Data Center Workloads

What runs in data centers · From microsecond trades to month-long training runs

Data center workloads range from trades matched in microseconds to AI training runs that last months. AI is now the largest and fastest-growing of them: inference is forecast to be two-thirds of AI compute in 2026, and training loads swing so hard that grid operators now regulate them.

Approximate positions of data center workloads by how fast they must respond and how much power a rack draws: exchange matching engines respond in microseconds on low-power racks, interactive AI inference in under a second on racks over 100 kW, and AI training runs for weeks on the densest racks HOW FAST IT MUST RESPOND vs. RACK POWER (APPROXIMATE) 1 µs 1 ms 1 s 1 hour 1 month 10 kW 30 kW 100 kW 300 kW Exchange matching Telecom and CDN Cloud and enterprise AI inference Analytics HPC simulation Storage archive AI training Response time, log scale. Positions are approximate, for comparison only.
Workload pages
10
Inference share, 2026
~2/3 (fcst)
Cloud growth, Q2 2026
43%
Fastest response
Microseconds
Hub
Workloads

Workloads

Ten Workloads, One Flagship Each

Workload pages describe what runs in data centers; Types pages describe the buildings it runs in.

Data center workloads and flagships
PageFlagshipCovers
AI TrainingColossus 1 Megapacks, MemphisPower swings and grid rules
AI InferenceCerebras, Oklahoma CityTokens, latency and custom chips
Cloud and EnterpriseAWS us-east-1, Northern VirginiaCloud market and Arm CPUs
Big Data and AnalyticsDatabricksWarehouses, streaming, AI retrieval
Storage and ArchivalSeagate HAMR drivesDrives, flash, tape and the shortage
Telecom and ContentEquinix AshburnCDNs, interconnection, 5G core
Government and DefenseAWS classified regionsClassified clouds and DOE land
Financial ServicesCME Group, Aurora, ILMatching engines and trading GPUs
HealthcareEpic hosting, Verona, WIHealth records and clinical AI
Energy and UtilitiesERCOT, Taylor and Bastrop, TXGrid operations and seismic HPC

Shift

AI Is Now the Largest and Fastest-Growing Workload

  • Inference: about half of AI compute in 2025 and a forecast two-thirds in 2026 (Deloitte); Google processed 3.2 quadrillion tokens in May 2026.
  • Training: single jobs now span hundreds of thousands of GPUs and swing building power by tens of megawatts, enough for grid operators to write new rules.
  • Everything else: cloud infrastructure still grew 43% in Q2 2026, much of it pulled by AI, and AI's demand for drives and flash has created a storage shortage.

Sectors

Sectors Add Their Own Rules

  • Government: classification levels decide where work can run.
  • Finance: microsecond latency and resilience rules such as the EU's DORA.
  • Healthcare: HIPAA and records that cannot go down.
  • Energy: grid control rooms with fully redundant sites.

Record

Dataset Record

Dataset
Index of data center workloads with flagships; data on each linked page
Sources
Listed on each Workloads page
Last verified
October 2026