Home โ€บ Blog โ€บ The cloud comes to you
AI Factory as a Service ยท Part 2 ยท Hyperscalers come to you

The Cloud Comes to You: How AWS, Microsoft, Google and Oracle Put AI Factories Inside Customers' Buildings

For fifteen years the big clouds told customers to move everything to them. Sovereignty rules and AI's hunger for sensitive data have reversed that for some of their most valuable customers. Every major cloud now offers a version of itself that runs inside the customer's own building, sometimes cut off from the internet entirely, with AI models included. This part compares the four offerings, the customers using them and the catch.

FL
FrontierAGI Team
โš ๏ธ How this was researched (as of 26 Oct 2026) Sources are the providers' own announcements (Amazon, Microsoft, Google, Oracle) and coverage by DCD, Network World, SiliconANGLE, CRN Asia, Techzine, SDxCentral and others. Product names and availability change often, and some Microsoft components are timed for its November 2026 Ignite event; where we could not confirm general availability we say so. Customer names are those announced publicly; most deployments are not. See Reading the numbers.

1. Four offerings

The hyperscalers' on-premises AI clouds

Tap a provider for its case file.

2. Timeline

The cloud moves on-premises, 2019 โ†’ 2026

One lane per provider. Solid dots are confirmed by the provider, 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.

The idea is not new. Oracle launched Dedicated Region, the full Oracle cloud in a customer's data center, in July 2020, with Nomura Research Institute and Oman's government among the first customers. AWS sold racks of its own hardware for customer sites as Outposts from 2019. What changed in 2025 is AI: the offerings now include large GPU clusters and frontier models, and some run with no connection to the provider at all.

3. The bet: keep the software, give up the building

A hyperscaler's profits come from its software and services as much as its hardware. If a ministry or bank will not move its data to the public cloud, the next-best outcome is that it runs the provider's software stack, tools and models in its own building, rather than a rival's or an open-source stack. That keeps the customer on the provider's platform, sells its chips (Trainium, for AWS), and positions it for the day rules relax.

๐Ÿ’ก The reverse of the Case StudiesIn the Case Studies, the big clouds built giant factories and invited customers in. Here they ship a smaller factory to the customer. AWS described its offering as working like a private AWS Region inside the customer's data center.

4. Side by side

AWS AI FactoriesMicrosoft sovereign private cloudGoogle Distributed CloudOracle Dedicated Region / Alloy
LaunchedDec 2025Azure Local (2024); disconnected AI 2026GDC (2021); Gemini GA on air-gapped Aug 2025Dedicated Region Jul 2020; Region25 Oct 2025
ChipsNVIDIA GPUs or AWS TrainiumCustomer or partner hardware, including NVIDIA GPUsNVIDIA GPUsNVIDIA, including Blackwell; customer-owned GPU option
AI modelsAWS AI services and foundation modelsFoundry Local, large models for qualified customersGeminiOCI AI services with NVIDIA AI Enterprise
Disconnected?Private-region style (connected)Yes: disconnected operationsYes: built to stay air-gappedConnected to Oracle for operations
Who runs itAWSCustomer, on Microsoft softwareCustomer (air-gapped) or GoogleOracle; partners run Alloy
Smallest size~$200M over 5 years (AWS executive)A few servers (Azure Local)Rack-scale configurationsThree racks (Region25)

5. From connected to air-gapped

How disconnected can it be?

Move the slider from a private region linked to the provider to a fully air-gapped site. Support levels are our reading of the providers' public descriptions, simplified.

An air-gapped AI factory has real costs. Models must be delivered and updated by hand, security patches arrive on physical media or through tightly controlled transfers, and the provider's engineers cannot log in to fix problems. Google says its air-gapped offering does not need a connection to Google Cloud or the public internet to manage its infrastructure, services, APIs or tooling, and is built to stay disconnected permanently. Microsoft says Azure Local's disconnected operations are available worldwide, with large models on Foundry Local available to qualified customers.

6. Inside the building

Who owns what on the customer's site

A simplified diagram. Tap a part.

Customer's data center Space ยท power ยท cooling (customer) Provider's racks Customer data & apps AI models link to provider (optional)
Tap a part.

7. Customers

Named customers and partners

Only publicly announced names. Filter by provider.

The pattern is clear even from the few public names: governments and security agencies (Singapore, Oman, Bangladesh), telcos (Vodafone), financial services providers (NRI) and national champions (Humain). Most deployments, especially in defence and intelligence, are never announced. Oracle says 60 Dedicated Region or Alloy regions are live or planned worldwide.

8. The catch

Minimum spendAWS quoted around $200 million over five years for its AI Factories. Oracle has pushed its entry point down to three racks, a sign that size was a barrier.
Lock-inThe point of the offering, for the provider, is that the customer builds on its stack. Moving to another cloud or to open-source tools later is hard.
Less of the cloudDisconnected sites get fewer services, slower model updates and no remote help. The customer still has to supply space, power and cooling, often the hardest part.
Cloud, on siteAll four big clouds now do it
Models includedGemini, Foundry Local, AWS models
Air-gapped AIGoogle and Microsoft run fully offline
For the fewBig budgets, long contracts
VerdictPutting the cloud inside the customer's building is how the hyperscalers keep governments, banks and telcos on their platforms despite sovereignty rules. Google and Microsoft have gone furthest with fully disconnected AI; AWS brings the most hardware choice, including its own chips; Oracle has the longest track record and the smallest entry point. All four share the same limits: high minimums, deep lock-in and a building the customer must still power and cool. Part 3 looks at the company whose blueprint sits underneath almost all of them: NVIDIA.

9. Reading the numbers

AWS
AI Factories launched at re:Invent, Dec 2025: NVIDIA GPUs or Trainium, AWS networking, storage and AI services in the customer's data center; customer supplies space, power and connectivity; ~$200M over five years (AWS executive, as reported); Humain AI Zone with up to 150,000 chips. Outposts from 2019.
Microsoft
Azure Local disconnected operations available worldwide; Foundry Local large models for qualified customers; some components timed for Ignite (17โ€“20 Nov 2026); GA of all parts not confirmed.
Google
Gemini on GDC air-gapped generally available Aug 2025; connected version then in preview; CSIT, GovTech Singapore and HTX among first users; Teradata sovereign cloud on GDC air-gapped in private preview (Oct 2026); Forrester Wave leader (2026). A later expansion with Blackwell Ultra and a newer Gemini model is reported by one source โš ๏ธ.
Oracle
Dedicated Region announced 8 Jul 2020 (NRI, Oman); Vodafone six regions (deal from 2022); Region25 from three racks (Oct 2025), Fujitsu a customer; 60 Dedicated Region or Alloy regions live or planned; customer-owned GPU option.