The Science of Foreshocks

研报英文原文证据摘录

The Science of Foreshocks

pends on a complex interplay between AI business models, demand for 105

memory, and alternative supplies. While there are other questions (e.g. 0

Keiretsu-style circular flow of spending, investment and revenues), these

are unlikely to be as critical. First, irrespective of immediate returns, AI labs US China

will likely continue their massive investment spree (i.e. ~US$850bn in '26 Source: OpenRouter; Macquarie Global Strategy

alone). Simply put, dangers of falling behind are perceived to be far greater

than overinvestment and lower LT returns. Second, there is evidence that Intelligence Index vs Price per 1m

Input tokens - China vs USadoption and monetization is running ~3x faster than recent technological

revolutions, already generating annualized revenues of US$175bn, clearing 65 Intelligence Index vs Price per 1m input token

the depreciation hurdle. Third, rising demand and fierce competition is 60 GPT 5.5 Claude Fable 5

translating into a huge contract backlog (>US$2trn) for at least three years 5550 GLM 5.2

while new chip capacity will take time. Four, however, there is also evidence 4540DeepSeek V4 Gemini 3.5

of users pushing back on costs (token, chips) by leveraging cheaper but 35 Grok

less capable China open-weight models while alternative (mostly Chinese) 3025

chipmakers are bridging performance gaps. Also, corporates (software to 0 0.5 1 1.5 2 2.5 3

biotech) are trying to slow their own commoditization by making access to US CHN

systems and databases more difficult. 3. Investors are facing a conundrum Source: OpenRouter; Macquarie Global Strategy

of revolutionary changes that create and destroy at lightening speeds by

API (US$ per 1m) - China vs USattacking "impregnable moats", eroding pricing power and values of those