SYNAPSE

模型學會「思考」之後，全球約三分之二的 AI 算力已花在推理上。每一分智能都以算力計價——推理基礎設施，成為新的主戰場。 Once models learned to think at run-time, roughly two-thirds of the world's AI compute moved to inference. Every unit of intelligence is metered compute — inference infrastructure is the new battleground.

一次任務、上百次調用，Agent 消耗的 token 以數十倍計。機器間的流量將遠超人類點擊——而這一切，都由中樞承載。 One task, hundreds of chained calls — agents burn tokens at tens of times human rates. Machine-to-machine traffic will dwarf human clicks — and the nexus carries all of it.

前沿研究與生產實踐指向同一結論：按任務把請求路由到最合適的模型，比押注任何單一模型更快、更省、更穩。 Research and production point the same way: routing each request to the right model beats betting on any single one — faster, cheaper, steadier.

數據主權已是董事會議題，監管正把三成以上的 AI 支出推向本地化。能把前沿能力裝進客戶邊界的人，拿下下一個十年。 Data sovereignty is now a boardroom issue, steering a third of AI spending toward localized platforms. Whoever puts frontier capability inside the customer's perimeter owns the next decade.