What's counted
The world's compute, split into slices that do not overlap and sum to the headline exactly once. Every figure is peak dense BF16 FLOP/s: where a vendor quotes FP4, FP8 or INT8 "AI TOPS" with 2:1 sparsity, it has been converted down. Each row states how its device count and per-device figure were derived, and what it deliberately excludes.
Headline constituents — summed to 80.08 zettaFLOP/s
-
NVIDIA datacenter GPUs → · accreted 17.06 ZFLOP/s
21.3% of the total · 17.2M H100e · +141% in a year · ai accelerator · ~34.0% utilized
Every A100, H100, H200, B200 and GB200 racked and powered on Earth — the single largest block of purpose-built AI compute, and about 62% of the world's AI accelerator capacity. Quoted at dense BF16: a B200's headline '20 PFLOPS' is FP4 with sparsity, which is 2.25 PFLOPS here.
How it's counted: ~9.3M H100-equivalents. Blackwell (B200/B300/GB200) now carries the majority; Hopper is the long tail.
Datacenter parts only. GeForce cards sold into gaming PCs are the separate consumer line, and the TOP500 machines built from these GPUs are the not-summed overlay.
-
Google TPUs → · accreted 6.96 ZFLOP/s
8.7% of the total · 7.0M H100e · +216% in a year · ai accelerator · ~36.0% utilized
The largest fleet of custom AI silicon ever built, and the reason Epoch AI puts Google ahead of every other single owner of compute. TPUs are deployed only inside Google — none are sold — so this slice is a private supply chain from wafer to workload.
How it's counted: ~2.7M H100-equivalents across TPU v5e, v5p, v6e (Trillium) and v7.
Google's NVIDIA GPUs sit in the NVIDIA line; this is TPU silicon only, so the two do not overlap.
-
AWS Trainium & Inferentia → · accreted 3.28 ZFLOP/s
4.1% of the total · 3.3M H100e · +257% in a year · ai accelerator · ~30.0% utilized
Amazon's answer to buying NVIDIA — in-house accelerators deployed at hyperscale across AWS regions, and the fastest-compounding vendor slice in the partition. Project Rainier alone is one of the largest single AI clusters on the planet.
How it's counted: ~1.2M H100-equivalents, dominated by Trainium2 and the Trainium3 ramp.
Amazon's very large NVIDIA fleet is counted in the NVIDIA line; this is Annapurna-designed silicon only.
-
AMD Instinct → · accreted 1.63 ZFLOP/s
2.0% of the total · 1.7M H100e · +181% in a year · ai accelerator · ~28.0% utilized
The only merchant-market alternative to NVIDIA at scale. Strong on memory capacity per package, which is why it lands disproportionately in inference fleets, and the anchor of the US Department of Energy's exascale machines.
How it's counted: ~750k H100-equivalents across MI300X, MI325X and MI350X.
Instinct datacenter parts only; Radeon gaming cards are in the consumer GPU line.
-
Other AI silicon → · accreted 2.47 ZFLOP/s
3.1% of the total · 2.5M H100e · +214% in a year · ai accelerator · ~25.0% utilized
Everyone else building accelerators: Huawei's Ascend line serving a market NVIDIA cannot legally sell into, Meta's MTIA and Microsoft's Maia serving their own datacenters, and the wafer-scale and deterministic-inference startups. Individually small, collectively larger than AMD.
How it's counted: ~1.1M H100-equivalents: Huawei Ascend, Meta MTIA, Microsoft Maia, Cerebras, Groq, Tenstorrent.
Defined as the residual of Epoch's fleet total after the four named vendors, so the five AI lines sum to the fleet exactly once.
-
Smartphones → · accreted 25.71 ZFLOP/s
32.1% of the total · 26.0M H100e · +19% in a year · personal compute · ~0.5% utilized
The largest single slice of the partition, and nobody planned it. Seven billion phones, each with a GPU and a neural engine, quietly out-compute every datacenter on Earth put together. Almost none of it is ever used for anything harder than rendering a feed.
