06-reference

stratechery nvidia earnings dollars per gigawatt hugging face

2026-09-01·reference·source: Stratechery·by Ben Thompson
stratecherynvidiaearningscapexdollars-per-gigawattopen-sourcehugging-facecapital-cycleinvestingphdata

Nvidia Earnings, Dollars Per Gigawatt, Open and Hugging Face (Stratechery Update, 2026-09-01)

Why this is in the vault

A fresh, direct data point on Nvidia's Q2 FY2027 earnings call (Aug 2026) — supply-constrained 70% forward growth guidance, the $/GW framing management used to justify TAM expansion beyond hyperscalers, and Nvidia's reported (unconfirmed) $12.9B move to acquire Hugging Face — all load-bearing evidence for RDCO's active memory/chip-fab capital-cycle investing thesis.

The core argument

Nvidia projected 70% sales growth for its next fiscal year (beating the ~45% analyst consensus), sending shares up 8.7% in the biggest one-day rally since April 2025. Thompson's throughline: Nvidia's results remain "boring" because they are a function of supply, not demand — CFO Colette Kress said customer forecasts point to growth "doubling next year," and CEO Jensen Huang confirmed the company deliberately gave a full-year (not just quarterly) guide because building AI infrastructure now requires securing land, power, and shell 2-3 years out, forcing Nvidia upstream into the power/land/cooling supply chain itself. Thompson repeats his standing analogy: Nvidia today is like Apple in the early iPhone years — revenue is capped by what the company can manufacture, not by what the market wants, and true demand won't be visible until supply becomes abundant (no sign of that yet, four years past ChatGPT's launch).

Second thread, "Dollars Per Gigawatt": management reframed growth as revenue captured per gigawatt of data-center buildout — $18B/GW at Hopper, $25B at Blackwell, $40B with Vera Rubin (now including Vera CPU, Rubin GPU, NVLink, InfiniBand/Ethernet, and the newly announced Groq LPU) — and argued this integration captures TAM disproportionately from non-hyperscaler buyers (sovereign AI, neoclouds, enterprises — the latter half of Nvidia's business, growing 100%/year) who lack the in-house skill to mix-and-match components themselves. Thompson notes the one credible threat to this margin structure is buyers with enough scale AND motive to disintermediate Nvidia's stack — hyperscalers, or frontier labs with high enough costs to justify building custom silicon (e.g., OpenAI's in-development chip). He calls Huang's answer defending against that threat "not particularly compelling" — a scattershot of five different justifications — and reads Huang's pivot to touting Nvidia's equity stakes in OpenAI and Anthropic as a tell that the direct competitive answer was weak.

Third thread, "Open and Hugging Face": Huang argued open-weight models are structurally good for Nvidia because CUDA's fungibility lets it capture nearly all open-model inference regardless of who wins the open-vs-closed race — as long as the market stays fragmented, Nvidia wins. Thompson connects this to The Information's report (unconfirmed by Nvidia) that Nvidia agreed to buy Hugging Face for $12.9B. He dismisses elaborate financial-engineering explanations (using it to backstop DGX Cloud commitments or absorb excess compute) in favor of a simpler read: Hugging Face is open source AI's Schelling point the way GitHub was for open source code, and Nvidia — like Microsoft with GitHub — doesn't need a clean revenue model to justify owning the piece of infrastructure that keeps the ecosystem it depends on alive.

Mapping against Ray Data Co

Concrete connection: the $/GW framing is a direct, quotable anchor for [[project_investing_markov_capital_cycle]] (chip-fab/memory capital-cycle thesis) — it gives a numeric ladder ($18B → $25B → $40B per gigawatt across Hopper/Blackwell/Vera Rubin) for tracking whether Nvidia's TAM-capture argument is holding up quarter over quarter, which is exactly the kind of anchor data investing-edgar-watch and investing-label-historical-phases are built to consume. The "boring because supply-constrained" argument is also a direct extension of the vendor-financing/capital-cycle risk thread already logged from "Nvidia's Risky Business" (2026-08-11) and the OpenAI/SB Energy Ohio deal (2026-08-18 note): Nvidia is now underwriting land/power/shell years in advance, meaning its 70% guide is itself a forward commitment on capex that has to show up in hyperscaler 10-Qs — a testable claim for the next investing-edgar-watch cycle. Secondarily, the Hugging Face read (buy the Schelling point, not the revenue model) is a useful cross-domain pattern for RDCO's own DSA/TAL positioning conversations: infrastructure moats get built by owning the point of developer gravity, not by mandating downstream tool choice — relevant framing for phData conversations about platform lock-in versus openness.

Related