Satya Nadella on AI: @NoPriorsPodcast x Latent Space Crossover Special at Microsoft Build 2026

Platforms must create more value than they capture

Nadella defines successful platforms by their ability to generate value for participants beyond what the platform itself captures, requiring an ecosystem approach rather than a single-model monopoly.

Every company can operate at the frontier

Microsoft aims to enable any enterprise or AI-native startup to participate as a first-class citizen by providing the stack, tooling, and recipes to create their own specialist models.

Clean lineage pre-training for MAI models

Microsoft's MAI models focus on rigorous data quality and ablations to build a 'cognitive core' that avoids the benchmark-maxing pitfalls common in open-weight models.

Private evals are the new enterprise IP

Companies should build proprietary evaluation datasets as their primary intellectual property, allowing them to 'hill climb' using any frontier model without leaking traces or depending on a single vendor.

Hill climbing scaffolds enable specialization

The platform provides scaffolding around base models so companies can collect traces, build RL loops, and train specialist agents on their private data.

The 'harness' combines models, data, and tools

Enterprise AI requires a multimodal harness integrating multiple models, rich context layers, and progressive tool disclosure to execute plans efficiently.

Context preparation is the new moat

Success depends on preparing the context layer so agents can execute tasks efficiently, with Microsoft's GitHub harness serving as an open template for enterprise use.

Multimodal harnesses outperform single-model training

Evidence from Microsoft's security tools demonstrates that a harness using multiple models with tools can find vulnerabilities that specialized single models miss.

Rebuilding the IDE for agentic coding

AI coding assistants have become so effective that they create cognitive overload, necessitating new UI paradigms like canvases to manage hundreds of concurrent agent sessions.

Autonomous agents handle 'glue work'

Durable, long-running agents will handle routine coordination tasks overnight, allowing human capital to focus on judgment while compounding value through trace collection.

Rebundling the SaaS stack

While underlying data models and business logic remain valuable, they must be unbundled from rigid UIs and repackaged to feed autonomous workflows rather than traditional apps.

Token capital joins human capital on the balance sheet

Companies will record 'token expertise' as assets by capturing tacit knowledge in company-specific veteran agents trained on accumulated traces.