How it's counted: 7.21 billion ACTIVE handsets (Ericsson; Counterpoint puts 2025 growth at 2 percent, with replacement cycles near four years). Per-device is a tier model over the installed base, not over new sales — the global median handset is several years old. A current flagship is ~25 TFLOP/s counting GPU and neural engine; a five-year-old budget phone is ~0.15. Crediting the NPU is a convention, so the low end of the range for this slice counts the GPU alone.
Handset SoCs only. Tablets and laptops are the personal-computer line; the cell networks behind them are not compute.
-
Consumer & gaming GPUs → · accreted 16.14 ZFLOP/s
20.2% of the total · 16.3M H100e · +14% in a year · personal compute · ~3.0% utilized
GeForce, Radeon and Arc cards sitting in gaming PCs — the same tensor cores as the datacenter parts, in smaller quantity, distributed across a third of a billion desks. This is the pool that mined Ethereum, and it remains the world's largest reserve of idle general-purpose parallel compute.
How it's counted: ~250 million discrete cards: desktop add-in-board installed base ~180M (Jon Peddie Research) plus laptop discrete. Per-device is a WEIGHTED AVERAGE over the 85 discrete models in the July 2026 Steam Hardware Survey at dense FP16 — 74.2 TFLOP/s — haircut to 60 for the non-gaming tail Steam under-samples and for its old-skewed 8.7 percent Other bucket. The most common card, an RTX 3060, is 51 TFLOP/s; an RTX 4090 is 330.
Discrete add-in cards in client machines. Datacenter GPUs are the AI-accelerator lines; integrated graphics is in the personal-computer line.
-
PCs & laptops → · accreted 3.96 ZFLOP/s
4.9% of the total · 4.0M H100e · +10% in a year · personal compute · ~2.0% utilized
Every laptop and desktop in the world, counted as its CPU vector units plus integrated graphics plus, increasingly, a dedicated neural engine. An M-series MacBook now carries more usable FP16 throughput than a 2016 datacenter node.
How it's counted: ~1.55 billion active machines x ~2.4 TFLOP/s. The integrated-GPU rows in the July 2026 Steam survey average 4.4 TFLOP/s dense FP16, but Steam skews modern and the global PC base is older, so the fleet average is set below it.
Excludes discrete add-in GPUs, which are the consumer-GPU line, so a gaming PC is not counted twice.
-
Game consoles → · accreted 2.08 ZFLOP/s
2.6% of the total · 2.1M H100e · +6% in a year · personal compute · ~2.0% utilized
PlayStation, Xbox and Switch — a quarter of a billion semi-custom AMD and NVIDIA parts, each roughly a small workstation, sold at or below cost and idle most of the day. The cheapest FLOP/s per dollar anyone has ever shipped at volume.
How it's counted: Sales are public and specs are exact, so this is the tightest consumer line. PS5 94.1M x 20.6 TFLOP/s, Switch 1 155.9M x 0.79, Xbox Series 35.2M x 16.0, PS4 117.2M x 2.6, Switch 2 22.6M x 6.2, Xbox One ~58M x 1.9 — 3.18e21 sold. The active rate is measured, not guessed: Sony reports 132M PSN monthly actives against 211M PS4 and PS5 sold, i.e. 63 percent.
Console APUs are distinct silicon from the PC and consumer-GPU lines.
-
Server CPUs → · accreted 328.37 EFLOP/s
0.4% of the total · 331.9k H100e · +13% in a year · general purpose · ~35.0% utilized
Xeon, EPYC, Graviton and the rest — the silicon that actually runs the internet. Tiny in FLOP/s next to a GPU rack and indispensable anyway: databases, networking and orchestration are latency problems, not throughput problems.
How it's counted: ~120 million servers. Intel AMX and Arm SVE2 give modern sockets real matrix throughput; most of the installed base predates them.
Host CPUs in accelerated nodes are counted here; their attached GPUs are counted in the AI lines. The two are separate silicon.
