth by 103x. Why each one is excluded →

How it's built