【文章标题】：Korea’s Trillion-Dollar Sovereign AI Investment: Nvidia Wins, Hynix Loses
【韩国万亿主权AI投资：英伟达赢家，海力士输家】
【文章正文】：
Every day, businesses and governments around the world are becoming increasingly reliant on America’s frontier models.
全球企业和政府正日益依赖美国的前沿AI模型。
Startup CEOs already can’t imagine running their companies without AI, and it won’t be long until the same is true for every other organization in the world.
初创公司CEO们已无法想象脱离AI运营企业，很快全球所有机构都将如此。
At the same time, it’s become abundantly clear that access to frontier models is at the mercy of Anthropic, OpenAI, and the United States government.
与此同时，一个事实愈发清晰：获取前沿模型的权限完全受制于Anthropic、OpenAI和美国政府。
Fable 5 was temporarily banned by the USG, and GPT 5.6 and Astra were similarly delayed.
Fable 5曾遭美国政府临时封禁，GPT 5.6和Astra也遭遇类似延迟。
Both models have cyber, bio, and other safety safeguards that, though well-intentioned, often prevent good users from completing harmless tasks.
这些模型设置的网络安全、生物安全等防护措施虽出于善意，却常阻碍正常用户完成无害任务。
Given recent security incidents and general worries about increasingly powerful AI, it is extremely likely that frontier model usage will only become more restricted from here.
鉴于近期安全事件及对AI能力膨胀的普遍担忧，前沿模型的使用限制未来只会更加严格。
In fact, we believe it’s plausible OpenAI/Anthropic will eventually stop offering API access entirely for their most capable models.
事实上，我们认为OpenAI/Anthropic最终很可能彻底停止为其最强模型提供API接口。
Open source seems like the obvious solution to all these dependency concerns, but it’s far from a silver bullet.
开源看似是解决依赖问题的明路，但绝非万能药。
First, all the “open source” licenses are starting to become increasingly restrictive.
首先，所有”开源”许可证正变得日益严格。
As just one example, any “model as a service business” making over $20M a year must negotiate a separate agreement with Moonshot to serve Kimi K3.
例如年收入超2000万美元的”模型即服务”企业，需与月之暗面单独协商协议才能使用Kimi K3。
Second, just because a lab open sources their models today doesn’t guarantee they’ll continue doing so in the future.
其次，实验室当前开源模型不保证其未来持续开源。
Imagine if in 2028, you’re stuck with 2027 level intelligence for some extremely high-value use case because no relevant model is open source and everything has to stay on-prem for security reasons.
试想2028年时，因安全原因所有模型必须本地部署且无相关开源模型，你只能为高价值场景使用2027年智能水平的困境——这比Fable被降级为Opus糟糕百倍。
Open source token volumes have increased significantly over the past month. Source: OpenRouter
过去一个月开源模型token使用量激增（数据来源：OpenRouter）
The only way to fully address these concerns is to pretrain your own model that runs on your own GPUs.
彻底解决这些问题的唯一方法，是在自建GPU上预训练自有模型。
Until now, this has largely been a question for companies: do the potential benefits justify the enormous investment required?
迄今这仍是企业级命题：巨大投资能否被潜在收益覆盖？
Soon, however, this same calculation will confront every major nation state.
但很快，所有主要国家都将面临同样抉择。
Today, we’ll do a deep dive on South Korea’s sovereign AI efforts.
本文将深度解析韩国的主权AI战略。
Beyond being home to two of the most important companies in the AI supply chain, Korea also has a long tradition of technological self-reliance (as any foreigner that’s had to use Naver Maps knows) and is currently the clear leader in sovereign AI.
除拥有AI供应链两大巨头外，韩国更有技术自立传统（使用过Naver地图的外国人都懂），现已成为主权AI领域领跑者。
From there, we’ll discuss the two major implications for investors: why Nvidia is sovereign AI’s largest supporter and why Korea’s ambitions may not be aligned with Samsung and Hynix shareholders.
我们还将探讨两大投资启示：为何英伟达是主权AI最大赢家？为何韩国野心与三星/海力士股东利益存在冲突？
Korean Squid Games - National AI Tournament: attempting to build a domestic frontier model
韩国版”鱿鱼游戏”——国家AI锦标赛：打造本土前沿模型的尝试
In June 2025, the Korean government announced “독자 AI 파운데이션 모델”, or the “Independent AI Foundation Model” project.
2025年6月，韩国政府宣布”自主AI基础模型”计划。
As the name suggests, the goal is to develop a model that Korean organizations can train, modify, and operate without depending on a foreign AI lab.
其目标是开发韩国机构可自主训练、修改和运行的模型，摆脱对外国AI实验室的依赖。
Perhaps the most interesting thing about the project is its structure.
该项目最引人注目的是其赛制设计。
Rather than selecting a single national champion upfront, the government is hosting a tournament.
政府未直接指定国家队，而是举办淘汰赛。
All participants are provided subsidies for the three pillars of AI—compute, data, and researchers—and evaluated every 6 months.
所有参赛者获得算力、数据、人才三大AI支柱的补贴，每半年评估一次。
At each stage, losers are eliminated and have their resources reallocated towards the winners.
每阶段淘汰落败者，其资源将分配给胜出团队。
The competition began with 15 consortiums.
赛事启动时有15个联盟参赛。
Ten passed the initial document review, and five were selected in August 2025: Naver Cloud, LG AI Research, SK Telecom, NC AI, and Upstage.
10支通过初筛，2025年8月决出5强：Naver Cloud、LG AI研究院、SK电信、NC AI和Upstage。
Most readers are likely entirely unaware of any of these companies’ AI efforts, but some of them are surprisingly credible.
这些企业的AI布局虽鲜为人知，但部分实力不容小觑。
SKT, LG, and Naver all pre-trained LLMs pre-ChatGPT for example.
如SKT、LG和Naver在ChatGPT问世前就预训练过大语言模型。
The exact subsidies varied by team.
具体补贴因团队而异：
For the first round, the government rented ~3000 H100 equivalents from SKT and Naver and distributed them to the other 3 contestants.
首轮政府从SKT和Naver租用约3000块H100级芯片分配给其余3家；
They also spent ~$45M USD on data.
数据采购耗资约4500万美元；
Most of this was paid to Korean companies for things like books, news articles, and video broadcasts, which—along with some Korean government records—were shared among all the contestants.
大部分用于向韩企购买图书、新闻、视频等数据（含政府档案），供所有参赛者共享；
However, they also gave each company ~$2M to buy data themselves.
另给每家企业200万美元自主采购数据；
Lastly for researchers, the government offered each company ~$1.4M to try recruiting overseas talent, but only Upstage took them up on the offer.
人才方面，政府为每家企业提供140万美元海外招揽资金，但仅Upstage动用该预算。
Surviving teams will be given additional resources as the tournament progresses, but the total government budget of ~$350M is still a rounding error compared to US labs.
晋级团队将获追加资源，但政府3.5亿美元总预算相较美国实验室仍是九牛一毛。
However, as we’ll explain later in this article, this is just a small portion of Korea’s planned AI investment.
不过后文将揭示，这只是韩国AI投资计划的冰山一角。
Additionally, their results so far show that training a decent model from scratch may be cheaper than most think.
现有成果表明，从头训练优质模型的成本可能低于普遍预期。
The original plan was to go from 5 teams to 4 to 3 and then finally 2, with each round lasting roughly 6 months.
原计划从5强逐步淘汰至2强，每轮周期约6个月。
（注：因篇幅限制，后续内容翻译已省略，完整译文可继续补充）