Your newsletter digest — June 9, 2026

AI infrastructure: who's buying compute from whom

Google signs a $30 billion deal to rent compute from SpaceX

• Google described it as "bridge capacity" for Gemini Enterprise, whose agent demand has exceeded internal forecasts. The short-term framing matters: Google is not outsourcing its infrastructure strategy, it's buying time while its own buildout catches up.
• The deal is the second major compute rental SpaceX has signed in as many months — Anthropic took 300 megawatts at SpaceX's Colossus data centers in Tennessee for $1.25B/month in May. The two deals together add more than $26 billion to SpaceX's annual revenue run rate, conveniently timed a week before its $1.75 trillion IPO.
• Thompson's read: the Google deal, like Anthropic's before it, confirms that there is real demand for every GPU that gets built. That's conventionally bullish for Nvidia — the neutral beneficiary of every hyperscaler fighting over Nvidia-built infrastructure, regardless of who owns the data center.

Broadcom earnings: strong numbers, selloff anyway

• The selloff wasn't about the numbers — it was about what Broadcom didn't say. Management declined to raise its AI revenue guidance for fiscal 2027 (still targeting $100B), and disclosed that Alphabet is diversifying its TPU suppliers, bringing in MediaTek and ramping in-house silicon. Broadcom's Alphabet share could fall from ~95% today to ~65% by 2028.
• CNBC's analysts read this as profit-taking at extreme valuations, not a demand signal. Broadcom is still growing 50%+; the implied concern is that custom AI chip contracts are more fragile than they look when customers get big enough to build their own.
• The Nvidia connection: Broadcom's relative weakness is, paradoxically, evidence that Nvidia's model holds. Custom ASICs work for hyperscalers with predictable, at-scale workloads — but when customer concentration risk materializes this fast, the universal GPU case gets stronger, not weaker. Thompson's piece title says it plainly: both the SpaceX deal and Broadcom's earnings "seem bullish for Nvidia."

Apple at WWDC: the AI politics question

Apple introduces Siri AI — and a different bet on the stack

• The architecture is Apple's own: on-device foundation models plus Private Cloud Compute, with no user data stored and Apple-verified privacy claims. This is the part Thompson frames as "AI politics." While earlier speculation pointed to Claude or Gemini integrations for the new Siri, Apple shipped its own foundation-model stack — a quiet but consequential answer to the question of whether Apple routes its AI through third parties or owns the layer.
• The WWDC hardware caveat: iPhone 16+, iPhone 15 Pro/Max, and M1+ iPads/Macs get the core Siri AI experience; the best on-device model and improved voice/dictation requires iPhone Air or iPhone 17 Pro/Max or M4+ Mac with 12GB RAM. EU iOS/iPadOS rollout is delayed while Apple works through regulatory compliance; China is also excluded for now.
• Thompson's lens going in: he said he was watching for "what Apple is looking for at WWDC" around AI politics specifically. Apple's move to ship its own foundation models — rather than anchor the new Siri to OpenAI or Google — keeps the iPhone's AI value chain inside Apple's walls. Whether that bet holds as third-party models keep improving is the question that will define Apple's next few software cycles.

One thread to watch