
亞洲多個中等強國正投入大量公共資源開發本土AI基礎模型以減少對美中依賴,但外媒警告,基礎模型訓練成本每年增長3.5倍,下一代模型成本將遠超目前補貼規模,且維持技術領先需持續巨額投資,落後難以追趕。新加坡選擇微調現有模型作為本地化替代方案,而專家建議亞洲國家應加深半導體、記憶體晶片及機器人等既有產業優勢,而非嘗試從零打造競爭模型,因為領先企業優勢具累積效應,且基礎模型更新快速,傳統產業政策模式在此領域效果有限。
AI-generated summary
亞洲多個中等強國正投入資金發展本土AI基礎模型,希望降低對美國與中國供應商的依賴。此舉延續亞洲自1970年代以來依靠產業政策追趕先進國家的傳統,但基礎模型與傳統產業如汽車、電子存在關鍵差異:更新速度極快、競爭壁壘高、領先優勢具累積效應。
亞洲砸錢追AI基礎模型,可能從一開始就選錯戰場。(路透資料照)
〔財經頻道/綜合報導〕亞洲多個中等強國正投入資金發展本土AI基礎模型,希望降低對美國與中國供應商的依賴。不過,外媒認為,AI基礎模型訓練成本每年約增加3.5倍,下一代基礎模型的訓練成本將是現在的數倍,很快就會超過亞洲追趕型國家目前能提供的補貼規模,警告亞洲中等強國恐陷入燒錢競賽。
外媒指出,單看模型訓練費用,仍低估了真正逼近技術前沿所需的成本,企業還必須爭奪人才、支付大量實驗費用,並承擔訓練失敗所造成的損失。基礎模型尤其難以複製亞洲過去產業追趕的成功模式,因為維持在技術前沿所需的資本投入不僅龐大,而且正在快速上升。
請繼續往下閱讀...
文章指出,訓練目前領先模型之一Grok 4,估計耗資4.9億美元(約新台幣155.3億元);研究機構Epoch AI估算,自2020年以來,前沿語言模型的訓練成本每年約增加3.5倍。這意味著,要持續保持競爭力,政府必須一代接一代投入更高金額。
儘管成本快速上升,亞洲多個中等強國仍正投入公共資源,希望建立本土基礎模型、降低對美中供應商的依賴。南韓2025年8月選出5家公司競逐國產模型計畫,提供5300億韓元(約新台幣124億元)資金,希望開發可與美中領先企業競爭的模型。
印度則透過IndiaAI計畫約11億美元(約新台幣348億元),其中相當部分用於開發以印度語言及本地數據訓練的模型;日本GENIAC計畫也向本土開發商提供運算資源與支援。
這種做法延續了亞洲自1970年代以來依靠產業政策追趕先進國家的傳統,但基礎模型與汽車、電子等傳統產業存在明顯差異。汽車工廠可以使用數十年,即使競爭對手推出更好的產品,由於產品週期較長、轉換成本較高,原有企業仍可能維持市場。但基礎模型更新速度極快,稍微落後的模型若沒有更低推論成本、更好的本地語言能力、更高隱私保護或特殊領域優勢,很難形成明顯差異。
而且領先企業的優勢會隨時間累積。能力較強的模型可以吸引更大規模的全球客戶,創造收入後再投入下一代模型。以Token計價及訂閱制為主的商業模式,也使消費者對模型能力的小幅差距十分敏感,因此落後企業能取得的收入更少,也更難持續投入研發。
AI基礎模型本身又高度依賴研究支出。Anthropic在2025年的研發支出約為公司營收的1.5倍,許多中國競爭對手的研發投入甚至達營收數倍。對亞洲中等強國而言,如果希望長期維持在前沿模型附近,所需投入將遠超目前政策計畫的規模。
新加坡選擇了不同路線,其SEA-LION v4.5模型專門針對東南亞語言及文化情境,並非從零開始訓練,而是在阿里巴巴及Google DeepMind既有開源模型基礎上進行微調,整項計畫預算僅7000萬新幣(約新台幣17.3億元),仍能取得相當程度的本地化效果。
不過,微調能做到的事情仍受到原始模型能力限制,在網路行動及科學研究等具有戰略意義的領域,仍高度依賴更先進的基礎模型。若亞洲國家真正擔心未來遭到前沿AI系統限制,將有限公共資源投入自身已具備長期比較優勢的環節,可能更具持續性,包括南韓的記憶體晶片、台灣的半導體,以及日本的機器人產業。這些產業需要龐大前期投資、技術能力與多年建立的供應鏈,競爭者要大規模複製的成本遠高於重新訓練一套模型。
文章指出,亞洲中等強國擔心依賴外國AI系統並非沒有理由,但如果試圖在本國重新打造能與美中前沿模型競爭的基礎模型,不僅成本高昂,也未必能真正確保取得最具戰略價值的AI能力。相較之下,加深既有產業優勢,可能是更具持續性的選擇。
一手掌握經濟脈動 點我訂閱自由財經Youtube頻道

Nvidia announced a record US$150 billion share buyback and launched the Open Agent Safety Platform, with CEO Jensen Huang urging US-China collaboration on AI safety despite political tensions, emphasizing that competitors can still work together to prevent AI harm.

When Jen-Hsun Huang attended the annual dinner of the South Korea-US friendship non-profit organization "Korea Association" in New York and received the Van Fleet Award, he once again emphasized that AI is not a living organism, but just a software. He opposed giving it too many human characteristics and called for an alignment mechanism to ensure the safe and responsible development of AI.

The result of the Joint Research Center of Tsinghua University Network Research Institute and Qihoo 360, "One Word Settles the World: Systematic Exploration and Exploitation of DNS Silencing Vulnerabilities in Domain Name Resolution" won the "Most Innovative Research Award" at The Pwnie Awards in 2026. This award is known as the Oscars of global white hat hackers and demonstrates China's international influence in the field of network security.

The first batch of units 1 and 2 of the Xinjiang Dashixia Water Conservancy Project invested and constructed by China Energy Construction Gezhouba Group were connected to the grid to generate electricity on September 29. The project breaks the world record for concrete-faced gravel dams with a maximum dam height of 247 meters. It has a total installed capacity of 750,000 kilowatts and an average annual power generation of 1.893 billion kilowatt-hours. It can meet the annual electricity needs of about 600,000 households and assume the responsibility of South China. The four prefectures and prefectures in Xinjiang carried out peak load regulation tasks for the power system, and simultaneously transported 3.42 billion cubic meters of water to the Tarim River ecology, alleviating the water shortage problem of more than 6.8 million acres of farmland in the Aksu River Basin, and raising the flood control standard for downstream rivers from a once-in-10-year to a once-in-50-year period.

Two artificial intelligence system-level chips (SoC), ZQ300 and ZQ320, independently developed by China Electronics Technology Research Institute 58, recently completed full-link evaluation and achieved comprehensive performance indicators. The chip adopts a heterogeneous computing architecture, integrates display control, graphics and intelligent computing power, supports multi-source data fusion scenarios, and can be used for intelligent terminals, unmanned platform edge computing and real-time image processing, promoting the engineering application and large-scale deployment of domestic intelligent computing power chips.

NASA announced the latest progress of Boeing Starliner, planning to conduct an unmanned Starliner-1 mission as early as December 2024 or January 2025 to verify improvements, and will conduct a manned mission in 2028. Previously, the Starliner's first manned test flight caused astronauts to be stranded on the space station due to a malfunction. It was deemed a Class A accident. Boeing has made improvements to the spacecraft's thermal environment.