Home โ€บ Blog โ€บ Four ways to get an AI factory
AI Factory as a Service ยท Part 1 ยท The landscape

Build, Install, Buy or Rent: The Four Ways to Get an AI Factory

Only a handful of companies can build a Colossus. Everyone else, from governments and banks to drug makers and telcos, still wants AI factories of their own, often inside their own walls. A new market has grown up to serve them: hyperscalers that install their cloud in your data center, hardware vendors that sell "AI factories" off the shelf, and neoclouds and marketplaces that rent capacity by the hour. This first part maps the four routes and what drives the choice between them.

FL
FrontierAGI Team
โš ๏ธ How this was researched (as of 26 Oct 2026) This market is dominated by vendor marketing, so we label vendor claims as such. Sources are Amazon's announcement of AWS AI Factories and DCD's coverage; NVIDIA's announcements of DGX Cloud Lepton; Dell's press releases and earnings coverage; Gartner's sovereign cloud forecast; an IDC survey; and vendor-sponsored surveys from Cloudian and Cloudera. We found no published customer counts for HPE's or Lenovo's AI factory offerings. See Reading the numbers.

1. Why this series

The AI Factory Case Studies followed the giants: xAI, OpenAI, Microsoft, Amazon, Meta and Google, each spending tens of billions of dollars on campuses measured in gigawatts. Most organisations that want AI compute look nothing like them. A ministry may be forbidden to send citizens' data to a foreign cloud. A bank may want predictable costs and full control of its models. A hospital may need answers in milliseconds next to its scanners. A telco may want to sell AI capacity to its own country. They need AI factories measured in megawatts, not gigawatts, and they usually do not want to design one from scratch.

2. Four routes

The four ways to get an AI factory

Tap a route for how it works, who sells it and who buys it.

3. Timeline

The managed AI factory market, 2023 โ†’ 2026

Solid dots are confirmed by a company, hollow dots come from the press, โš ๏ธ marks claims we could not confirm and dashed dots are planned. Click a dot for its source.

Click a milestone to see where it comes from.

4. Who owns what

The same stack, four ways

Pick a route. Green layers belong to the customer; grey layers belong to the provider. Simplified: real contracts vary.

5. What drives the choice

SovereigntyGartner: $80B of sovereign cloud IaaS spending in 2026, +35.6%
Cost & utilisationOwning pays only if the GPUs stay busy
SkillsFew organisations can run a GPU cluster
SpeedRenting is fastest; building is slowest

Sovereignty is the newest and fastest-growing driver: governments and regulated industries increasingly require that data, and sometimes the models, stay within national borders and under local control. Gartner forecasts worldwide sovereign cloud infrastructure-as-a-service spending of $80 billion in 2026, up 35.6% from 2025. Cost cuts both ways: a GPU you own is cheaper per hour than one you rent only if you keep it busy, which is the same arithmetic as CoreWeave's loans in Case Study 7. Skills are scarce: the Who Runs the AI Factory series shows how many specialist roles a cluster needs. Speed favours renting, which can start in days, over buying, which takes months, and building, which takes years.

6. Decision tree

Which route fits?

An illustration, not advice. Answer five questions; the tool scores each route and explains why.

7. The repatriation debate

Vendors selling on-premises AI like to say that companies are moving AI workloads out of the public cloud. The surveys they cite are real but come from interested parties. An independent IDC survey of 1,015 decision-makers paints a more mixed picture: most say their AI workloads need a hybrid of on-premises and public cloud.

What the surveys say

Each figure is labelled with who ran the survey. They ask different questions and are not directly comparable.

โš–๏ธ Read surveys by sponsorA storage company's survey finding that most firms are moving AI on-premises may well be accurate, but it is also good for storage sales. The safest reading across all of them is the IDC one: most organisations will use more than one route at once.

8. Market map

Who sells what

Selected providers by route. Tap a card. Customer counts are vendor claims where shown.

9. The series

The takeawayThe AI factory is becoming a product as well as a project. The same four routes recur everywhere: build it, have a cloud install it, buy it from a vendor, or rent it. Most large organisations will mix them, and the deciding questions are about data rules, utilisation, skills and time, not just price. Part 2 looks at the newest and most surprising route: hyperscalers putting their cloud inside the customer's own building.

10. Reading the numbers

AWS
AWS AI Factories launched at re:Invent (Dec 2025); NVIDIA GPUs or Trainium installed in the customer's data center, customer provides space, power and connectivity; "around $200 million over five years" quoted by an AWS executive; Humain "AI Zone" with up to 150,000 chips as the named partner. We found no other named customers.
NVIDIA
DGX Cloud Lepton announced May 2025 with ten NVIDIA Cloud Partners; AWS, Azure and European providers added at GTC Paris (Jun 2025). Lepton routes customers to partners and owns no hardware.
Dell (vendor claims)
AI Factory customers: 3,000+ (Jan 2026), 4,000 (Feb), 5,000 (May), 6,500+ (Q2 FY2027). Dell does not publish how it defines a customer.
Market
$80B sovereign cloud IaaS in 2026 (Gartner). Repatriation figures: 79% (Cloudian, storage vendor, 203 IT leaders), 66% (Cloudera, as reported); IDC (1,015 decision-makers) finds most want hybrid.
Illustrations
The decision tree, ownership chart and route descriptions are simplified.