Nvidia Earnings, The AI Stack, Nvidia's New Reporting

Why this is in the vault

Thompson reads Nvidia's Q1 (April-quarter) earnings through a business-model lens, and the load-bearing event for RDCO is a reporting-structure change: Nvidia split Data Center revenue into hyperscale vs ACIE (AI Clouds, Industrial, Enterprise). That split is a direct map onto the central tension in our AI-infra investing theses — where Nvidia fights commoditization (hyperscalers building their own chips/networking) vs where it sells and locks in the full vertically-integrated stack (everyone else). The note also surfaces a fresh data point on the Anthropic-on-Nvidia reversal and the Vera CPU TAM claim. This is investing-thesis evidence, not slop.

The core argument

• Earnings were predictable in a supply-constrained world. Record ~$81.6B April-quarter sales (up 85% YoY), net income ~$58.3B, networking tripled YoY to ~$14.8B. Two shareholder-friendly moves: $80B buyback and a dividend hike (1c to 25c/quarter), with intent to return 50% of FCF this year. Demand "gone parabolic," attributed to agentic AI.
• Nvidia's moats are real but not unassailable. Thompson lists three: (1) CUDA lock-in — but most app developers operate above the layer where CUDA matters, and Anthropic + Google have proven CUDA isn't essential to train/serve a frontier model on TPU/Trainium; (2) first-in-line supply from TSMC logic+packaging and memory — durable short-to-medium term, erodes as supply comes online; (3) selling networking and full systems, not just chips.
• The frontier-lab wrinkle. OpenAI signed a Trainium deal but for now trains/infers on Nvidia, so more OpenAI usage = more Nvidia. Anthropic is the opposite: Nvidia's coverage of Anthropic was "largely zero" because Google/Amazon investments tied it to TPU/Trainium — but that is now reversing fast (Nvidia securing Anthropic capacity across Azure, AWS, CoreWeave). Thompson frames the prior near-zero Anthropic exposure as an under-appreciated bullish factor (Nvidia grew without the fastest-growing lab) while conceding it shows the position isn't unassailable.
• The reporting change is the real signal. Hyperscalers are ~50% of Data Center revenue and are "frenemies" — the players most motivated and able to reduce Nvidia to a commodity chip vendor (custom silicon, forcing Nvidia chips into their own networking). The new ACIE submarket (neoclouds, 250k enterprises, industrial/edge, sovereign AI) is the segment that is, relative to hyperscalers, already commoditized in Nvidia's favor: these buyers take whole systems, don't roll their own chips/networking, and have no real alternative. ACIE's share of sales has been rising over the trailing 9 quarters.
• Thompson's verdict: the trend is bullish. Bear read: shrinking hyperscaler share = the most direct competitor for stack value is winning. Bull read (his): every point of share that sits in ACIE is current revenue plus multi-year lock-in across chips, networking, and full systems at high margin. He flags a caveat — much neocloud capacity ultimately serves the same hyperscalers/labs, so this is leverage-into-lock-in, not true demand diversification.
• Vera CPU. Custom-ARM CPU codesigned with Rubin/NVLink, claimed 1.5x perf/core, 2x perf/watt, 4x density/rack vs x86; a claimed new ~$200B TAM with ~$20B visibility this year. Thompson reads the natural customer as ACIE/neoclouds (already all-in on Nvidia), not hyperscalers (who have their own CPU programs + AMD/Intel ties). He treats "world's leading CPU supplier" as a long-run aspiration, not a $20B-implied claim.

Mapping against Ray Data Co

This maps strongly and is mostly confirming, with one useful refinement.

• Confirms the two-thesis split is the right frame. Our AI-infra bucket was deliberately split into [[2026-05-18-nvidia-supply-chain-v1]] (picks-and-shovels monetizing Nvidia's own revenue: CoWoS, HBM, foundry) and [[2026-05-18-nvidia-adjacent-v1]] (names competing for the same accelerator wallet). Thompson's hyperscale-vs-ACIE cut is the demand-side mirror of that supply-side split: ACIE = the lock-in/full-stack base that protects Nvidia revenue (good for supply-chain thesis durability); hyperscale = the contested-commoditization layer where AMD/Trainium/TPU pressure lives (the battleground the adjacent thesis is positioned around).
• New monitorable: the ACIE-vs-hyperscale share ratio. Going forward, Nvidia's own segment disclosure gives a quarterly read on whether commoditization pressure is winning (hyperscale share rising on custom-silicon substitution) or whether lock-in is winning (ACIE share rising). That is a cleaner, free signal than inferring from TrendForce/10-Q triangulation alone. Candidate to add as a tracked input to both theses' monitor lists.
• Refines, doesn't change, the bull case. Thompson's caveat is the important nuance for us: ACIE growth is partly neoclouds reselling capacity to the same hyperscalers/labs, so rising ACIE share is NOT independent demand diversification. Don't over-credit the segment split as de-risking; it is lock-in leverage, which still helps the supply-chain thesis (units + attach) but doesn't insulate against a frontier-lab/hyperscaler capex air-pocket. This sits alongside the demand-concentration risk already flagged in the [[2026-05-17-power-cycle-v1]] and hyperscaler-capex anchor work.
• Anthropic-on-Nvidia reversal is a thesis-relevant tape. Earlier coverage ([[2026-04-07-stratechery-anthropic-tpu-deal-google-alliance]]) had Anthropic moving toward TPU/Trainium; this note shows Nvidia clawing back Anthropic capacity. Net: the CUDA-isn't-essential structural risk is real, but Nvidia's commercial/allocation muscle is partially offsetting it in the supply-constrained window. Watch whether this holds once supply loosens.
• Vera CPU = optional adjacency, not a new thesis. The ~$200B TAM claim is aspirational; the credible near-term is ACIE attach. Note it, don't allocate on it.

Related

• [[2026-05-18-nvidia-supply-chain-v1]] — picks-and-shovels thesis; ACIE lock-in is the demand-side durability argument for it.
• [[2026-05-18-nvidia-adjacent-v1]] — competitor/parallel-track thesis; the hyperscale-commoditization battleground is exactly its terrain.
• [[2026-03-04-stratechery-anthropic-revenue-nvidia-earnings]] — prior Stratechery Nvidia-earnings read; provides the trend line on Anthropic revenue + Nvidia exposure.
• [[2026-05-17-power-cycle-v1]] — hyperscaler AI-capex magnitude/concentration context behind the whole stack.
• [[2026-05-04-stratechery-google-meta-earnings]] — adjacent Stratechery big-tech-earnings note on capex/agent-infra economics.