Leaders | Seek and ye shall find
經濟學人社論 | 上下求索，終得其果
英文部分選自經濟學人20250201期社論版塊
Leaders | Seek and ye shall find
The real meaning of the DeepSeek drama
The Chinese model-maker has panicked investors. But it is good for the users of AI
這家中國AI模型公司引發了投資者的恐慌，但對於AI使用者而言，卻是利好。
The market reaction, when it came, was brutal. On January 27th, as investors realised just how good DeepSeek’s “v3” and “R1” models were, they wiped around a trillion dollars off the market capitalisation of America’s listed tech firms. Nvidia, a chipmaker and the chief shovel-seller of the artificial-intelligence (AI) gold rush, saw its value fall by $600bn. Yet even if the Chinese model-maker’s new releases rattled investors in a handful of firms, they should be a cause for optimism for the world at large. DeepSeek shows how competition and innovation will make AI cheaper and therefore more useful.
市場反應如洪水猛獸般襲來。1月27日，投資者們逐漸感受到了DeepSeek的“V3”和“R1”模型所擁有的強勁實力，之後美國上市科技公司的總市值便瞬間蒸發約一萬億美元。作為AI “淘金熱”中的頭牌“賣鏟人”，晶片巨頭英偉達（Nvidia）更是眼睜睜地看著自己的市值揮發了6000億美元。然而，儘管Deepseek的新產品讓小部分企業的投資者感到慌張，但對全球而言，這些新的AI模型應該值得讓大家歡欣雀躍。DeepSeek證明，競爭與創新將降低AI成本，因此也使AI更具實用價值。
DeepSeek’s models are practically as good as those made by Google and OpenAI—and have been produced at a fraction of the cost. Barred by American export controls from using cutting-edge chips, the Chinese firm undertook an efficiency drive, even reprogramming the chips it used to train the model to eke out every drop of power. The cost of building an AI model that can stand toe-to-toe with the best has plummeted. Within days of its release, DeepSeek’s chatbot was the most downloaded app on the iPhone.
在實際使用中，DeepSeek的模型幾乎可與Google和OpenAI開發的模型相媲美，但其研發成本只相當於這兩家公司的一小部分。由於受美國出口管制的限制，Deepseek無法使用尖端晶片，於是便踏上了提升效率之路，甚至重新程式設計用於模型訓練的晶片，以完全榨乾每一滴算力。如今，打造一個能與頂級AI並駕齊驅的模型，成本已大幅下降。DeepSeek的聊天機器人在釋出後短短幾天內，就成為iPhone應用商店中下載量最高的應用。
The contrast with America’s approach could not be starker. Sam Altman, the boss of OpenAI, has spent years telling investors—and America’s new president—that vast sums of money and computing power are needed to stay at the forefront of AI. Investors have accordingly been betting that a handful of firms stand to reap vast monopoly-like rents. Yet if fast followers such as DeepSeek can eat away at that lead for a fraction of the cost, then those potential profits are at risk.
Deepseek的發展與美式路徑形成了極為鮮明的對比。OpenAI執行長山姆·阿爾特曼（Sam Altman）多年來一直向廣大投資者乃至美國新任總統強調，要想保持美國在AI領域的領先地位，就必須投入鉅額資金和海量算力。因此，投資者普遍押注少數幾家企業，預期這些企業將獲得近乎壟斷的超額回報。然而，如果像DeepSeek這樣的“後起之秀”能夠以極低的成本蠶食這一領先優勢，那麼這些企業的潛在鉅額利潤便岌岌可危。
Nvidia became the most valuable listed company in the world thanks to a widespread belief that building the best AI required paying through the nose for its best chips (on which its profit margins are reported to exceed 90%). No wonder DeepSeek’s success led to a stockmarket drubbing for the chipmaker on January 27th. Others in the data-centre business are also licking their wounds, from Siemens Energy (which would have built the turbines to power the build-out) to Cameco (which would have provided the uranium to fuel the reactors to turn the turbines). Had OpenAI been listed, its stock would surely have taken a tumble as well.
要想打造頂級AI，就要不惜重金購入英偉達效能最好的晶片——這一深入人心的觀念讓英偉達成為全球市值最高的上市公司。據報道，這些晶片的利潤率超過90％。難怪DeepSeek的成功導致英偉達股價在1月27日暴跌。包括西門子能源（Siemens Energy）和卡梅科（Cameco）在內的其他資料中心企業也在恢復元氣。按計劃，西門子要建造渦輪機為資料中心擴建供電，卡梅科要用鈾燃料驅動反應堆來運轉渦輪機，但都沒做成。要是OpenAI上市了，它也同樣難逃暴跌命運。
註釋：
1. pay through the nose: pay much more than a fair price花大錢
2. drubbing: a resounding defeat in a match or contest潰敗
3. 資料中心（data center）是專門用於存放計算機系統和相關元件（如伺服器、儲存系統、網路裝置等）的物理設施。它們為企業或組織提供資料儲存、管理和處理的能力，確保各類線上服務和應用能夠持續執行。本文提到的資料中心業務涉及的是與支援人工智慧模型訓練和執行相關的基礎設施。
4. lick one's wounds: to spend time trying to get your strength or confidence back after a defeat or disappointment （失敗或失望後）恢復元氣，重整旗鼓
5. tumble: a sudden fall 跌倒；滾落；暴跌
Yet there are far more winners than losers from the DeepSeek drama. Some of them are even within tech. Apple will be cheering that its decision not to throw billions at building AI capabilities has paid off. It can sit back and pick the best models from a newly commoditised selection. Smaller labs, including France’s Mistral and the Emirati TII, will be racing to see if they can adopt the same improvements, and try to catch up with their bigger rivals.
