The Investor

Today's repricings share one mechanism: revenue existed because something was scarce, restricted or subsidized, the condition ended, and the mark still capitalizes it. The cleanest print is Shein's Hong Kong listing at roughly $26B — 73% below the $98B private valuation it carried in 2022. Watsco sits more than 45% below peak on the same logic.

a16z crypto's post-crisis dataset: the gap between the official dollar and the crypto dollar collapsed from over 100% in 2023 to roughly 4% as of August 28, 2026, after Argentina lifted most dollar-purchase restrictions in April 2025. The deep dive below separates the macro-levered series from the product-levered one, and names what to underwrite instead of spread capture.

Tencent open-sourced Hy4 under Apache 2.0 — 770B total parameters, roughly 49B active, 1M+ token context — with hosted inference quoted at $0.834 per million input tokens. Andrew Ng separately released OpenWorker under MIT, a local-first desktop agent with 25+ integrations that runs free on local Ollama. The sources date neither release, so treat both as standing conditions rather than this week's news. Frontier-class agentic coding is now free to own and cheap to rent. Any app-layer mark underwritten on model access or a proprietary desktop-agent feature set is pricing a moat that two permissive licenses already gave away.

Netflix's new four-year NFL deal takes it from two games to five, including two Christmas games and a Rams-49ers game in Australia, against a reported ~$75M per game under the prior arrangement, per Morning Brew. The league minted those windows to fill the order, so aggregate spend rises while per-slot scarcity erodes. The same Morning Brew item reports Pac-12 annual TV fees fell from $381M to $3M after ten teams left, with the ACC named as the next candidate. Volume-levered exposure holds up better than scarcity-levered exposure.

An engineer who ran an exabyte-scale storage system for eight years reports his daily dashboard shrank to seven numbers, surfaced in Lex Neva's reliability roundup. That is the mature-buyer end state, and it is hostile to per-gigabyte and per-metric pricing that assumes telemetry compounds forever. Separately, PlanetScale — the industry's most Vitess-committed vendor — now ships Postgres on dedicated NVMe inside AWS and GCP. Price Postgres as the default engine, and split observability expansion into new workloads versus telemetry per workload.