06-reference

stratechery nvidia vs tpu broadcom miss

2026-06-08·reference·source: Stratechery·by Ben Thompson
chip-capital-cyclenvidiabroadcomtpuhyperscaler-capexapple-aiagent-deploymentinvesting-thesis

"Google Buys Compute From SpaceX, Broadcom's Outlook, Apple's AI Politics" — Ben Thompson

Why this is in the vault

Two of the three segments are direct reads on the chip-capital-cycle thesis — Thompson's net call is that both the Google/SpaceX compute deal and Broadcom's "miss" are quietly bullish for Nvidia and bearish for the TPU/custom-ASIC story.

The core argument

A three-topic Stratechery Update, with a single connecting thread Thompson states up front: the news flow is bullish for Nvidia.

  1. Google buys compute from SpaceX ($920M/month, ~$30B through mid-2029). Google's Cloud backlog nearly doubled QoQ to >$460B; this is framed as "bridge capacity" for Gemini Enterprise demand. Thompson dismisses the index-inclusion conspiracy (the deal conveniently gives SpaceX the profitability S&P 500 requires) as not Google's primary motive. His sharper read: a huge share of that backlog wants Nvidia chips specifically, not TPUs. The SpaceX deal is contingent on SpaceX delivering Nvidia chip access by Sept 30 or Google can terminate. He floats that the cluster could even be for training next-gen Gemini, since Nvidia's scale-up networking is better for Mixture-of-Expert training than TPUs.

  2. Broadcom's outlook disappointed. AI-chip guidance of $16B (vs ~$17.2B expected) for the fiscal Q3, $56B full-year vs higher estimates. Triggered a ~$1.3T chip-sector selloff Friday (PHLX chip index down 10.3%, worst day since March 2020), hitting Nvidia, Micron, AMD. CEO Hock Tan conceded Google will diversify its custom-chip sourcing (MediaTek, in-house). Thompson's key reframe: Broadcom's most important AI business is TPUs; the "miss" may not be an AI miss but an Nvidia win — TPUs are at risk of commoditization via a thin customer base (just Google + Anthropic), while Nvidia is the default for everyone else.

  3. Apple's AI politics (WWDC preview). Drawing on Gurman's reporting of a consequential early-2025 exec meeting (Federighi, Rockwell, Giannandrea) over the Apple Intelligence flop and Siri delay — internally framed as palace intrigue, with Rockwell volunteering as the AI "fixer." Thompson notes the people now in charge of AI at Apple are the ones who didn't know generative AI existed pre-ChatGPT. He argues Apple may be fine (low infra spend, devices needed regardless) but needs a software story. His three WWDC questions, all reducing to the control-vs-capability tradeoff: (a) will Apple let vibe-coded services deploy apps to an iPhone without the App Store; (b) will Apple open Siri App Intents to third-party AI services; (c) will macOS add provisions for agentic services like OpenClaw. His view: Apple should signal it won't stand in the way of users who want more capability.

Mapping against Ray Data Co

Strong reinforcement — investing / chip-capital-cycle thesis. This is the load-bearing connection. Our active thesis ([[2026-05-27-markov-equities-pipeline-spec]]) places RDCO at Phase 2 (capacity-announcements), and the Markov phase tracker's whole job is reading exactly these signals. Two things to register:

Medium — agent-deployer positioning. Thompson's three WWDC questions map almost one-to-one onto RDCO's agent-deployer interests, especially Q1 (deploy apps to iPhone without the App Store) and Q3 (macOS provisions for agentic services like OpenClaw). If Apple opens the operating system to third-party agentic capability, that expands the action-surface available to agent-deployers; if it doubles down on control, the App Store remains the chokepoint. This is the same control-vs-capability tension we track in the agent-action-surfaces cluster (cf. the Robinhood-via-MCP datapoint in the May digests). Worth watching the actual WWDC keynote outcome against these three questions.

Note on contradiction risk: none here — this Update extends rather than contradicts prior vault belief. It is the explicit continuation of the SpaceX/compute-fungibility argument in [[2026-05-12-stratechery-spacex-anthropic-xais-two-companies]] (model labs treating each other's compute as fungible) and the Cursor angle in [[2026-04-22-stratechery-john-ternus-spacexai-cursor]].

Related