DeepSeek——這家以低成本、高性能模型震驚矽谷的新創公司——突然宣佈將「大幅」調漲價格，此舉可能削弱中國人工智慧產業的競爭力。過去近兩年間，業界驚嘆於中國模型在快速演進的同時卻能維持極低定價，彷彿無視經濟引力。其價格底線之激進，被形容為美國AI模型難以生存的「死亡區」。DeepSeek的超低費率之後，Alibaba、Z.ai與Moonshot等公司也相繼推出價格親民的前沿級中國模型，進一步加劇了外界對中國已破解AI經濟密碼、可能顛覆美國高成本模型領先地位的憂慮。其定價策略被比作中國低價傾銷太陽能板、衝擊國際市場價格的先例。（關鍵數字：nearly two years）

然而，本週的漲價公告表明，這種策略在規模化後可能難以為繼。此事影響深遠，因為DeepSeek的任何價格變動都可能重新定義整個中國AI產業的定價格局。這家位於杭州的AI實驗室及其同業的廉價token，不僅對正朝萬億美元估值公開上市邁進的OpenAI和Anthropic構成競爭威脅，其低成本更補貼了一整個下游開發者與應用建構者生態系統——這些人在設計產品時，預設推論成本會持續維持在低位。DeepSeek的模型採開放權重發佈，任何人皆可下載、修改並在本地運行，這與其美國競爭對手截然不同，在一定程度上為使用者提供了更多AI控制權，但並未改變成本正在上升的事實。

DeepSeek低調的創辦人Liang Wenfeng及其國內競爭者，終究必須像西方同業一樣面對成本問題。對於那些圍繞高性能但低價AI設計產品的開發者和新創公司而言，這是一記警鐘：依賴政策隨時可能變動的第三方平台存在固有風險。這也提醒世人——經濟引力終將追上所有人。

DeepSeek, the startup that stunned Silicon Valley with its low-cost, capable models, abruptly announced "significant" price increases, a move that could blunt the competitiveness of the hard-charging Chinese AI industry. For nearly two years, the industry marveled at how Chinese models evolved rapidly while remaining extraordinarily cheap, as if defying gravity. The pricing floor was so aggressive it created a "death zone" for US AI models. DeepSeek's rock-bottom rates were followed by affordably priced, frontier-level Chinese models from Alibaba, Z.ai, and Moonshot, intensifying concerns that China had cracked the AI economics code and could disrupt the leadership of costlier American models. The pricing strategy drew comparisons to Chinese "dumping" of low-cost solar panels that collapsed international market prices.

However, this week's price hike announcement signals that approach may not be sustainable at scale. The implications are significant because any changes to DeepSeek's pricing are likely to reset pricing across the entire Chinese AI industry. The Hangzhou-based AI lab and its peers' cheap tokens were not merely a competitive threat to OpenAI and Anthropic—both heading toward public listings at trillion-dollar valuations—but also subsidized an ecosystem of downstream builders and developers who priced their own products assuming inference costs would remain low. DeepSeek's models are open weight, meaning anyone can download, modify, and run them locally, unlike their American counterparts. This gives users more control over their AI, somewhat reducing the sting, but it does not change the fact that costs are rising.

DeepSeek's reticent founder, Liang Wenfeng, along with his local competitors, will eventually have to account for their costs, just like their Western counterparts. For developers and startups that designed their products around capable but cheap AI, this serves as a stark reminder of the risks inherent in relying on third-party platforms whose policies can change at any time. It is also a reminder that gravity catches up with everyone.