No PriorsWe Need An Ecosystem in AI, And Every Company Can Win A Place In It
CHAPTERS
- 0:00 – 3:12
Why AI needs an ecosystem (not one model): platform value and participation
Satya frames Microsoft Build’s core thesis as an ecosystem play where many companies can create and own meaningful AI value, not just consume it. He defines a strong platform as one that creates more value on top of it than it captures inside it, and argues Microsoft’s job is to make the recipe and stack for participation clear.
- •Platform shifts reward ecosystems that enable others to build durable businesses
- •Goal: every company can point to AI it created, not only AI it uses
- •Microsoft’s role: provide the stack/tooling so enterprises and startups can participate first-class
- •Developer conference should empower builders—not “worship at the altar of one model”
- 3:12 – 5:48
Microsoft’s MAI training approach: clean lineage, cognitive core, and company-specific specialization
Satya describes how Microsoft is training MAI models with emphasis on clean data lineage and ablations to ensure real-world performance, not benchmark spikes. He explains a strategy where a base model becomes a specialist via scaffolding, traces, and private evaluations, enabling smaller models to “hill climb.”
- •Clean lineage and careful ablations matter more as training data gets noisier
- •Open-weight models can look good on benchmarks but fail in practice
- •Specialization comes from scaffolds + traces + RLE + private evals, not just bigger models
- •Frontier capability can be achieved by distilling/optimizing from frontier traces into smaller reasoning models
- 5:48 – 7:33
What Microsoft underestimated: real-world deployment complexity and proving value beyond tokens
Reflecting on scaling laws and early compute bets, Satya says the industry underestimated the complexity of deploying AI to deliver measurable outcomes. Benchmarks matter, but the true evaluation is whether customers achieve unique, valuable real-world results—and whether the economics of token usage map to value creation.
- •Scaling laws drove rapid capability gains, but deployment is the hard part
- •Benchmarks are insufficient; real-world outcomes are the real eval
- •Token cost skepticism reflects weak linkage between tokens and value delivered
- •Industry mindset shift: design systems that create value at each step of token spend
- 7:33 – 9:37
Where customers see value: agentic coding, new IDEs, and automating “glue work”
Satya highlights coding as a major value driver, but notes it’s forcing a rethinking of developer UX (canvas, new IDE patterns, managing many agent sessions). Beyond code, he points to durable, long-running agents that automate coordination and “glue work,” compressing workflows across enterprises.
- •Coding copilots work so well they create new UI/IDE needs (sessions, canvas)
- •Agentic work increases cognitive load unless the interface evolves
- •Large opportunity in automating glue work and task completion across orgs
- •Long-running agents with delegated authority become “overnight coworkers”
- 9:37 – 12:23
The enterprise harness: models + data + tools in a feedback loop (and why context prep is ‘the magic’)
Satya generalizes the notion of an agent ‘harness’ for enterprise work: orchestrating multiple models, tool access, and rich context with an iterative loop. He argues the hardest, most valuable work is preparing the context layer so plans execute efficiently, and notes Microsoft aims to offer open, multi-model harnesses through Foundry and product surfaces.
- •Harness components: models, tools, data/context, connected in a continuous loop
- •Progressive tool disclosure improves token efficiency
- •Context preparation is a major hidden workload and key differentiator
- •Microsoft products (Copilot, security, science) are built as multi-model tool-using harnesses
- 12:23 – 15:50
Frontier intelligence for everyone: private evals as modern IP and the ‘hill climbing’ advantage
Satya presents a key platform promise: every company should operate at the frontier with its own frontier intelligence. He argues private evaluations and proprietary traces become central IP—enabling companies to swap models, improve performance, and stay in control without leaking sensitive learning signals.
- •Tagline: operate at the frontier with your own frontier intelligence
- •Stable equilibrium requires companies to compound value on top of the platform
- •Private evals may be the most defensible IP in the agent era
- •Control comes from being able to switch models and still hill climb using your own evals/tools/context
- 15:50 – 18:23
Redefining company value: traces, ‘company veteran agents,’ and capturing tacit knowledge
The conversation shifts to how enterprises compound human and token capital. Satya suggests traces between humans and agents can train ‘company veteran agents’ that encode tacit operational knowledge—potentially becoming balance-sheet-worthy assets because they are now capturable and reusable.
- •Future value comes from compounding human judgment with token/agent labor
- •Enterprise traces (human↔agent work) become critical training context
- •‘Company veteran agents’ encode tacit knowledge learned over time
- •Provocative idea: agent-learned tacit knowledge could become a recognizably valuable asset
- 18:23 – 19:20
Vendor vs. enterprise agents: unbundling SaaS into data models, business logic, and new bundles
Satya argues classic SaaS captured workflows by stacking schema, business logic, and UI; the agent era forces a re-litigating of that stack. Stable underlying schemas (e.g., general ledger) and semantic models remain valuable, but vendors will need to unbundle/rebundle offerings and find new business models as agents consume these layers differently.
- •Traditional SaaS = schema + business logic + UI + configuration
- •Some layers should remain stable and reused (schemas, semantic models)
- •Agents change how value is accessed—driving unbundling and rebundling
- •New business models will emerge as usage shifts from end-users to agents
- 19:20 – 21:47
Work IQ and the M365 ‘hidden database’: turning collaboration exhaust into agent-ready context
Satya explains that Microsoft 365 contains one of the most important ‘databases’ in a company—previously captive to apps like email and Teams. With Work IQ, agents can query transcripts, documents, and meetings to produce actionable plans (e.g., mapping meeting discussions to code changes), but serving agents requires re-architecting backend systems built for mailboxes and human interaction.
