Writing
Blog
Findings from the dataset, and notes on how it is built. Every claim here links back to the rows it came from, and where a post reports a number it also reports how that number could be wrong.
-
Google Willow: what it has and has not shown
Willow cut logical error 2.14x per code-distance step, to 0.143% per cycle at distance 7. Its sampling and Quantum Echoes claims rest on classical-cost estimates, and no logical gates have been shown.
deep-divequantumbenchmarks -
Microsoft Fairwater: two sites, few numbers
Microsoft says Fairwater links Wisconsin and Atlanta with hundreds of thousands of Blackwell GPUs. Epoch AI estimates about 480,000 across the two, at 369 MW and 636 MW of IT power, but Microsoft has published no count or megawatts.
deep-diveai-clusterspower-and-coolingaccelerators -
Meta Hyperion and Prometheus: tents and gas
Meta's Prometheus in Ohio is already training models, with Epoch AI estimating 471 MW of IT power. Hyperion in Louisiana is a $27 billion Blue Owl joint venture that Meta now says will reach 5 GW, with no GPU count disclosed.
deep-diveai-clusterspower-and-coolingprocurement -
No TOP500 for quantum: what each metric is worth
Quantum computers have no shared ranking. Our table holds 75 benchmark rows for 35 machines, and all 75 are vendor-reported or peer-reviewed, none labelled as an independent measurement. Here is what each metric means and what it can't tell you.
deep-divequantumbenchmarks -
Quantinuum Helios: what the claims support
Helios has 98 barium-ion qubits at 99.921% two-qubit fidelity and a reported 48 error-corrected logical qubits. The H-series reached QV 2^25. Here is what is measured, what is postselected, and who pays.
deep-divequantumbenchmarks -
Project Rainier: Amazon's Anthropic cluster
Amazon says Project Rainier reached nearly 500,000 Trainium2 chips in October 2025; Epoch AI now estimates about 1.04 million at New Carlisle, Indiana alone. Neither figure is audited, and the Mississippi campuses are not confirmed as part of it.
deep-diveai-clustersacceleratorspower-and-cooling -
Anthropic and Google TPUs: stated vs modelled
Google and Anthropic say Anthropic can use up to one million Ironwood TPUs and well over 1 GW in 2026. No site, chip count or power figure has been disclosed, and the campus-level numbers in circulation are Epoch AI estimates.
deep-diveai-clustersacceleratorsprocurement -
Isambard-AI versus Alps: two GH200 routes
Bristol's Isambard-AI has 5,448 GH200 chips (168 in phase 1, 5,280 in phase 2) in a 5 MW modular site; CSCS's Alps has 10,752 in Lugano. Alps and phase 1 both score near 60 GFlops per watt in HPL, but the facilities differ.
deep-diveai-clusterspower-and-coolingprocurementsovereignty -
How Nebius and Nscale build GPU clouds
Nebius owns Finland and rents in Iceland and the UK; Nscale mostly leases, with Microsoft and OpenAI as anchors. Their PUE and WUE figures are targets or company statements, and only one Nscale cluster has a benchmark.
deep-diveai-clusterspower-and-coolingsovereignty -
Fugaku: the CPU-only No. 1 and its successor
Fugaku led four benchmark rankings with no accelerator and has posted the same 442 PFLOPS result since 2020. Its successor, FugakuNEXT, adds NVIDIA GPUs.
deep-divehpchistoryacceleratorspower-and-coolinginterconnect -
Frontier: the first exaflop and its price
Frontier cleared 1.102 exaflops on HPL in May 2022 at 21.1 MW, meeting the 20 MW-per-exaflop goal as a ratio, not as a power budget. The route took a missed schedule and a hunt for bad hardware.
deep-diveexascalehpcpower-and-coolingaccelerators -
LineShine: a CPU-only 2.2 exaflop machine
LineShine tops the June 2026 TOP500 at 2.198 EFlop/s with no accelerators. The headline is documented; the chip's designer, fab and memory suppliers are not.
deep-divehpcexascalesupply-chainaccelerators -
Stargate Abilene: 1.2 GW, four of eight live
Stargate Abilene is stated at 1.2 GW across eight buildings. Epoch AI counts four operating at 421 MW of IT load in July 2026, on a stack of leases, loans and on-site gas turbines.
deep-diveai-clusterspower-and-coolingacceleratorsprocurement -
xAI Colossus: fast to build, hard to count
Colossus 1 went from site start to 100,000 GPUs in a claimed 122 days. Its GPU counts, its power sources and Colossus 2's size rest mostly on executive statements and court records.
deep-diveai-clusterspower-and-coolinginterconnectsupply-chain -
Aurora: eight years from contract to first list
Aurora took eight years from a 2015 award to its first TOP500 entry, after Knights Hill was cancelled. It now holds 1.012 exaflops, and about 96% of recorded Intel Max GPUs in this dataset.
deep-diveexascalehpcacceleratorshistory -
JUPITER: Europe's exascale, American silicon
JUPITER hit 1.000 exaflop/s in November 2025 on a 500 million euro budget. The Booster's CPUs, GPUs, fabric and flash are American; the European CPU reached sample stage only in September 2026.
deep-diveexascaleprocurementsupply-chainpower-and-cooling -
El Capitan: building the NNSA exascale system
El Capitan put CPU and GPU on one package, cost about $600 million, and runs classified weapons work. Its public numbers disagree on nodes, power and peak, and the lab does not reconcile them.
deep-divehpcexascaleacceleratorsprocurement -
Same hardware, a different machine
El Capitan and xAI's Colossus 1 both run liquid-cooled accelerators from the same two silicon generations. One reports a benchmark result to the world twice a year. The other has never run one. That gap is the whole story.
explainerai-clustershpcpower-and-coolinginterconnect -
The merchant fabric tier is one company
Of the 21 interconnect edges in our dataset where the fabric came from someone other than the machine's integrator, 14 belong to NVIDIA. Two announcements this week land on exactly that segment.
analysisinterconnectnvidiacxl -
Europe's domestic substitution index is zero
Not one European system in our dataset is fully domestically supplied, even counting the EU as one home country.
analysiseuropesupply-chainsovereignty -
The fabric layer consolidated twice in a year
HPE bought Cray in September 2019, NVIDIA bought Mellanox in April 2020, and because a spec string names the product rather than the parent, almost nobody charted it.
analysisinterconnectentity-resolutionm-and-a -
What a holdout corpus told us
Our parser scored 100% against strings we wrote and 47% against strings we had not seen. Publishing the second number is why the first one means anything.
methodologyingestmeasurement