06-reference

stratechery nvidia openai datacenter anthropic spirit data

2026-08-18·reference·source: Stratechery·by Ben Thompson
stratecherynvidiaopenaianthropicgooglecapital-cyclehyperscaler-capexdata-valueinvestingphdata

Nvidia Backs OpenAI Data Center, Anthropic News, Google Buys Spirit Airlines Data (Stratechery Update, 2026-08-18)

Why this is in the vault

Three items in one Update, each a live data point on RDCO's active theses: Nvidia's vendor-financing pattern (direct continuation of the 2026-08-11 "Nvidia's Risky Business" capital-cycle argument), Anthropic's revenue trajectory ($65B run-rate, up 7x YoY, IPO expected this fall — MAIN-bet dependency), and a concrete market test of the "data is oil" framing via Google's $10M bankruptcy-auction purchase of Spirit Airlines' internal data (100M emails, 500M Teams chats) for AI training use.

Issue contents

1. Nvidia Backs OpenAI Data Center

OpenAI signed a 10-gigawatt, 20-year data-center lease in Ohio with SoftBank's SB Energy, partly backed by Nvidia. Nvidia backstops the value of the completed data center (starting with the first ~5GW phase) so OpenAI can raise debt without exposing Nvidia to full project risk; in exchange Nvidia becomes exclusive chip supplier for the first half of the site and takes a $1.5B equity stake in SB Energy, with an option to expand later. This is a reworked version of a late-July rumor that Nvidia would backstop the entire $250B project — the market balked, so Nvidia narrowed to the data-center portion only, leaving chip-purchase risk with SoftBank. Thompson frames this as economically rational for Nvidia even though it expands bubble blast radius: if AI revenue growth is real, financing customers' capex now locks in future Nvidia demand; "no risk, no reward."

2. Anthropic News

Bloomberg reports Anthropic's annualized revenue run-rate surpassed $65B by end of July, up more than 7x from year-end 2025, ahead of an expected fall IPO (before OpenAI's). Thompson flags the growth is decelerating slightly in absolute-increment terms — April→May added ~$17B, May→July added $18B over a longer window — while still "amazing, but not perfect." Separately, Anthropic is reportedly in talks to buy Decart AI ($6B), an Israeli world-models startup, for its Decart Optimization Engine (claimed training/inference efficiency gains) — a departure from Anthropic's text/code focus that Thompson reads as Anthropic paying for efficiency expertise it hasn't had time to build in-house given how recently demand outstripped supply.

3. Google Buys Spirit Airlines Data

Google won a bankruptcy auction (beating Mercor) for $10M, acquiring de-identified Spirit Airlines data — 100M emails, 500M Teams messages, revenue/operations/audit/fraud records spanning 26 years — explicitly to improve its AI. Personal data (97.5M passenger profiles, 50.2M loyalty records) is excluded. Box CEO Aaron Levie invoked "data is oil" on X; Thompson counters with his own 2018 "Data Factories" argument (raw data isn't valuable until transformed — which is what training actually does) but grants Levie's underlying point: pre-LLM, no one would pay $10M for a defunct airline's Slack/Teams corpus. Fungible, undifferentiated corporate data now has standalone market value for training, independent of its original business context.

Mapping against Ray Data Co

Concrete connection: Thompson's "Data Factories" pushback — data isn't valuable until transformed into something useful — is the literal thesis embedded in Ray Data Co's own name and phData's DSA/TAL positioning; this issue is a live market test of that argument (Google paying real money for raw, undifferentiated corporate data specifically because transformation pipelines now exist to make it useful) and is directly quotable in phData conversations about why "we have lots of data" isn't itself a moat — the transformation layer is. Secondarily, this Update is another data point for project_l5_north_star_strategic_direction (phData = MAIN bet, downstream of Anthropic's capability trajectory): a $65B run-rate and looming IPO is bullish evidence for the bet's foundation, while the growth-deceleration nuance Thompson flags is a calibration input, not a green light — Ray should track it the same way project_investing_markov_capital_cycle tracks capex anchors, not as a headline win. The Nvidia/OpenAI Ohio financing structure is the same vendor-financing pattern already logged from "Nvidia's Risky Business" (2026-08-11) and belongs in the same investing capital-cycle file as further evidence of hyperscaler debt-financed buildout risk migrating to non-bank capital.

Related