06-reference

stratechery bajarin apple ai compute

2026-06-11·reference·source: Stratechery·by Ben Thompson (interview with Ben Bajarin, Creative Strategies)

"An Interview with Ben Bajarin About Apple, AI, and Compute" — @benthompson

Interview with Ben Bajarin, CEO/Principal Analyst at Creative Strategies (semiconductor-focused market research, co-host of The Circuit podcast). Third Bajarin appearance (prior: Sep 2024, Oct 2025). Covers WWDC 2026, Apple's cloud-AI architecture, Nvidia's PC chip, agentic CPUs in the data center, Intel's reversal, and the compute capacity shortage.

Why this is in the vault

Direct feed for two active RDCO threads: the chip-fab/memory capital-cycle thesis (this is one of the densest supply-demand discussions Stratechery has run this year) and hyperscaler compute economics (a genuinely new claim about why Google/Amazon are shifting internal workloads off their own ASICs). Pulled from the full email body via Gmail; no web reconstruction needed.

The core argument

WWDC = platform re-architecture, not features. Bajarin's read: Apple spent the keynote on "boring" plumbing (CPU scheduler, a rebuilt system index) because the 2024 Apple Intelligence promises required re-architecting the platform first. Thompson's framing: Apple is "perfecting a 2024 AI story" — the Siri demos that work are fancy search, and that is what consumers actually want.

Consumers won't pay for agents. Thompson's recurring thesis, restated hard: enterprises pay for productivity, consumers pay for entertainment (Netflix, not Dropbox Carousel; OpenAI as "the modern Dropbox"). Bajarin partially dissents via Jobs to Be Done — Siri as a consumer "control plane" that hires out small outcomes — but both agree the endgame is human-in-the-loop, not autonomous consumer agents. Both think Apple shipping good-enough on-device AI makes OpenAI's device-level consumer ambitions "very, very difficult, if not nil."

Apple's cloud architecture: Gemini base + Google Cloud + Nvidia. On-device Apple Foundation Models are Apple's own; the cloud Pro model is base Gemini with Apple post-training. Inference runs on Nvidia GPUs (both lean "largely H100s — nothing Apple does needs a GB200") in Google Cloud, with Intel as confidential-compute head node. Thompson's take: Apple Silicon in the cloud is effectively dead; Nvidia-on-GCP preserves portability to AWS/Azure, whereas TPUs would lock Apple to Google. The ~$1B/yr Gemini deal makes sense once you see Google landed a massive GCP customer.

The ASIC lock-in mechanic (the standout claim). Bajarin, from Creative Strategies' customer-chain work: Google is deliberately moving internal workloads onto GPUs (including SpaceX/neocloud capacity) so it can offer scarce TPU capacity to third parties at priority pricing — because once a customer's stack runs on TPUs (or Trainium, where Amazon mirrors the play), "you're stuck." The capacity crunch is being used as a customer-acquisition weapon for custom silicon. Thompson called it mind-blowing; it reframes the Google–SpaceX GPU deal as TPU go-to-market, not GPU shortage relief.

Agentic inference brings back the CPU. "Humans click, agents swarm" (Jeetu Patel, Cisco): agent orchestration is CPU work, so expect dedicated CPU racks beside GPU racks (Nvidia Vera, Arm selling chips, AMD/Qualcomm following). Concurrency-per-megawatt ("cores per megawatt") becomes the metric. Bajarin: CPU inference is already happening more than people realize because GPUs are consumed by training. Software moats (maybe even CUDA) erode when agents can rewrite your software cheaply.

The capacity shortage was seeable and everyone missed it. TSMC slowed capex growth post-ChatGPT; memory makers underbuilt for years and new capacity takes ~2 years to arrive. The industry's overcapacity scars beat the Samsung/Morris Chang lesson (the two most valuable fab franchises were built by investing into downturns). Bajarin's estimate: TSMC could have used "five more foundries" and still not met demand. Consequence: TSMC created its own competition — Intel will get foundry customers ("a matter of when, not if"), and Intel's EMIB advanced packaging (mix-and-match tiles, capacity where CoWoS is rationed) is becoming a multi-billion business. Supply-demand does not balance "before 2030," possibly 2035.

Is AI software or graphics? Bajarin's framing for the cycle-length question: if AI is like internet/client software, it gets "good enough" and infrastructure spend plateaus (bearish). If it's like graphics — an industry that has never had enough compute, chasing fidelity for 30 years, "tokens and pixels are the same thing" (Jensen's insight) — the buildout runs decades. Bajarin leans graphics.

Two notable bear-case asides from Thompson: (1) memory — 40 years of never optimizing for memory means huge low-hanging fruit for demand-side optimization "to the detriment of the memory providers"; (2) hyperscaler-as-meter dominance could leak if token generation migrates to the edge/on-prem (OpEx vs CapEx logic for enterprises with token budgets).

Sponsor/bias scan

No third-party sponsors; Stratechery is subscription-funded. Standard podcast/Stratechery Plus plugs in footer — not structurally biasing. Perspective note: the closing segment is a friendly promo for The Diligence Stack, Creative Strategies' new $300/mo research Substack. No commercial relationship with Stratechery disclosed, so not a sponsor flag — but Bajarin's authority claims ("vast number of conversations with CIOs," "we hear in the customer chain") are also the sales pitch for his paid product. His customer-chain claims (e.g., the TPU lock-in mechanic) are plausible, sourced from non-public client work, and unverifiable from here — treat as informed analyst signal, not confirmed fact.

Mapping against Ray Data Co

Related