Microsoft CEO Nadella: Over-Reliance on OpenAI Led to Passivity, Will Strengthen Control Through In-House MAI Models
Miles Bennett
Nadella admitted for the first time that Microsoft's deep dependence on OpenAI had left the company exposed, unveiling in-house MAI models to reclaim control over pricing and compute — a dual-track strategy of self-built plus partnership is now official.
Why is Nadella admitting vulnerability now?
Nadella was blunt: "We are now competing with a bunch of people whose names I didn't even know in 2018."
This means → the AI race has shifted from "Microsoft + OpenAI vs. Google" to a crowded free-for-all, and betting on a single partner carries far more risk than it used to.
Microsoft's response: a from-scratch MAI model family with what Nadella calls "clean lineage" — in plain terms = no OpenAI code or architecture underneath, so pricing power and compute allocation stay entirely in Microsoft's hands.
How can self-built and OpenAI coexist?
Microsoft retains intellectual-property access to OpenAI's models and uses reverse knowledge distillation (RKLD) — compressing a large model's capabilities into a smaller one — plus reinforcement learning (RL) to boost its own models.
This means → Microsoft hasn't broken up with OpenAI. It has shifted from full dependence to "own frontier + OpenAI frontier" running in parallel.
Nadella frames the end-state as a "multi-tenant learning system" — in plain terms = every enterprise client gets its own self-improving AI engine; Microsoft provides the platform, each customer runs its own race.
GPUs are scarce — where is Microsoft sending its compute?
Nadella explicitly refused to sell raw GPUs to emerging AI labs for quick Azure revenue: "If you want short-term cloud revenue, it's easy — just sell compute to those labs. We won't do that."
The three-tier allocation priority: ① hyperscale cloud customers → ② high-margin proprietary apps (GitHub, M365) → ③ internal R&D.
This reflects a deeper bet — frontier labs like OpenAI and Anthropic will eventually build their own hardware infrastructure, making them risky anchor tenants.
Why is the SaaS subscription model breaking?
Nadella argues the traditional SaaS bundle — data, business logic, and UI locked into one subscription — no longer works. AI continuously calls backend data, and inference costs are surging.
His answer is hybrid billing: "Per-seat subscription plus consumption-based pricing. This is 100% the future."
In plain terms = software used to be "all-you-can-eat monthly." Now every AI call costs real compute, and without usage-based pricing, vendor margins blow up.
Agents and edge devices — what is Microsoft betting on next?
Nadella is staking future growth on AI agents, with priority domains: coding, security, and knowledge work.
He also revealed Project Solara — running native agents on Windows and edge devices to break the smartphone's lock on user access.
This means → Nadella envisions "unmetered intelligence": instead of watching cloud bills climb, buy a Windows device and amortize the cost — using on-device compute to offset cloud inference bills.
Content is for reference only, not financial advice.