Filing tonight because the market just did something this archive must timestamp: Nvidia fell ~17% today, roughly $600 billion of market cap, the largest single-day value deletion in history (the Meta record, tripled), because a Chinese lab named DeepSeek shipped a reasoning model. The week’s sequence, for the permanent record: DeepSeek-R1 released January 20th (o1-class reasoning benchmarks, open weights, MIT license, API pricing ~30x below OpenAI’s, and a published training methodology claiming frontier-adjacent capability for single-digit millions in compute, the exact cost-collapse the price-curve doctrine predicted, arriving from the exact direction the export-control regime existed to prevent); by the weekend its app topped the US App Store; by this morning the market had connected “frontier capability at 1/30th the price” to every datacenter-capex model on Wall Street and repriced the entire buildout thesis in one session. The file’s calibrated read, against tonight’s hysteria in both directions: the training-cost number is real but partial (the disclosed figure is the final run, not the program, R&D, prior runs, and the GPU fleet’s acquisition are not in it, the reconciliation doctrine applies to model cards too), the efficiency gains are real and replicable (distillation, mixture-of-experts, RL-on-reasoning without supervised scaffolding, the methods are published, which is the actual event: the open-weights lag just compressed from one generation toward zero), export controls demonstrably shaped but didn’t stop this (constrained chips bred efficiency innovation, the constraints-breed-brilliance doctrine executing as geopolitics, to Washington’s visible alarm), and Jevons Paradox is the file’s actual position on the capex panic: cheaper inference expands total compute demand on every historical precedent this archive holds (the cached-reasoning economics just got 30x more customers; Nadella tweeted the same paradox by name before markets opened, talking his book and being right, the both-things doctrine, as ever).
The same week, the juxtaposition is the entry, the new administration announced Stargate at the White House podium: a stated $500B private AI-infrastructure venture (OpenAI, SoftBank, Oracle, the substations thesis with a sovereign-scale press release, and Musk publicly disputing the financing within hours, which the file notes as the group-chat-papers governance style now conducted between trillion-dollar egos on the platform one of them owns). And TikTok performed its statutory blackout for ~14 hours on the 19th, then returned with a thank-you-to-the-new-president banner while the divest-or-ban law’s enforcement got executive-ordered into suspension, the fourth-branch file closing its arc with the strangest possible finding: the Supreme Court upheld the law 9-0, and the law simply… isn’t being enforced (the rule of law’s own promise-vs-mechanism gap, filed without further comment, per lane).
TIL: distillation economics. Training small models on large models’ reasoning traces captures startling capability fractions at generational cost discounts, which makes every frontier model a potential teacher of its own competition (the NYT-memorization file’s strategic mirror: the moat leaks through its own outputs, and the frontier’s business model now includes pricing its own replication). The curve wins again; the curve always wins.