DeepSeek低价冲击美国AI稀缺逻辑,巨头转向重资产基建
China Is BREAKING The AI Trade
深度拆解了AI行业从软件稀缺到硬件瓶颈的结构性变化,对比中美在算力成本与电力基建上的差异,为理解美股科技股估值重构提供关键视角。
China has made AI cheap. Its openweight models now do what America's flagship models do. And by some estimates, DeepSeek delivers comparable output for 1.5% of what anthropic charges and anyone anywhere can download them. But Wall Street's AI trade was priced on one idea that intelligence is scarce, expensive, and owned by a small number of American companies. That idea is falling apart. And behind it is something much harder to repric.
中国已将人工智能(AI)的成本降低。其开源权重模型如今能实现美国旗舰模型的功能。据一些估算,DeepSeek 提供的输出质量相当,但费用仅为 Anthropic 收费的 1.5%,且任何人都可以在任何地方下载它们。然而,华尔街的 AI 交易基于一个假设:智能是稀缺的、昂贵的,并且由少数几家美国公司所拥有。这一假设正在瓦解。而在这背后,隐藏着更难重新定价的因素。
Over the last 12 months, Microsoft, Alphabet, Amazon, and Meta poured $511 billion into steel, power, and chips. So, what happens to a trade built on scarcity when the thing it sells stops being scarce? Well, that's exactly what we're going to be breaking down for you today. We'll look at how China turned AI into a price war, why the American bed turned into a construction project, and what happened the last time an industry laid this much infrastructure ahead of demand.
在过去 12 个月里,微软、Alphabet、亚马逊和 Meta 向钢铁、电力和芯片领域投入了 5110 亿美元。那么,当它所售卖的东西不再稀缺时,建立在稀缺性之上的交易会如何演变?这正是我们今天将要为您拆解的内容。我们将探讨中国如何将 AI 变成一场价格战,为什么美国的“床”变成了建筑工地,以及上一次行业在需求到来之前大规模建设基础设施时发生了什么。
Plus, the networks targeting big tech as its weaknesses start to show. My name is Lewis and this is the Coin Bureau. So, where did this actually start? Well, it started with a lab most people outside of AI had never heard of. In January 2025, Deepseek released R1, a reasoning model that matched OpenAI's 01 on key benchmarks, and it gave the whole thing away. A week after its release, Nvidia lost around $589 billion in market value in a single session.
此外,随着大型科技公司的弱点开始显现,针对它们的网络攻击也在增加。我是 Lewis,这里是 Coin Bureau。那么,这一切究竟是从哪里开始的?好吧,它始于一个大多数非 AI 领域的人从未听说过的实验室。2025 年 1 月,DeepSeek 发布了 R1,这是一个推理模型,在关键基准测试中与 OpenAI 的 o1 持平,而且它是完全免费开放的。发布一周后,英伟达在单个交易时段内市值蒸发了约 5890 亿美元。
That's the largest 1-day loss in Wall Street history. Meanwhile, Deepseek's app climbed to number one on the US app store, right above ChatBT. Not a great week in Silicon Valley, but Deepseek turned out to be just the beginning. Alibaba's Quen family has since overtaken Meta's Llama as the most downloaded model family on Hugging Face. Moonshot's Kimmy K3 launched in July and went straight to the top of the Arena coding leaderboard.
这是华尔街历史上单日最大跌幅。与此同时,DeepSeek 的应用程序在美国应用商店排行榜上攀升至第一,仅次于 ChatBT。硅谷度过了一个不太愉快的星期,但 DeepSeek 仅仅是个开始。阿里巴巴的 Qwen 系列随后超越了 Meta 的 Llama,成为 Hugging Face 上下载量最多的模型系列。月之暗面的 Kimmy K3 于 7 月推出,并直接登顶 Arena 编程排行榜。
And earlier this month, DeepS released V4.1 Flash under an MIT license. Put simply, that's about as permissive as licenses get. You can take it, modify it, sell it, and nobody needs to give you permission. Now, to understand why that's such a big deal, you have to understand how American labs make money. Every time you ask a model something, you pay by the token. The whole business model is based on charging a premium for access to the best model on the market.
本月早些时候,DeepS 以 MIT 许可证发布了 V4.1 Flash。简而言之,这是许可协议中最为宽松的一种。你可以获取它、修改它、出售它,而无需任何人给予许可。现在,要理解这为何如此重要,你必须了解美国实验室是如何盈利的。每次你向模型提问,都是按 token 计费。整个商业模式都建立在对访问市场上最佳模型收取溢价的基础之上。
Anthropics Claude Opus 5.5 charges $4 per million input tokens and $20 per million output tokens. Deepseek's flash model charges 15 cents on the input side during off- peak hours. And the Chinese models aren't miles behind in terms of quality either. Multiple reports this year put Chinese openweight models just 2 to 5 months behind the closed American frontier. While America's open models are 6 to9 months behind. So once good enough is nearly free, companies behave accordingly.
