Compute Operations

Running GPU fleets · Utilization, prices and depreciation · Flagship: Meta research clusters

A GPU fleet earns its keep only when it is busy. Meta keeps its research clusters about 85% utilized, H100 rental prices rose nearly 40% in late 2025 after two years of falling, and companies now depreciate GPUs over four to six years, a period critics call too long.

The H100 one-year rental index rose from $1.70 an hour in October 2025 to $2.35 in March 2026, reversing the fall; on-demand list prices range from about $1.40 on marketplaces to over $12 on Azure; and companies depreciate GPUs over 4 to 6 years while critic Michael Burry argues 2 to 3 H100 PRICE PER GPU-HOUR, MID-2026 LIST AND INDEX $0 $4 $8 $12 1-yr contract indexOct 2025 low$1.70 1-yr contract indexMar 2026$2.35 MarketplacesVast.ai and others$1.45 Nebius, LambdaOn-demand$3.90 AWS p5On-demand$6.88 Azure ND H100 v5On-demand$12.29 GPU DEPRECIATION LIFE, YEARS Microsoft, Alphabet, CoreWeave6 Meta5.5 Amazon (some servers)5 Nebius4 Michael Burry's view2-3
Meta utilization
~83-85%
H100 index, Mar 2026
$2.35/hr
GB200, CoreWeave
~$10.50/GPU-hr
Depreciation
4-6 years
Hub
Operations

What it is

Keeping Expensive GPUs Busy

Compute operations runs the GPU fleet itself: scheduling jobs, keeping utilization high, pricing capacity and deciding how long hardware is worth keeping.

  • Utilization: Meta's research clusters stay about 83-85% busy; large training jobs use only 38-55% of the chips' peak math, with the rest lost to communication and memory waits.
  • Prices reversed: the H100 one-year rental index fell to $1.70 an hour in October 2025, then rose nearly 40% to $2.35 by March 2026 as inference demand grew. In China, Alibaba Cloud raised GPU prices 25-34% in April 2026.
  • Idle chips: in November 2025 Microsoft's Satya Nadella said some GPUs sat in inventory because there was no power to install them.
  • Depreciation: Microsoft, Alphabet and CoreWeave depreciate servers over six years, Meta over 5.5 and Nebius over four. Michael Burry argues the real life is two to three years.

Flagship

Meta's Research Clusters: What a Busy Fleet Looks Like

  • Data: 11 months of records from Meta's RSC-1 (16,000 GPUs) and RSC-2 (8,000 GPUs), published in 2024.
  • Utilization: about 83% and 85%, running 7,200 and 4,400 jobs a day.
  • Big jobs: jobs of 4,000 GPUs are under 1% of jobs but use 12% of GPU time.
  • Losses: preemption, scattered free capacity and hardware failures cost the most useful work. See Hardware Fleet Management.

Data

Fleet Utilization, Prices and Depreciation

GPU fleet operations data
ItemFigureStatus
Meta research clusters (RSC-1, RSC-2)~83-85% utilization over 11 months; 4,000-GPU jobs are under 1% of jobs but 12% of GPU timeMeasured
ByteDance MegaScale55.2% of peak math throughput on 12,288 GPUsMeasured, 2024
Meta Llama 3 405B38-43% of peak math throughput on 16,000 H100sMeasured, 2024
H100 one-year contract index$1.70 (Oct 2025) to $2.35 an hour (Mar 2026)Index
GB200 NVL72 per GPU-hour~$10.50 CoreWeave; ~$16 Oracle; ~$27 AzureList prices
China cloud pricesAlibaba Cloud GPU lines up 25-34%; Tencent up 5%; H100 rentals up 20-30%April-May 2026
Depreciation livesMicrosoft, Alphabet, CoreWeave 6 years; Meta 5.5; Nebius 4Filings
Burry critiqueReal GPU life 2-3 years; depreciation understated by $176B over 2026-2028Claim

Record

Dataset Record

Dataset
GPU fleet utilization, rental prices and depreciation
Sources
Meta (arXiv 2410.21680, Llama 3 paper), ByteDance (arXiv 2402.15627), SemiAnalysis, Google, AWS, IntuitionLabs, TrendForce, company filings, CoreWeave
Last verified
October 2026