Microsoft used its July 29 fiscal fourth-quarter earnings call to sharpen a strategy it has been building in public for more than a year: keep the enterprise agent layer under Microsoft’s control while allowing the underlying AI model to change. CEO Satya Nadella presented model substitutability as a way to improve cost, resilience and business continuity, not as a declaration that OpenAI or Anthropic had become enemies.
The numbers behind the strategy
Microsoft closed its fiscal year on June 30 with $331.8 billion in revenue and $133.7 billion in net income, according to the company’s Q4 earnings release. The same release disclosed a $3.2 billion gain from Microsoft’s Anthropic investment and separately detailed the effect of its OpenAI investments on reported results.
That leaves Microsoft in a complicated but increasingly deliberate position. It is an investor, infrastructure provider and distribution partner to frontier AI laboratories while also developing its own models, chips and enterprise software to reduce dependence on any single supplier.
Nadella’s pitch: swappable models, durable platform
In the official earnings-call transcript, Nadella said enterprises should keep the model separate from the harness containing their memory, context and action space. That architecture, he argued, makes any given model swappable.
He did not tell customers to abandon frontier systems. He said they could use frontier models for the tasks where those models perform best, lower-cost models for other workloads and internally trained models when organisations do not want to rely on an external provider.
Microsoft said its cloud catalogue now includes more than 11,000 models from providers including OpenAI, Anthropic, Mistral and xAI, alongside Microsoft’s own MAI family. It also reported a fivefold increase since the start of 2026 in customers building with models from multiple providers.
The strategic logic is clear. When Copilot, GitHub Copilot, Foundry and Microsoft’s security systems remain the stable layer, Microsoft can continue selling infrastructure and software regardless of which underlying model performs best on a particular task.
What the Hugging Face incident demonstrated
Nadella used the recent Hugging Face security incident to illustrate the operational case for model diversity. Hugging Face’s incident disclosure said its responders initially tried to analyse the attack using frontier models supplied through commercial APIs, but the providers’ safety controls blocked requests containing real attack commands and exploit material.
The company instead used the open-weight GLM 5.2 model on its own infrastructure. Hugging Face also said the model used by the attacking agent system remained unknown.
Nadella’s point was that an enterprise can become operationally constrained when one model refuses a legitimate security task. A second model may be required to investigate or remediate a problem that the first model cannot process. The episode supports Microsoft’s multi-model architecture, but it does not establish that OpenAI or Anthropic had become corporate threats to Microsoft.
MAI models, Maia silicon and MDASH
The names describe different parts of Microsoft’s stack. MAI is the company’s homegrown model family, while Maia 200 is Microsoft’s custom AI accelerator. Nadella said Maia 200 was delivering 30% better performance per dollar than the latest-generation hardware in Microsoft’s fleet and was supporting both OpenAI and MAI workloads.
Microsoft also said running its MAI models on Maia 200 produced 40% better performance per watt. These are company-reported benchmarks, but they show how Microsoft intends to connect its model development with its own infrastructure economics.
Codename MDASH is another distinct component. It is Microsoft Security’s multi-model agentic scanning harness, which coordinates specialised agents and different models across vulnerability discovery, validation and remediation tasks.
MDASH illustrates the architecture Nadella described to analysts. The surrounding pipeline, accumulated context and validation process remain in place even when Microsoft changes the models working inside it.
What actually changed
The earnings call was not Microsoft’s first public move towards a multi-model product strategy. In September 2025, the company announced model choice in Microsoft 365 Copilot, allowing its Researcher agent to use either OpenAI or Anthropic models and enabling Copilot Studio customers to mix models for specialised tasks.
What changed on July 29 was the directness with which Nadella connected that design to Microsoft’s economics. Swappable models can lower inference costs, protect business continuity and give Microsoft more leverage as it expands its own MAI models and Maia silicon.
Nadella did not publicly label OpenAI or Anthropic as threats. The more defensible reading is that Microsoft is building its platform so that no external model provider, including the frontier laboratories it has funded, becomes indispensable to the products and enterprise relationships Microsoft controls.