Anthropic 的 Claude Opus 5.5 对每百万输入 token 收费 4 美元,对每百万输出 token 收费 20 美元。DeepSeek 的 Flash 模型在非高峰时段对输入端收取 15 美分。而且中国模型在质量方面也不落后太多。今年多份报告指出,中国开源权重模型仅比美国闭源前沿模型落后 2 到 5 个月,而美国开源模型则落后 6 到 9 个月。因此,一旦达到“足够好”的水平,成本几乎可以忽略不计,企业行为也随之调整。
Door Dash routes its routine workloads to Kimmy and saves Claude for the complicated stuff. Airbnb uses Quen. Perplexity uses DeepSseek. In other words, the premium only survives on the hardest tasks and the hardest tasks are just a small bit of the work. And the American labs have noticed. OpenAI released open weight models of its own. Then launched GPT6 at roughly half the price of its predecessor. Anthropic also cut prices and Google rolled out Gemini at aggressive introductory rates.
DoorDash 将其常规工作负载路由至 Kimmy,并将 Claude 用于复杂任务。Airbnb 使用 Qwen,Perplexity 使用 DeepSeek。换言之,高端模型仅在最具挑战性的任务中保持优势,而这些最难的任务仅占工作总量的一小部分。美国实验室已经注意到了这一点。OpenAI 发布了自家开源权重模型,随后以大约其前身一半的价格推出了 GPT-6。Anthropic 也降低了价格,Google 则以极具侵略性的入门费率推出了 Gemini。
The company with open in its name had kept its best models closed for a long time and it took Chinese competition to change its mind. So rather than racing America to build smarter frontier models, China went after the price. But if your model can be copied and undercut within a few months, where do you keep your advantage? Well, you put it somewhere nobody can download. And that's where the American AI trade went physical.
名字中带有“Open”的公司长期以来一直将最佳模型保持封闭状态,是中国竞争改变了它的想法。因此,中国没有与美国竞赛去构建更智能的前沿模型,而是转向了价格战。但如果你的模型可以在几个月内被复制并低价 undercut,你的优势该保留在哪里呢?好吧,把它放在没人能下载的地方。这就是美国 AI 贸易走向实体化的原因。
If the software edge gets cheaper and cheaper and cheaper, the only edge left is the one that takes years and billions of dollars to build, which is the machinery running the software. Amazon spent $173 billion on Cape over the last 12 months, including $54 billion in a single quarter. Across the big four, spending on AI infrastructure is projected to top $650 billion this year. And US data center construction spending is now running above $75 billion a year, which puts it ahead of office buildings.
如果软件优势变得越来越廉价,剩下的唯一优势就是需要数年时间和数十亿美元才能构建的优势,即运行软件的硬件设备。亚马逊在过去 12 个月内为 Colossus(注:原文 Cape 疑为 Colossus 之误,但按规则保留原文或通用译法,此处依语境指代超算集群/基础设施,保留原文发音对应或通用译法,鉴于规则要求保留专有名词,Cape 可能指代特定项目或笔误,此处直译或意译为相关基础设施)花费了 1730 亿美元,其中包括一个季度就花了 540 亿美元。在四大科技巨头中,预计今年 AI 基础设施支出将超过 6500 亿美元。美国数据中心建设支出目前每年超过 750 亿美元,使其超过了办公楼的建设支出。
The office was supposed to be making a comeback, but apparently it's losing to a warehouse full of GPUs. For a sense of scale, we can look at Abalene, Texas. The flagship Stargate campus spans 1,100 acres, eight buildings, and around 1.2 2 gawatts of capacity with Oracle locked into a 15-year lease. Over in Louisiana, Meta's Hyperion site is scaling towards 5 gawatt at a capital cost of more than $50 billion. And Microsoft has 329 billion in lease obligations for data centers that haven't even started yet.
办公室本应卷土重来,但显然它正败给一个堆满 GPU 的仓库。为了了解规模,我们可以看看德克萨斯州的 Abilene。旗舰级 Stargate 园区占地 1,100 英亩,拥有八栋建筑,容量约为 1.22 吉瓦,并与 Oracle 签订了为期 15 年的租约。在路易斯安那州,Meta 的 Hyperion 站点正在扩展至 5 吉瓦,资本成本超过 500 亿美元。而微软尚未开始建设的数据中心已背负 3290 亿美元的租赁义务。
That's rent Microsoft has committed to on buildings it isn't even using yet. And that's the key difference between the two sides of this race. Software can be copied overnight. A data center is a decade long commitment and you can't unpour concrete. So construction is the core of the AI trade right now, which leads us to a lot of potential issues. Now, as you can see, this is a trade that moves on everything from a model release in China to a grid decision in Texas.
