Who Buys an AI Factory? Telcos, Drug Makers, Banks, Hospitals and Governments
Parts 2 to 4 followed the sellers. This part follows the buyers: a telco turning exchanges into its country's AI cloud, a drug maker training models on proteins, a bank keeping its data in-house, a hospital speeding up pathology, a national lab buying its next supercomputer. Each case file answers the same questions: what they bought, from whom, why they did not just rent, and what they run.
1. The customer map
2. Timeline
Announced customer deployments, 2024 โ 2027
One lane per sector. Solid dots are confirmed by the customer or vendor, hollow dots come from the press, dashed dots are planned. Click a dot for its source.
3. Case files
Named customers
Filter by sector. Tap a card for detail.
4. Telcos: becoming the national AI cloud
Telcos are the most visible new buyers. They have what AI factories need, buildings, power connections, fiber and a national customer base, and a reason to want new revenue. NVIDIA counts 18 telcos on five continents building sovereign AI infrastructure on its platform, naming SoftBank and KDDI in Japan, Indosat Ooredoo Hutchison in Indonesia, Singtel's RE:AI, Swisscom and Telenor. At GTC Berlin in October 2026, it said about ten operators in Europe, the Middle East and Africa had signed on. SK Telecom plans a gigawatt-scale AI cloud in Korea on NVIDIA's DSX design, with its first facility due in 2027. Vodafone, from Part 2, runs six Oracle Dedicated Regions for its own IT.
Will a telco AI factory pay back?
An illustration of the critique raised by Fierce Network: GPUs lose value as new generations arrive. Set the investment's usable life and how much of the capacity the telco sells.
5. Pharma and health
Drug discovery is one of the clearest cases for owning compute: the data is proprietary, the models (proteins, molecules, genomes) are specialised, and the work runs continuously. Eli Lilly's LillyPod (Part 3) is a DGX SuperPOD of 1,016 Blackwell Ultra GPUs, assembled in about four months and live from February 2026. In Denmark, the Novo Nordisk Foundation funded Gefion, a DGX SuperPOD of 1,528 H100 GPUs operated by the Danish Centre for AI Innovation and hosted by Digital Realty, which went live in October 2024 with about DKK 600 million of initial funding and is shared with Danish researchers, startups and public bodies. Mayo Clinic announced a DGX SuperPOD of B200 systems in July 2025 for pathology, drug discovery and precision medicine, and says it cut the time to train some pathology models from four weeks to one.
6. Manufacturing
Manufacturers mostly buy AI for their factories rather than an AI factory: digital twins, robotics and production planning that run on GPUs in the plant or the cloud. Siemens and NVIDIA are building what they call an Industrial AI Operating System, with Siemens's electronics factory in Erlangen, Germany, as the first reference site from 2026; a report in October 2026 said engineering effort there fell by 85% โ ๏ธ (company-reported). Foxconn is using Siemens and NVIDIA tools for a robotic plant that builds NVIDIA's own AI hardware, HD Hyundai says it cut design iterations from days to hours, and KION and PepsiCo are evaluating the technology.
7. Governments and banks
Governments buy for science and sovereignty. The US Department of Energy's next NERSC flagship, Doudna, is a Dell system on NVIDIA's Vera Rubin platform due for delivery in late 2026, promising more than ten times the performance of its predecessor. Singapore's CSIT, GovTech and HTX were among the first users of Gemini on Google's air-gapped cloud (Part 2); NxtGen in India built what Dell called India's largest AI factory; and Humain is building an AI Zone with AWS. Banks are the quietest buyers. BNY announced in March 2024 that it was the first major bank to deploy a DGX SuperPOD, built with dozens of DGX H100 systems, for uses such as deposit forecasting, payment automation and trade analytics. We found no more recent announced bank deployment, which likely reflects secrecy rather than absence.
8. Why buy, not rent
Reasons by sector
Our editorial reading of why each sector owns compute, from the case files. Darker = stronger reason. Tap a cell.
9. What we can't see
Dell alone claims more than 6,500 AI Factory customers (Part 4). This article names about twenty organisations. The gap is the point: most buyers, especially banks, defence and intelligence agencies and many manufacturers, never announce what they run, and vendors rarely name them without permission. Treat any list of customers, including this one, as the visible tip.
Named vs claimed
10. Reading the numbers
- Telcos
- 18 telcos on five continents (NVIDIA developer blog, date uncertain); ~10 EMEA operators at GTC Berlin (Oct 2026, SDxCentral); SK Telecom gigawatt-scale plan, first facility 2027 (NVIDIA); Vodafone six Oracle regions.
- Pharma & health
- LillyPod 1,016 B300 GPUs, live Feb 2026; Gefion 1,528 H100 GPUs (191 DGX H100), live 23 Oct 2024, ~DKK 600M; Mayo Clinic DGX SuperPOD with B200, Jul 2025, "four weeks to one" (Mayo).
- Manufacturing
- SiemensโNVIDIA Industrial AI OS, Erlangen from 2026; 85% engineering reduction reported โ ๏ธ; Foxconn, HD Hyundai; KION and PepsiCo evaluating.
- Government & finance
- Doudna (Dell, Vera Rubin) due late 2026, >10ร Perlmutter (DOE); Singapore agencies on GDC; NxtGen (Dell); BNY DGX SuperPOD (Mar 2024).
- Illustrations
- The payback model and heatmap are simplified and editorial.
- Telcos
- NVIDIA: telcos across five continents
- SDxCentral: Europe's telcos as sovereign AI factories
- NVIDIA: SK Telecom AI infrastructure
- Fierce Network: telco AI factories could backfire
- Pharma and health
- DCD: Lilly DGX B300 SuperPOD
- DCAI: Gefion operational
- DCD: Novo Nordisk Foundation SuperPOD
- HLTH: Mayo Clinic deploys NVIDIA infrastructure
- Manufacturing
- NVIDIA: Siemens Industrial AI Operating System
- TechTimes: Erlangen engineering down 85%
- Government and finance
- DOE: Doudna supercomputer
- NERSC: Doudna system
- BNY: first global bank with a DGX SuperPOD
- HPCwire: NxtGen and Dell
- Part 4: the hardware vendors