"Thank God For Data Centers" — Packy McCormick
Why this is in the vault
McCormick's "Buyer of Capabilities" frame is a clean intellectual spine for the RDCO Power Cycle thesis: it explains why hyperscaler capex de-risks the hard-tech / energy layer faster than markets price, by analogy to how Apollo/DoD procurement pulled the integrated circuit down its cost curve. Directly reinforces the innermost-loop AI-infra thesis and the SpaceX/data-centers-in-space note filed the same day.
⚠️ Sponsorship & bias disclosure
Two separate layers, disclosed independently:
- Explicit paid sponsor — Deel. The issue carries a Deel ad block (Employer of Record / global hiring). Standard Not Boring sponsor placement, unrelated to the essay's data-center argument. No bias on the thesis content itself.
- Portfolio-adjacency ambiguity (Not Boring Capital). This is a single-author Packy essay — NOT the co-written "founder hands the keyboard" pattern, so no GTM-collateral flag on that axis. However, the essay is enthusiastic about the exact hard-tech categories Not Boring Capital invests in (enhanced geothermal, nuclear, solid-state transformers, silicon photonics, modular construction, HVDC). Verification result: Not Boring Capital does not publish a public holdings list — LP updates are redacted to protect private-company info, so individual positions can't be confirmed or ruled out. Treat the category enthusiasm as thesis-aligned advocacy from an investor talking his book, not neutral analysis. No specific named companies are pitched (categories only), which keeps it on the softer end of the disclosure-ambiguity spectrum.
The core argument
There is a large pool of technologies that could beat today's incumbents but are stuck in a local maximum: they need scale and a learning curve to get cheap, and in normal times no buyer will stick their neck out to fund that scaling. Historically two mechanisms break the stalemate — Alpha Products (e.g. the Sony Handycam pulling lithium-ion batteries down the curve; the calculator for microcontrollers) and extraeconomic Buyers of Capabilities (DoD/NASA buying capability almost irrespective of price).
McCormick's thesis: AI data centers are now a third kind of Buyer of Capabilities — "a commercial analog operating on DoD-style procurement logic but commercial timescales." With backlogs in every traditional input, developers have a near-bottomless bid and will pay up for anything that delivers fast, giving novel tech a shot at scale and cost-down years earlier than otherwise. The list of beneficiaries extends well beyond GPUs/DRAM to supersonic turbines, enhanced geothermal, modular construction, HVDC, solid-state transformers, silicon photonics, optical fiber, lasers, batteries, and nuclear. This funding is dilution-free (real revenue, negative working-capital cycle) and rescues companies that would otherwise die in the Valley of Death.
Supporting arc:
- Apollo analogy. The Moon program was unpopular in its time (1961 Gallup: only 33% support), yet its absurd ambition + bottomless budget incidentally pulled forward a long list of terrestrial technologies. DoD procurement is "the closest thing the US has to a national industrial policy," and funded nearly every general-purpose technology of the past century.
- The integrated circuit (cleanest case study). Fairchild's first big customer was the military (Minuteman); NASA's Apollo Guidance Computer became the anchor demand. Noyce slashed IC prices 87.5% ($120→$15, later to $1) while charging NASA premium prices to fund scale, betting volume would crack the commercial market. By 1963 MIT alone consumed 60% of US IC production; Moore's Law followed. Data centers may be the 2026 analog.
- Scale of the bid. Western hyperscalers/labs/neoclouds spend ~$750B this year, $1T+ next; Goldman pegs AI capex at $7.6T across 2026–2031. This year's spend is ~2.4% of US GDP — bigger than Manhattan Project (0.4%), Apollo (0.4%), or the late-90s telecom bubble (1.2%). One data point: Anthropic's reported ~$15B/yr to SpaceX for two Colossus data centers is ~60% of NASA's annual budget.
(Email body truncated at the paywall after the GDP-context section; argument above is complete through the thesis statement.)
Mapping against Ray Data Co
Strong. This is a higher-resolution version of the mechanism the Power Cycle thesis already bets on.
- Power Cycle thesis ([[2026-05-17-power-cycle-v1]], [[2026-05-18-pressure-test-and-winners-survey]]). Power Cycle's structural claim is that AI-infra capex is the picks-and-shovels demand pulling the power/energy layer forward. McCormick supplies the why it works engine — the "Buyer of Capabilities on commercial timescales" frame — and the historical proof-of-mechanism (IC cost curve). Useful as conviction-support, but note the pressure-test caveat already in the vault: the thesis is most contagious (and least alpha-rich) exactly where retail piles into the obvious pure-plays. McCormick's enthusiasm is the contagious-narrative side; weight accordingly.
- Innermost-loop AI-infra thesis ([[2026-05-12-innermost-loop-ai-infrastructure]]). McCormick's beneficiary list (HVDC, solid-state transformers, geothermal, nuclear, photonics) overlaps the innermost-loop's data-center-infrastructure + power layers (Vertiv, GE Vernova, Bloom, Oklo, Fluence; Talen, Constellation, Vistra). The "bottleneck migrated from chips to power" framing in that thesis is the same insight McCormick generalizes into "data center as meta-Alpha-Product."
- SpaceX / data-centers-in-space ([[2026-05-27-stratechery-spacex-ipo-data-centers-in-space]], filed same day). Direct overlap — both cite the Anthropic↔SpaceX/Colossus data-center spend, both treat AI compute demand as the forcing function for otherwise-uneconomic hard infrastructure. McCormick frames it terrestrially (data center as procurement engine), Thompson frames the orbital escape hatch. Read as a pair.
- Hyperscaler-capex anchor. McCormick's $750B-2026 / $1T-2027 / $7.6T-cumulative numbers (Goldman) are quotable capex-anchor context; the
investing-edgar-watchskill computes the bottom-up version from 10-Q filings. (Note: theinvesting/anchors/hyperscaler-capex/directory was not yet populated as of this filing — flag for the next edgar-watch run.)
Net: file as thesis-reinforcing narrative and a clean explanatory frame to cite, not as new evidence. The mechanism is sound; the conviction is an investor's.
Related
- [[2026-05-17-power-cycle-v1]] — the RDCO thesis this essay's mechanism most directly supports
- [[2026-05-18-pressure-test-and-winners-survey]] — Power Cycle hype-vs-structural pressure test (the skeptical counterweight)
- [[2026-05-12-innermost-loop-ai-infrastructure]] — chips→power bottleneck migration; overlapping beneficiary list
- [[2026-05-27-stratechery-spacex-ipo-data-centers-in-space]] — companion same-day note; shared Anthropic/SpaceX data-center anchor
- [[2026-01-08-stratechery-interview-power-for-ai]] — hyperscalers building own on-site power generation