“Microsoft Earnings, Apple Earnings” — @Ben Thompson

Why this is in the vault

Two earnings recaps in one Update, but the Microsoft segment is the load-bearing piece for RDCO. Thompson surfaces Nadella and Hood explicitly reframing Microsoft 365 as “seats plus consumption” - the first time the company has publicly admitted that pure per-seat licensing breaks under agents. This is the business-model echo of the WTI agent-boss framing filed yesterday and is directly relevant to how RDCO positions agent-priced services. Apple segment is secondary but adds a useful data point: Mac Mini/Studio shortages are being driven by agentic harness workloads (CPU-bound), not on-device inference.

The core argument

Microsoft segment

Q3 FY26 beat top and bottom line ($82.9B revenue, +18% YoY; $31.8B net income). Azure grew 40%. Capex guided to $190B for the year, +61% over 2025. Copilot paid users hit 20M, +33% since January. Microsoft also let Anthropic models into Copilot.

Thompson’s actual frame: Nadella and Hood used the call to debut a new business-model description because agents structurally break the per-seat model. Agents (1) accomplish more than a person, (2) consume vastly more compute than a person, and (3) reduce headcount, which collapses seat count. The new framing is “seats plus consumption” - a license business plus a metered consumption business, billed like Azure. Hood explicitly said bookings will “look a little different” and that usage may not flow through bookings the same way.

Thompson’s tartest line: he “loves” Microsoft framing this as if the model is “magically appearing out of thin air, instead of it clearly being something that Microsoft needs to happen.” Translation - this is defense, not vision.

Three downstream implications Thompson highlights:

Sales motion problem: Microsoft has to retrain its own field and partners away from seat-coverage motions toward intense-usage motions. Internal lift is real.
Customer pressure: usage-based pricing forces Microsoft to actually deliver value or churn accelerates.
The real existential threat is not customers refusing AI - it’s customers adopting AI that isn’t Microsoft’s, which routes spend away. Thompson cross-references his own “Agents Over Bubbles” model+harness thesis: if model+harness is the integration point, Microsoft’s defense is harness + proprietary company data (SharePoint, Teams, mail, docs) that sits as the “constantly updated” context layer underneath whatever model wins.
Tail-end aside: Microsoft gets OpenAI IP free for a long time, which Hood admitted will fatten margins. Thompson notes this may quietly motivate Microsoft to make one model work best despite multi-model bluster.

Apple segment

Q2 FY26: $111.2B revenue, iPhone +21.7% YoY to ~$57B (iPhone 17 cycle), gross margin 49.3% (iPhone-era record). Forward guide 14-17% revenue growth vs analyst 10%. Two structural problems Cook flagged:

Memory costs will be “significantly higher” beyond June quarter and impact will grow.
Chip shortages: not enough A19 Pro for iPhones, and the MacBook Neo blew through the A18 Pro stockpile (per Tim Culpan). TSMC has no spare capacity for either.

Thompson’s frame: Apple is finally leaning into “Mac as the best platform for AI,” and he was right a year ago in “Apple AI’s Platform Pivot Potential” - but wrong about why. He thought Macs would matter for on-device inference. Reality: Mac Mini/Studio shortages are driven by agentic harness workflows (he cites OpenClaw) that lean on CPU more than GPU because the harness uses cloud LLMs. Apple did not see this coming and is months behind on Mini/Studio supply.

The bigger structural takeaway, in Thompson’s voice: “even Apple is a 2nd priority to the insatiable demand for everything from memory to TSMC fab capacity.” Apple has never operated in a world where it couldn’t dictate supply terms - John Ternus is inheriting that for the first time. AI’s gravity is now strong enough to displace Apple’s most-favored-customer status.

Mapping against Ray Data Co

Strong relevance via the Microsoft segment. This is the corporate-finance confirmation of yesterday’s WTI “agent-boss” pattern - same direction, different artifact. Microsoft is publicly conceding that the per-seat model collapses under agents and is reaching for “license + consumption” hybrid pricing. That has direct implications for how RDCO prices anything agent-shaped:

The “seats plus consumption” frame is the right shape for any RDCO agent-as-service offering. A flat-license-only model under-monetizes when an agent runs hot; pure consumption metering scares buyers. Microsoft’s hybrid is the de-risking template.
The harness + proprietary data thesis Thompson reasserts is the same defensibility argument behind the 2026-04-15-thariq-claude-code-session-management-1m-context / “thin harness, fat skills” pattern that already informs RDCO’s COO agent. RDCO’s moat is not the model; it is the harness wrapped around founder-specific context (vault, Notion board, Gmail, calendar, channels). Same shape Microsoft is betting on at corporate scale.
Microsoft’s internal-motion problem (“not pure seat coverage type of motions like in the past”) is a tell that even hyperscalers are scrambling to retrain sales orgs around agent pricing. Smaller teams selling agent services have a simpler path here because there is no entrenched seat-motion to undo. Cross-link to bet-pricing thinking when RDCO begins productizing.

Weak relevance via the Apple segment. The Mac-as-AI-harness-host point is mildly interesting (validates the local-Mac-as-control-plane pattern that the channels agent uses on the Mac mini), and the memory/chip squeeze is a useful macro data point - but neither moves any RDCO decision today. File for trend tracking, do not act on.

Decision needed: none. Status note. The Microsoft pricing-model shift is a vault-priority concept though - worth a follow-up concept article on “seats plus consumption” as the agent-pricing default, citing this Update plus the WTI piece. Logging that as a queue candidate, not a now-decision.