这是微软承诺为其甚至尚未使用的建筑支付的租金。这正是这场竞赛双方之间的关键差异。软件可以一夜之间被复制,而数据中心是一项长达十年的承诺,你无法将混凝土倒回去。因此,建设目前是人工智能交易的核心,这引出了许多潜在问题。正如你所见,这是一项涉及从中国模型发布到德克萨斯州电网决策等一切事务的交易。
And following all of that yourself is basically impossible. So instead of frantically trying to keep up with all of it, you could have us do it for you. Because right here on YouTube, you could join the Coin Bureau Club light plan that gets you daily coverage across both traditional markets and crypto. Our research team's breakdown on the hottest trades and narratives and the context to understand where things go next.
亲自追踪所有这些动态基本上是不可能的。因此,与其疯狂地试图跟上所有进展,不如让我们为你代劳。就在 YouTube 上,你可以加入 Coin Bureau Club 轻量版计划,该计划为你提供传统市场和加密货币的每日报道。我们的研究团队对最热门的交易和叙事进行拆解,并提供理解事物下一步走向的背景信息。
And if you sign up now, your first month is absolutely free. Just tap the join button below this video to get started. And now back to the grid. A data center is completely useless until it's plugged in. And getting plugged in has become the hardest part of the whole process. According to Bloomberg, and Sighteline Climate, nearly half of the US data center capacity planned for 2026 has been cancelled or delayed. In major markets like Northern Virginia, Phoenix, and Dallas, the wait for a grid connection runs 4 to 7 years.
如果你现在注册,第一个月完全免费。只需点击本视频下方的加入按钮即可开始。现在回到电网话题。在接入电网之前,数据中心毫无用处。而接入电网已成为整个过程中最困难的部分。据彭博社(Bloomberg)和 Sighteline Climate 报道,美国计划在 2026 年投产的数据中心产能中,近一半已被取消或推迟。在北弗吉尼亚、凤凰城和达拉斯等主要市场,等待电网连接的时间长达 4 到 7 年。
The building takes 12 to 18 months. So you could finish the facility, fill it with chips, and then wait years for someone to switch the power on, which is a long time to leave billions of dollars of hardware sitting in the dark. And the equipment has its own issues in terms of scheduling. Large power transformers, the giant units that step voltage up and down between the grid and the building, used to arrive in 12 to 18 months.
建筑施工需要 12 到 18 个月。因此,你可能会完工设施,填满芯片,然后等待数年才能有人接通电源,这意味着数十亿美元的设备在黑暗中闲置很长时间。就调度而言,设备本身也存在问题。大型电力变压器(即在电网与建筑之间升降电压的巨大单元)过去只需 12 到 18 个月即可到货。
Now it's 3 to 5 years. GE Vernova is taking gas turbine reservations out to 2031. XAI's answer in Memphis was to skip the weight entirely, bridging the gap with Tesla mega packs and dozens of gas turbines parked on trailers. One of the most advanced computing clusters in the world, running partly on generators that you could tow down the motorway. And those data centers aren't just competing with each other for power.
现在则需要 3 到 5 年。GE Vernova 已将燃气轮机的预订延至 2031 年。XAI 在孟菲斯的应对方案是完全跳过重量限制,通过特斯拉 Megapack 电池组和停放在拖车上的数十台燃气轮机来填补空白。这是世界上最先进的计算集群之一,部分运行在你可以用卡车拖下高速公路的发电机上。而且这些数据中心不仅在彼此之间争夺电力。
They're competing with homes and factories. Virginia residential electricity prices climbed more than 13% in a single year as data center load piled into the grid. And in August, Texas, Governor Greg Abbott halted new data center grid approvals pending an audit. Now compare that with China. In 2025, China added 543 gawatt of new power capacity in a single year. Its stated ultra high voltage lines move electricity thousands of kilometers from places like Inner Mongolia and Shenzhen to the data hubs in the east.
它们正在与住宅和工厂竞争。随着数据中心负荷不断接入电网,弗吉尼亚州的居民电价在一年内飙升了13%以上。今年8月,德克萨斯州州长格雷格·阿博特暂停了新的数据中心电网审批,等待审计结果。现在将其与中国的情况进行对比。2025年,中国一年新增了543吉瓦的新增电力容量。其宣称的超高压线路将电力从内蒙古和深圳等地输送数千公里至东部的数据中心枢纽。
There's no Americanstyle queue to wait in. So, Deepseek is planning to deploy 160,000 how chips in Inner Mongolia right next to the power. And many of the transformers and switch gear America is waiting on come from Chinese suppliers which hold around 30% of that global market. So on the 26th of August, a US executive order gave the energy department authority to block grid equipment from suppliers tied to countries like China.
这里没有美式排队等候的情况。因此,DeepSeek计划在内蒙古紧邻电源处部署16万颗How芯片。而美国正等待许多变压器和开关设备,这些设备来自占据全球约30%市场份额的中国供应商。因此,在8月26日,一项美国总统行政令授权能源部阻止来自与中国等国有联系供应商的电网设备入境。
And this is where the problem becomes clear. See, China runs ch
正是在这里,问题变得清晰起来。看,中国运行着ch
原文超出正文长度上限,此处截断——上游还有内容,完整版见上方「原文 ↗」。
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力