但在這場DeepSeek風雲中，贏家遠多於輸家，其中一些甚至來自科技行業。蘋果（Apple）肯定會慶幸自己當初沒有投入數十億美元去自建AI能力。如今，它可以坐享其成，從新興的商業化模型中挑選最優者。此外，包括法國Mistral和阿聯酋TII在內的小型實驗室也將競相檢測自己是否可以進行類似的技術改進，以嘗試追上行業巨頭的步伐。
Moreover, efficiency gains are likely to result in greater use of AI. The Jevons paradox—the observation that greater efficiency can lead to more, not less, use of an industrial input—may come into play. The possible applications for a language model with computing costs as cheap as DeepSeek’s ($1 per million tokens) are vastly more numerous than those for Anthropic’s ($15 per million tokens). Many uses for cheaper AI are as yet unimagined.
此外，效率的提升可能會進一步推動人工智慧的普及。傑文斯悖論（Jevons paradox）——即更高的效率會導致資源使用量增加而非減少——可能在此得以印證。與Anthropic的模型（每百萬分詞的成本為15美元）相比，DeepSeek的計算成本（每百萬分詞的成本為1美元）要低得多，如此便大大拓展了應用場景。還有許多低成本的AI應用場景尚未被髮掘。
Even Nvidia may not suffer too much in the long run. Although its market clout may be diminished, it will continue to sell chips in vast quantities. Reasoning models, including DeepSeek’s R1 and OpenAI’s o3, require much more computing power than conventional large language models to answer questions. Nvidia will be hoping to supply some of that.
從長遠來看，即使是英偉達也不一定會損失太多。儘管英偉達的市場影響力可能會有所削弱，但其仍將繼續大量銷售晶片。包括DeepSeek-R1和OpenAI-o3在內的推理模型，在回答問題時所需的算力遠超傳統的大語言模型，而英偉達則有望為滿足這些算力需求繼續提供支援。
註釋：
clout: influence and power影響力
However, the real winners will be consumers. For AI to transform society, it needs to be cheap, ubiquitous and out of the control of any one country or company. DeepSeek’s success suggests that such a world is imaginable. Take Britain, where Sir Keir Starmer, the prime minister, has unveiled a plan to use AI to boost productivity. If he does not need to pay most of the efficiency gains back to Microsoft in usage fees, his proposal has a better chance of success. When producers’ rents vanish, they remain in the pockets of users.
然而，真正的受益者將會是消費者。要讓AI真正改變社會，它必須價格低廉、普及率高，而且不能被任何國家或公司壟斷。DeepSeek的成功表明，這樣的未來並非遙不可及。以英國為例：首相基爾·斯塔默（Keir Starmer）提出了一項利用AI提升生產力的計劃。如果他無需將效率提升帶來的大部分收益作為使用費交還給微軟的話，那麼這項計劃的成功率會更高。當生產者的超額收益消失時，廣大使用者會更受益。
Some have begun to suggest that DeepSeek’s improvements don’t count because they are a consequence of “distilling” American models’ intelligence into its own software. Even if that were so, R1 remains a ground-breaking innovation. The ease with which DeepSeek found greater efficiency will spur competition. It suggests many more such gains are still to be discovered.
一些人開始表示DeepSeek的最佳化並不作數，認為它不過是將美國AI模型“蒸餾”到了自己的軟體中。即便如此，R1依然極具開創性。DeepSeek能夠如此輕鬆地實現效率提升，這將激發更多競爭，也說明還有許多潛在的效率提升機會尚待發掘。
For two years the biggest American AI labs have vied to make ever more marginal improvements in the quality of their models, rather than models that are cheap, fast and good. DeepSeek shows there is a better way.
在過去兩年裡，美國幾家最大的AI實驗室搶破了頭只為讓模型在質量上取得邊際提升，而不是優先考慮讓模型更便宜、更高效、更出色。DeepSeek則證明了還有更好的發展方向。
翻譯組：
Ryan，學海無涯，譯無止境
Yo，女，種下過流星，立志不做大鴕鳥
Mai，男，經濟學博士，世界那麼大，我想活得久一點
審校組：
Charlie，剃刀邊緣無比鋒利
Collin，男，崇拜科比的一枚小翻譯
Clark，願以此功德迴向給眾生，離苦得樂
