"Models Become Utilities" Just Went From Metaphor to Price Sheet — Facts: announced June 29, DeepSeek V4's full release (industry dailies report it live July 15) introduces time-of-day API pricing: rates double during 9:00-12:00 and 14:00-18:00 Beijing time, off-peak unchanged — the first mainstream frontier lab to adopt peak-valley pricing at the API layer; the timing is telling: on May 22 it converted a limited-time 60%-off promo into permanent pricing, and a month later claws the peak-hour price back via time-of-day rates; dailies also report 60-80% faster inference and deep Ascend 910B/950 adaptation, plus unconfirmed IPO-prep chatter. Judgment: (1) her March call — "models go from moat to utility" — just got fulfilled literally: supply-side commoditization (March) → telco-like share structure (June 21) → demand-side quotas (Tesla's $200 meter, July 6) → supply-side time-of-day rates (today): metering, quotas, and time-of-day pricing — the utility trifecta is complete; time-of-day pricing appears only when the cost structure has become grid-like (fixed capacity, load peaks, near-zero marginal cost of idle capacity), so for the first time a model company admits it sells capacity utilization, not intelligence; (2) intelligence now has a time-of-day attribute: enterprises will shift non-realtime workloads (batch refactoring, data cleaning, regression tests, content generation) to off-peak — after "cost-switchable" (ENTRY 95) comes "time-switchable"; the precedent is AWS spot instances at ~70% off spawning an entire scheduling-tools layer — a token-scheduling / inference load-balancing layer is a fundable product today (an agent framework with built-in "queue at peak, run at trough" scheduling saves clients a large slice of their inference bill); (3) DeepSeek's play is smarter than it looks: nominal prices unchanged, the May price cut quietly recovered at peak hours, and price leverage flattens its load curve — saving GPU capex; combined with Ascend adaptation and mature private deployment, its listing narrative is crystallizing as "China's utility-grade intelligence supplier" — the first model company to voluntarily price like a telco, which sharpens the question her June 21 entry left open: at IPO, does the market anchor on a telco P/E or a platform P/S? Rollback: IPO timeline and valuation figures are aggregator-sourced and unconfirmed — only the pricing mechanism is treated as hard fact; if no second major lab adopts time-of-day pricing within three months, this is DeepSeek's load management, not an industry inflection — she'll downgrade the call; and whether a 2x peak-valley spread actually moves enterprise workloads is untested. Watch: whether ByteDance/Alibaba/Zhipu/Kimi follow, and the night-time share of token consumption.