- •M365 data becomes a platform for agents, not just for end-user apps
- •Work IQ enables cross-artifact retrieval (meetings → repo changes)
- •Agent usage may exceed end-user usage, changing scaling assumptions
- •Backends built to serve mailboxes/inboxes must evolve to serve agents efficiently
- 21:47 – 24:02
Pricing in the near term: subscriptions, per-user entitlements, consumption meters, and outcome tensions
Satya predicts multiple pricing models will coexist: per-user subscriptions persist because budgeting needs certainty, but consumption pricing will expand as agent intensity grows. He’s skeptical of pure outcome-based pricing at scale because customers resist sharing upside once outcomes materialize, and he uses GitHub Copilot’s evolution as an example of adjusting to agent-driven usage patterns.
- •Per-user pricing persists as a budgeting/entitlements mechanism
- •Consumption metering becomes necessary as agents run continuously at scale
- •Outcome-based pricing is attractive until customers see the ‘royalty-like’ downside
- •GitHub Copilot shifted pricing as usage moved from interactive to large-scale agent sessions
- 24:02 – 25:58
Will SaaS endure? The build-vs-buy pendulum and the true cost of maintenance in an agent world
Elad raises ‘agent euphoria’ where enterprises try to rebuild vendor software; Satya expects a budget cycle will clarify equilibrium. He emphasizes that while generating software becomes cheaper, maintaining it—security fixes, ongoing token costs, operational responsibility—still drives many teams back toward flexible vendors and composable solutions.
- •Enterprises may overbuild internally at first, then rebalance toward buy/compose
- •Decision rule: build when build+maintain marginal cost is lower than acquiring
- •Security and maintenance remain ongoing costs, even with coding agents
- •Vendors must be flexible; enterprises want agency over composition and workflows
- 25:58 – 28:17
What Satya is building: long-running Foundry agents that publish into Teams
Satya shares hands-on experimentation building long-running ‘autopilot’ agents using Work IQ, Foundry, and a memory backend, then publishing them directly into Teams. The point is not that everyone becomes an engineer, but that leaders and generalists can now inspect and shape software artifacts they previously couldn’t touch.
- •Building is increasingly accessible, even for non-traditional builders
- •Example: chief-of-staff-style autopilot agent monitoring work continuously
- •Foundry agents + memory backend + Teams distribution create end-to-end workflows
- •Copilot/Sessions expands agency over codebases and internal tooling
- 28:17 – 30:53
Future engineering roles: broader scopes, new infrastructure demands, and leverage for generalists
Satya anticipates experimentation in how roles evolve, citing shifts like ‘full-stack builder’ models that blend disciplines while keeping specialties. He notes new infrastructure challenges (RLE environments, distributed systems needs even in app teams) and argues the biggest returns may accrue to generalists who can span knowledge work and app building.
- •Roles may broaden in scope while preserving specialist ‘edges’
- •RLE/reward environments are hard infrastructure problems, even for app teams
- •Distributed systems skills become more important across the org
- •Generalists gain leverage: knowledge work increasingly includes building apps
- 30:53 – 34:30
Ambition and ‘meta work’: building agentic systems that do the work (Azure networking as an example)
Sarah asks how Microsoft can be more ambitious; Satya argues ambition comes from new conceptual models that make previously impossible outcomes feasible. He describes Azure networking teams reconceiving their mission from operating networks to building an agentic system (“Miles”) that operates the network—shifting from headcount to tokens and turning operational work into meta work.
- •True ambition: make the impossible possible, not just make hard things easier
- •Organizations need permission to reconceive workflows around new capabilities
- •Azure networking example: build an agent to run fiber/network operations
- •Shift in constraints: requesting tokens instead of headcount signals a new operating model
- 34:30 – 37:56
Data center scale and community permission: energy, water, jobs, and tangible local benefits
Satya acknowledges unprecedented datacenter build-out and stresses that continued progress depends on earning community trust. He argues benefits must be concrete—energy grid improvements, responsible water systems, training and jobs, and tax base growth—otherwise the industry loses ‘permission’ to expand.
- •Build-out is extraordinary, redefining hyperscaler scale and operations
- •Community skepticism is valid; tech must answer hard questions with facts
- •Key impact areas: energy prices/grid, water stewardship, job creation, tax base
- •‘Permission’ to build depends on tangible benefits, not PR narratives
- 37:56 – 42:26
Societal impact and education: proving broad participation and rethinking credentials and pedagogy
Satya says the biggest societal challenge is making it real that everyone can participate as a first-class actor in the new economy, because public trust in ‘trust us’ narratives is low. On education, he argues AI changes access to information, but incentives, credentials, and pathways to jobs must be reinvented—opening the door for new institutions or ‘new universities.’
- •Public trust requires tangible benefits visible within 12–18 months
- •Broad participation is essential to legitimacy and political sustainability
- •Education needs more than tutors: rework incentives, credentials, and employment links
- •Opportunity for startups to build new pedagogy or new institutional models (a ‘new university’)