-
Embedded & edge → · accreted 462.08 EFLOP/s
0.6% of the total · 467.0k H100e · +30% in a year · general purpose · ~10.0% utilized
Compute bolted to something that moves or watches: driver-assist stacks, warehouse robots, security cameras, smart TVs, drones. The fastest-growing slice outside the datacenter, because every new vehicle now ships with a small supercomputer.
How it's counted: Tesla HW4 is ~50 INT8 TOPS per car; NVIDIA Thor, Orin and Jetson carry the rest, plus billions of small NPUs.
Devices outside the datacenter and outside the personal-device lines; microcontrollers with no vector unit are excluded as immaterial.
The exclusions
Five things that look like compute, are frequently counted as compute, and are not part of this headline. Three traps in particular. The Bitcoin mining fleet is by raw switching throughput the largest computer ever built and can do exactly one thing — hash a block header — so its operations per second do not belong in the same total. Vendor peak "AI TOPS" is FP4 with a sparsity assumption, four to nine times the dense figure a chip sustains; it is a reference number, like a derivatives notional. And announced capacity is not compute: a signed order and a substation under construction compute nothing until they are energized.
-
Marketing headroom → 119.43 ZFLOP/s
+178% in a year
If you add up the numbers printed on the datasheets, the world has roughly 7.5e22 FLOP/s of AI accelerators rather than 1.5e22. The gap is precision and sparsity: FP4 instead of BF16, and a 2:1 structured-sparsity assumption that real weights rarely satisfy. It is a reference figure, not throughput — the compute equivalent of a derivatives notional.
How it's counted: A ~5x average multiplier over the AI-accelerator fleet. A B200 is quoted at 20 PFLOPS (FP4, sparse) and delivers 2.25 PFLOPS dense BF16.
The discarded portion of a number already counted once at dense BF16. Adding it would count the same transistors twice.
-
Announced, not built → 54.10 ZFLOP/s
+169% in a year
Signed orders, announced campuses, and gigawatts of substation capacity under construction. Larger than everything currently installed — and computing precisely nothing until it is energized. Power interconnects, not wafers, are the binding constraint.
How it's counted: Roughly a year of forward shipments. Five individual AI datacenters are expected to cross 1 GW during 2026.
Becomes headline compute the day it is racked and powered; counting it now would double-count it then.
-
Bitcoin mining ASICs → · live 2.27 YFLOP/s
+4% in a year
By raw switching throughput this is the largest computer humanity has ever built — roughly fifty times the world's entire floating-point capacity. It can do exactly one thing: hash a block header twice with SHA-256. It cannot run a model, solve a PDE, or add two decimals. The purest illustration of why operations per second is not compute.
How it's counted: ~900 EH/s at ~28 J/TH, drawing on the order of 150-200 TWh a year. Marked live from the Bitcoin network.
Fixed-function integer hashing, not floating point. Shown at its notional op/s and never summed — the arithmetic simply does not belong in the same total.
-
TOP500 supercomputers → 19.07 EFLOP/s
+27% in a year
The world's five hundred fastest ranked machines, measured the way science measures: FP64, sustained, on a real solver. Fifteen exaflops sounds enormous and is a thousandth of the headline — because the headline is half precision, and because these machines are built from the very GPUs already counted.
How it's counted: ~15 EFLOP/s aggregate Rmax at double precision. LineShine leads at 2.198 EFLOP/s; El Capitan and Frontier follow.
Overwhelmingly assembled from NVIDIA and AMD datacenter GPUs already in the AI-accelerator lines. Summing it would count the same racks twice.
-
Human brains → 8.24 YFLOP/s
+1% in a year
Eight billion brains at roughly a petaflop each, running on twenty watts. Still about two hundred times all the machine compute on Earth, and the only line in this table that shrinks per capita when it works harder. It is here as the scale anchor: the machines have not caught up, and the gap is closing at 3.3x a year.
How it's counted: Per-brain estimates span 1e13 to 1e17 FLOP/s. We use 1e15, the middle of Joseph Carlsmith's Open Philanthropy range.
Not machines, not silicon, and not a FLOP/s in any strict sense — a translated estimate shown for scale and never summed.