Your newsletter digest — June 5, 2026

One source today: Ben Thompson's interview with Microsoft CEO Satya Nadella, recorded right after Build 2026. The big theme is Microsoft figuring out what it actually is — not "the OpenAI partner," not "a second-tier frontier lab," but a platform company that wants every enterprise to have its own AI learning machine.

Nadella's framing at Build was unusually candid about the recalibration Microsoft has been doing. His core argument: the future of AI for enterprises isn't about picking the best frontier model — it's about building what he calls a hill-climbing machine, a system where a company's private data, private benchmarks, and reinforcement learning environment continuously improve a model toward that company's specific goals. Microsoft wants to be the platform layer that makes this possible for everyone.1

Three things came out of Build that anchor this vision:

1. Seven new MAI models — with clean lineage. Microsoft launched seven models built from scratch, without distillation from other labs' outputs, specifically so enterprises can fine-tune and RL-train them on proprietary data. The pitch: if your company's moat is tacit knowledge, you need models you can actually own and shape, not just API keys. Nadella was explicit that MAI models are one option in their multi-model harness (alongside OpenAI, Anthropic, and open-weight models) — enterprises plug whichever model wins their private evals.1

2. The OpenAI partnership is being actively managed down. Thompson pressed Nadella on whether Microsoft got "lulled to sleep" by over-relying on OpenAI. Nadella's answer was careful but clear: the original arrangement — where OpenAI was essentially the sole key tenant of Microsoft's AI infrastructure — was a concentration risk, and Microsoft has spent the last year diversifying its hyperscale book of business. The partnership extends to 2032 and remains a major IP and compute relationship, but Microsoft's own model lineage is now running in parallel.1

3. Project Solara: the ambient device bet. The quieter announcement was Project Solara — a platform for building agent-first ambient devices (think enterprise hardware in healthcare rooms, factory floors, meeting spaces) that connects to Microsoft's agent infrastructure rather than running local compute. Nadella's read is that wearables and ambient hardware fail when users have to actively interact with them; the value unlocks when a device can dispatch an agent and let the user go do something else. He explicitly positioned this as something Apple's vertically integrated stack can't easily replicate.1

Thompson asked directly whether software is dead. Nadella's answer: not dead, but the coupling that made SaaS defensible — data model + business logic + UI bundled together — is breaking apart. WorkIQ, Microsoft's newly opened M365 data layer, is now an MCP skill that any agent can query. That's the same database SaaS vendors previously had to build expensive integrations to reach.

The business model implication is that per-seat pricing alone doesn't survive this transition. Microsoft is moving to a hybrid: per-seat for budgetable usage entitlements, consumption-based for everything agents generate on top. The E7 SKU (roughly double the price of E5) is the first explicit attempt to bundle enough agent usage that enterprises can plan around it without getting surprised by a runaway compute bill.1

Nadella made one point here that stuck: real marginal costs in software are now unavoidable. Unlike the PC era — where developers assumed the next processor would solve performance problems for free — every agent inference costs something. That forces a kind of optimization discipline the industry never developed. Companies that don't measure what their AI stack actually produces will find they've signed up for a very large cloud bill without a matching revenue line.

Thompson didn't let Nadella off the hook on GitHub Copilot — Microsoft launched code completion first, held an early lead, and then watched Cursor and especially Anthropic's agentic models eat into that position. Nadella acknowledged it plainly: the team was thinking "code completions in the IDE" while the world moved to agent loops that can run a full task without a human in the loop. The recovery plan is to use the same multi-model harness across GitHub, M365, and security — and to lean into GitHub as the platform where external coding agents (Codex, others) are already showing up to work anyway.1

There's a subtext here that Thompson didn't fully surface: GitHub's reliability issues have been a real problem for this recovery. Nadella mentioned it briefly and said they'd work it. That's the quiet threat to the whole harness story — if developers can't trust the platform's uptime, the agent-loop value proposition falls apart before it starts.

The "hill-climbing machine" framing points at something the broader enterprise AI market is still figuring out: the difference between using AI and owning AI. Right now most companies are in the first category — calling APIs, plugging in LLMs, evaluating outputs manually. What Nadella is describing is a second phase, where the company's own feedback loops and private evals become the asset that compounds over time. That's the Microsoft pitch, but it's also the pitch from Salesforce, Workday, and every vertical SaaS vendor that still has deep customer data locked in. The race isn't just about which model is best on a public benchmark — it's about who controls the reinforcement environment where models actually learn to do your specific job.