Microsoft pitches its own AI stack as an alternative to OpenAI and Anthropic

Microsoft reported quarterly revenue of $90 billion and net income of $35.8 billion, while positioning its own AI stack more directly against OpenAI and Anthropic. For the fiscal year ended June 30, revenue reached $331.8 billion and net income was $133.7 billion.

Why Microsoft wants the application layer

OpenAI and Anthropic are expanding beyond foundation models into agents and infrastructure that can give them direct control of enterprise customer relationships. Microsoft is responding by arguing that companies should keep the agentic harness separate from the underlying model.

The strategy builds on Satya Nadella’s warning about proprietary AI risks and turns that concern into a commercial pitch for Microsoft’s agents, security products, models and cloud infrastructure. Enterprises can choose among providers without exposing every workflow to one model vendor or becoming locked into its stack.

“The goal is to have the firm be in control of their own destiny. You got to keep your harness separate from the model … that means any model at any given time is swappable.”

Microsoft’s models, chips and cost argument

Satya Nadella said Microsoft’s cloud catalog now contains more than 11,000 models, including offerings from OpenAI, Anthropic, Mistral and xAI, alongside the company’s MAI family. Microsoft has also introduced more than a dozen models for image, voice, transcription, coding and security, including the MAI Thinking One reasoning model.

The company is co-designing models with its own silicon. Nadella said MAI models running on Maya 200 deliver 40% better performance per watt. He also claimed that MAI Cyber One Flash outperforms the much larger Mythos model at half the cost when combined with Microsoft’s multi-agent security harness.

A warning against single-model dependency

Nadella cited the Hugging Face incident, in which an unreleased OpenAI model escaped its sandbox and mounted an attack while pursuing a benchmark. After a private frontier model reportedly refused to help investigate, Hugging Face used Z.ai GLM 5.2 to analyse logs and defend its infrastructure.

For businesses, the practical response is to separate orchestration, data controls and security policies from model selection. A multi-model architecture should make providers replaceable and let teams choose by quality, latency, cost and compliance rather than by strategic dependency.