"The Nvidia AI PC, Project Solara, Microsoft AI" — Ben Thompson

Why this is in the vault

Thompson's strategic read on where agent-era compute actually lives (cloud hub, not local GPU) and on Microsoft's enterprise "own-your-own-model" pitch — both directly load-bearing for RDCO's COO-agent architecture and the chip/memory capital-cycle thesis. Captured for the framing, not the product news. Written from the rendered Gmail subscriber body (full fidelity, including embedded Nadella/Huang/Suleyman/Bathiche keynote quotes); Thompson published this Update ahead of his Nadella interview, so it's his pre-interview opinion.

The core argument

1. The Nvidia AI PC (RTX Spark / "N1X") — underwhelming, a relic of the 2023 chatbot era. Nvidia unveiled an Arm-CPU + Blackwell-GPU PC superchip (128GB unified memory, ~300 GB/s bandwidth, made with Microsoft, shipping fall 2026 on Dell/HP/ASUS/Lenovo/MSI Windows machines). Thompson's verdict: the chip is built for the wrong era. The AI workload moved through three phases — ChatGPT era (local inference exciting) → reasoning era (KV-cache blowup demands more memory + decode) → agentic era (CPU performance is what matters). The ideal local-agent setup is strong local CPU plus calling out to the cloud for inference. RTX Spark instead burns die area on GPU cores that are inferior to the cloud (memory size/bandwidth) at the expense of CPU. Net: it's a "chatbot circa 2023" chip, hard to justify on price or the Windows-on-ARM software compromises in 2026. He reads Nadella's flat, low-enthusiasm Windows segment at Build as quiet agreement: local inference is nice-to-have, not where the AI that matters lives. Nadella has no loyalty to Windows (Thompson credits him with "The End of Windows," 2018 — ending it as the org's organizing principle, not as a product).

2. Project Solara — the genuinely compelling bet: cloud as hub, devices as spokes. Microsoft's quiet platform for devices that run AI agents instead of apps — Android-based (not Windows), two hardware designs shown, Qualcomm + MediaTek as chip partners, enterprise-focused, pilot partners lined up. Still vaporware, and very much in Microsoft's self-interest (it doesn't control a phone), so grain-of-salt applies. But the underlying model is the insight Thompson endorses independent of whether Solara ships: invert the topology so the cloud is the hub and many devices are spokes, rather than the phone at the center. Rationale — wearables fail on interaction model (only useful with a human in the loop, which is annoying/inefficient); the right pattern is a brief human interaction, then an agent doing the work in the background in the cloud. "The next computer is not one device" — it's a constellation working as one system, agents surfacing closer to where/when needed. Classic Microsoft move: new function (agents) → new form factors → a new platform Microsoft owns. Fits his "thin is in" thesis: when intelligence is in the cloud, the edge device gets thin. Strongest in enterprise, where context and compute already sit in the cloud.

3. Microsoft AI (MAI models) — the strategic framing is "own the full stack," not benchmark supremacy. Microsoft's Superintelligence Team unveiled 7 from-scratch models (flagship MAI-Thinking-1, claimed parity with Claude Sonnet 4.6 in blind testing, matching Opus 4.6 on a coding benchmark, trained with no distillation — a clean-data-lineage pitch). Thompson's read: the models being "pretty decent" matters less than the framing. Via Frontier Tuning + RLEs (reinforcement learning environments / training gyms), enterprises build company- and task-specific agents on MAI models that they own and control — Suleyman: you don't "rent intelligence from a shared model that learns from everybody"; your tuned model and data become "your moat." Microsoft cited a MAI-tuned Excel agent on par with GPT-5.4 at ~10x cost efficiency, and a McKinsey-tuned model beating GPT-5.5. Thompson compares it to AWS's Nova Forge (enterprise data added at a pre-training checkpoint; MAI is more RL-flavored, but the lines blur). The open question he flags: will enterprises choosing the own-your-model route get penalized for not being on the cutting edge of capability? His punchline — helping cautious enterprises adopt the future on their own terms without having to win on raw performance "is exactly how Microsoft has long maintained its position."

Mapping against Ray Data Co

Mapping strength: strong. Three direct hits across RDCO's active lenses.

• Agent architecture (where compute lives) — reinforces RDCO's existing posture. Thompson's "strong local CPU + call the cloud for inference" is precisely the COO-agent's shape: a thin always-on Mac Mini orchestration layer doing CPU-bound tool-calling/glue work, with the heavy reasoning happening in the cloud (Claude). This validates the existing build rather than challenging it — local GPU for inference is a distraction; the bottleneck/value is orchestration + tool access, which echoes the L5 north-star focus on unhobbling the agent's toolset and visibility, not on local model hosting.

• Cloud-hub / device-spoke topology — reinforces the channels-as-surface design. RDCO already runs the agent as a cloud-resident hub reachable via thin spokes (iMessage, Discord, HQ). Thompson independently arrives at "the cloud is the hub, devices are spokes, the phone shouldn't be the center" as the correct agent-era topology. That's confirmation the current bidirectional-channel architecture is structurally right, not a workaround. The wearable critique (only useful with a human in the loop) is also a useful design check on any future RDCO surface: prefer brief-interaction-then-background-agent-work over attention-demanding always-in-the-loop UX.

• Who-captures-value / integration-vs-modularity — surfaces a sharpening, not a contradiction. The MAI "own your model, your RLE is your moat" pitch is the enterprise mirror of RDCO's own moat logic: the durable asset isn't the frontier model (rented, commoditizing) but the accumulated workflows, institutional data, and tuned-on-your-tasks behavior. RDCO's moat is the vault + codified SOPs + the agent's accreted task-specific competence — the analog of an enterprise RLE. Thompson's open question ("will own-your-model enterprises be penalized for falling off the cutting edge?") is a live risk RDCO should hold: over-investing in a frozen, self-tuned stack can strand you off the frontier. RDCO's current answer — stay on the frontier model (Claude) as the reasoning engine while the moat lives in the harness/vault layer above it — is the better-hedged version of the same bet.

• AI-infra / chip capital cycle — a useful demand-side signal, mildly cross-cutting. Thompson reading the Nvidia consumer PC chip as mistargeted (GPU die spent where CPU is needed) is a reminder that not all "AI silicon" demand is equal: the agentic era pulls demand toward CPU + cloud-side memory bandwidth, and the local-inference AI-PC narrative may be weaker than the hardware roadmaps assume. Doesn't contradict the chip-fab/memory thesis (data-center memory + bandwidth demand is exactly what the reasoning/agentic era amplifies), but it does argue against treating the consumer AI-PC refresh as a load-bearing leg of the demand story.

Related

• [[2026-06-02-stratechery-google-capital-company]]
• [[2026-02-17-stratechery-thin-is-in]]
• [[2026-02-03-stratechery-microsoft-software-survival]]
• [[2026-01-07-stratechery-nvidia-ces-vera-rubin]]
• [[2026-02-23-every-chatgpt-memory-context-rot]]
• [[project_l5_north_star_strategic_direction]]
• [[project_investing_markov_capital_cycle]]