Microsoft is being pushed to own the stack while trust and cost tighten the terms of AI use

July 29, 2026

This week suggests Microsoft is trying to become less dependent on any one model supplier while becoming more useful as the place where enterprise AI is governed, deployed and paid for. Microsoft launched MAI image, voice and cyber models in Foundry and began moving workloads from OpenAI models to its own MAI models in products including Bing Image Creator, PowerPoint image features, OneDrive image editing, Dynamics 365 Contact Center and Azure Voice Live. It also added Anthropic’s Claude Opus 5 to both Microsoft 365 Copilot and Microsoft Foundry, while Manulife signed an expanded five-year partnership covering Microsoft 365 E7 Frontier Suite, Agent 365 governance and Copilot for more than 30,000 employees. This appears to extend last week’s line that competition is shifting from a single chatbot or base model toward the full stack of model choice, governance, memory, security and workflow deployment. It also suggests that for Microsoft, control of the customer environment may matter more than exclusive control of the underlying model.

The same week also showed why that stack is being built around constraints rather than around novelty alone. SAP decided to restrict AI token spending after internal use drove costs higher, and Alibaba Cloud described an agent optimisation aimed at cutting context-related token use. OpenAI launched Presence for governed enterprise agents, while Microsoft released a Cosmos DB-based memory provider for its Agent Framework and AWS updated AgentCore Gateway to the latest MCP specification with enterprise authorization and lifecycle controls. This likely means the next phase of enterprise AI buying will turn less on whether a company wants agents at all and more on whether those agents can be run within budgets, permissions and audit rules.

Trust pressure is rising at the same time. uniVersa reported that an OpenAI-linked crawler reached personal customer data during an IT migration, 404 Media reported that public Claude share links were being indexed by Google, Codeberg changed its terms to block AI training uses and heavily autogenerated projects, and the EU AI Omnibus entered into force. NIST also launched a sequestered evaluation programme, and the Reserve Bank of India issued draft model-risk guidance for regulated financial firms. We read this as evidence that AI is moving deeper into normal institutional control systems: privacy incidents, provenance signals, blind testing and formal risk management are no longer side issues, but part of the conditions under which deployment is allowed.

Underneath all of this, the infrastructure build-out still points in the same direction. Alphabet reported faster Google Cloud growth tied to AI infrastructure and enterprise AI demand, AMD launched its Helios rack-scale system and announced an Anthropic partnership for future GPU deployment, Intel raised capital spending after AI data-centre chip demand lifted revenue, and NVIDIA-backed projects in Korea outlined further large AI-factory expansion. This is likely to reinforce the line that AI demand is not behaving like a short burst around one model generation. The more immediate consequence for Microsoft, though, may be simpler: if compute stays expensive and abundant model supply keeps spreading across clouds, the durable advantage is likely to sit with whoever can package models into governed, costed and trusted operating systems for work.