How is AI changing azure cloud?
Azure has shifted from renting servers to renting intelligence. Microsoft Foundry — formerly Azure AI Foundry — is now the unified enterprise AI platform, offering access to over 11,000 models spanning OpenAI, Anthropic, Meta, Google, xAI, and Hugging Face, with built-in tracing, monitoring, and evaluations. Even the branding changed: "Azure AI Foundry" became "Microsoft Foundry."
Why — the first-principles explanation
The deepest change is what Azure actually sells. Classic cloud sold capacity: virtual machines, storage, bandwidth. You rented a computer by the hour and did whatever you wanted with it. The unit of value was a resource. AI inverted this. Azure now sells capability by the token — you don't rent a GPU and install a model, you call an endpoint and pay per unit of intelligence consumed. The customer no longer manages the machine, and often does not know or care which machine ran the request.
This matters because it changes the economics of scarcity. Ordinary cloud resources were effectively unlimited — need another VM, get another VM. Frontier AI compute is not. GPUs are physically constrained, power-constrained, and expensive, which is why AI services introduced things classic cloud never had: quotas, registration requirements for the largest models, capacity waitlists, and regional availability gaps. Microsoft's own docs reflect this — some models require registration while smaller variants like gpt-5-mini, gpt-5-nano, and gpt-5-chat do not. Scarcity came back to the cloud, and it reshapes how you architect.
The second structural shift is from single models to a model marketplace. Azure was once the Azure OpenAI cloud. Microsoft Foundry now spans over 11,000 foundational, open, reasoning, multimodal, and industry-specific models from OpenAI, Anthropic, Meta, Google, xAI, and Hugging Face. Microsoft draws a meaningful line here: models sold directly by Azure are billed through your Azure subscription, covered by Azure SLAs, and supported by Microsoft — versus models merely hosted from partners. That distinction determines who you call when it breaks, and it is the kind of boring detail that decides real enterprise architecture.
Third, the platform grew a conscience layer. Foundry unifies agents, models, and tools with tracing, monitoring, evaluations, and enterprise setup configuration. This exists because non-deterministic software broke traditional operations. You cannot unit-test a model the way you test a function — the same input can produce different output. So evaluation became infrastructure, not an afterthought. That is genuinely new: classic cloud never needed to ship tooling for "is the answer good?" because deterministic code either worked or threw an error. Finally, the naming churn is itself a signal. Azure AI Foundry became Microsoft Foundry, with older portals now labeled "classic." When a company renames its flagship platform mid-cycle and re-labels the previous generation, it is telling you the strategy is still moving. Expect further change.
An example that makes it click
Imagine Azure used to be a company that rented you an empty commercial kitchen. You brought your own recipes, hired your own chef, and cooked whatever you liked. They charged by the hour for the room. Simple.
Now they still rent kitchens — but the popular product is a window where you order a finished dish and pay per plate. You don't see the kitchen. You don't hire the chef. And here's the twist: the best chefs are genuinely scarce. There's a waitlist for the famous one, while the junior cooks are available instantly. That never happened with the old empty-room business, where there was always another room. Oh, and they hung tasting stations by the door — because unlike a rented room, which either has power or doesn't, a dish can come out wrong in ways nobody can automatically detect. Somebody has to taste it.
Key facts
- Azure AI Foundry has been renamed Microsoft Foundry; earlier portals are now documented as "classic" versions.
- Microsoft Foundry provides access to over 11,000 foundational, open, reasoning, multimodal, and industry-specific models spanning OpenAI, Anthropic, Meta, Google, xAI, and Hugging Face.
- Microsoft Foundry is a unified Azure platform-as-a-service for enterprise AI operations, unifying agents, models, and tools with tracing, monitoring, evaluations, and enterprise setup configuration.
- Models sold directly by Azure are billed through your Azure subscription, covered by Azure SLAs, and supported by Microsoft — a distinct category from partner-hosted models.
- Available OpenAI image models include gpt-image-1, gpt-image-1-mini, gpt-image-1.5, and gpt-image-2; video models include sora and sora-2 (as of 2026).
- Some models require registration to access, while gpt-5-mini, gpt-5-nano, and gpt-5-chat do not — reflecting genuine capacity scarcity rather than arbitrary gatekeeping.
▶ The 60-second explainer (script)
How is AI changing Azure? The deepest change is what Azure actually sells. Old cloud sold capacity. You rented a virtual machine by the hour, installed what you wanted, and did your thing. The unit of value was a resource. AI flipped that. Azure now sells capability, metered by the token. You don't rent a GPU and install a model — you call an endpoint and pay per unit of intelligence. You don't know which machine ran it. You don't care. And that changes something fundamental: scarcity came back. Ordinary cloud resources were effectively unlimited. Need another VM? Here's another VM. Frontier AI compute is not like that. GPUs are physically limited, power-limited, expensive. So Azure now has things classic cloud never had — quotas, capacity waitlists, regional gaps, and registration requirements for the biggest models. Microsoft's own docs show it: some models need registration, while the smaller ones like gpt-5-mini and gpt-5-nano don't. Second shift: Azure used to basically be the OpenAI cloud. Now Microsoft Foundry spans over eleven thousand models — OpenAI, Anthropic, Meta, Google, xAI, Hugging Face. And Microsoft draws a line that matters: models sold directly by Azure are billed on your subscription, covered by Azure SLAs, and supported by Microsoft. Partner-hosted models aren't the same deal. That determines who you call at 2 a.m. Third: the platform grew an evaluation layer — tracing, monitoring, evals. Why? Because non-deterministic software broke traditional ops. You can't unit-test a model. Same input, different output. So checking whether the answer is any good became infrastructure. Classic cloud never needed that; code either worked or threw an error. And notice the branding: Azure AI Foundry is now Microsoft Foundry, and the old portals are labeled classic. When a company renames its flagship mid-cycle, the strategy is still moving.
What authoritative sources say
People also ask
Is Azure AI Foundry the same as Microsoft Foundry?
Yes — it was renamed. Microsoft Foundry is the current name, and the earlier portal and its documentation are now labeled "classic." Older tutorials referencing Azure AI Foundry describe the same lineage of product.
Can I use non-Microsoft models on Azure?
Yes. Foundry spans over 11,000 models including Anthropic, Meta, Google, xAI, and Hugging Face. But check whether a model is "sold directly by Azure" — that category is billed on your Azure subscription and backed by Azure SLAs and Microsoft support.
Why do some Azure AI models require registration?
Because frontier compute is genuinely scarce and expensive. Larger models gate access through registration and quotas, while smaller variants like gpt-5-mini, gpt-5-nano, and gpt-5-chat are available without it.
What is the biggest practical difference for developers?
Non-determinism. The same input can yield different output, so you cannot test AI features the way you test ordinary code. That is why Foundry ships tracing, monitoring, and evaluation tooling as core infrastructure.
Did AI change Azure's pricing model?
Fundamentally. Classic cloud billed for resources over time — a VM per hour. AI services bill per token consumed, so cost scales with usage volume and prompt size rather than with how long a machine stays on.