Enterprise Data Centers

Company-owned data centers · On-premises AI · Flagship: Eli Lilly's LillyPod, Indianapolis

Company-owned data centers are losing share to the cloud and colocation, falling to 44% of enterprise workloads in 2026. AI is pulling some work back: Dell's AI Factory passed 6,500 customers, and Eli Lilly runs 1,016 Blackwell Ultra GPUs on its own Indianapolis site.

Share of enterprise IT workloads in company-owned data centers, by Uptime Institute surveys: 58% in 2020, 48% in 2023, 44% in 2026 and a projected 42% in 2028; and Dell AI Factory customers rising from 4,000 in February 2026 to more than 6,500 in September as companies build AI on their own premises WORKLOADS IN COMPANY-OWNED DATA CENTERS (UPTIME) 0% 20% 40% 60% 58%2020 48%2023 44%2026 42%2028 proj. BUT COMPANIES ARE BUYING AI SYSTEMS FOR THEIR OWN SITES Dell AI Factory customers, 2026 February4,000 May5,000 September6,500+
On-prem workloads
44%
Dell AI Factory
6,500+ customers
LillyPod
1,016 GPUs
Avg rack
11+ kW
Hub
Types

What it is

Data Centers Companies Own for Themselves

Enterprise data centers are owned and run by a company for its own IT, from a server room in an office to a campus for a bank, retailer or drug maker.

  • Shrinking share: on-premises capacity fell from about 56% of the world's total around 2019 to 34% in early 2025, and Synergy Research forecasts 22% by 2030.
  • Workloads: in Uptime Institute's 2026 survey, company-owned sites ran 44% of enterprise workloads against 46% in third-party venues, the first year outside venues led.
  • Density: average racks passed 11 kW in 2026, and nearly a quarter of operators run some racks above 30 kW, mostly for AI.

On-premises AI

Companies Are Building AI Systems on Their Own Sites

  • Dell: its AI Factory line grew from 4,000 customers in February 2026 to more than 6,500 by September, with a $51.3 billion AI backlog and about $74 billion of AI server sales guided for its fiscal 2027.
  • HPE: Private Cloud AI, which HPE says cuts token cost by up to 60% against public cloud; HPE has not published a customer count.
  • Why on-premises: data that cannot leave the building, from patient records to classified work, and steady workloads that cost less to own than to rent. Lockheed Martin keeps its core AI system on-premises and sends unclassified overflow to the cloud.

Flagship

Eli Lilly's LillyPod in Indianapolis

  • System: 1,016 Nvidia Blackwell Ultra GPUs in DGX B300 systems, the first DGX SuperPOD built on DGX B300, liquid cooled, on Spectrum-X networking.
  • Power: more than 9,000 petaflops of AI performance, owned and run by Lilly, inaugurated February 2026.
  • Work: protein design, genomics models and molecule screening; Lilly shares some models with biotech companies through its TuneLab service.

Deployments

Enterprise AI Systems on Company Sites

Enterprise on-premises AI deployments
CompanySiteSystemSizeDateStatus
Eli LillyIndianapolis, INLillyPod: DGX SuperPOD on DGX B3001,016 Blackwell Ultra GPUs; 9+ EF AIFebruary 2026Operating
Mayo ClinicRochester, MNDGX SuperPOD on DGX B200Not disclosedJuly 2025Operating
Lockheed MartinInternal and classified sitesDGX SuperPOD AI factoryNot disclosedOctober 2025Operating
Hyundai Motor GroupKoreaNvidia Blackwell AI factory50,000 GPUs, ~$3BOctober 2025Announced
Dell AI Factory customersUS and worldwidePowerEdge with Nvidia6,500+ customersSeptember 2026Shipping

Record

Dataset Record

Dataset
Enterprise data center share and on-premises AI deployments
Sources
Synergy Research, Uptime Institute Global Data Center Survey 2026, Dell, HPE, Nvidia, Eli Lilly, Mayo Clinic, Lockheed Martin, Hyundai Motor Group
Last verified
October 2026