博威---云架构决胜云计算

 找回密码
 注册

QQ登录

只需一步,快速开始

搜索
查看: 172|回复: 5

台积电美国工厂晶圆投入成本飙升,NVIDIA GPU 价格“全线上涨”-Digitimes

[复制链接]
发表于 前天 09:06 | 显示全部楼层 |阅读模式
台积电美国工厂晶圆投入成本飙升,NVIDIA GPU 价格“全线上涨”-Digitimes

供应链消息人士透露,NVIDIA 正面临多重危机。据报道,由于降级版 H20 芯片被禁止向中国销售,其季度收益减记了 55 亿美元。首席执行官黄仁勋一直往返于中美之间,希望将影响降至最低,并说服美国政府放宽对 AI 芯片的限制。

同时,市场传闻称,为了保持稳定的盈利能力,NVIDIA近期已上调了几乎所有产品的官方价格,并允许合作伙伴相应提高价格。

例如,华硕今年初推出的高端RTX 5090显卡,零售价约为9万元新台币,最近经销渠道价格已飙升至10万元新台币,涨幅超过10%。H200/B200 GPU模块和服务器的价格也随之上涨。

鉴于中国以外的AI芯片需求依然强劲,且云服务提供商(CSP)继续扩大资本支出,预计NVIDIA将在本月底前实现预期内的强劲季度收益。

自4月以来,美国总统特朗普对华加征关税,并收紧AI芯片出口规则,有效阻止了NVIDIA降级版H20芯片进入中国市场,NVIDIA因此成为这些贸易限制措施的主要目标。

黄仁勋正在尽一切努力缓和局势,并将影响降至最低。

为了响应特朗普的“美国制造”倡议,英伟达宣布将在四年内投资5000亿美元,与台积电、富士康等公司合作,在美国建立人工智能服务器制造基地。台积电位于亚利桑那州的工厂已经开始生产4纳米Blackwell芯片。黄仁勋也与特朗普一同出席,表示美国的政策应该侧重于提升美国企业的竞争力。他警告称,限制对中国和其他国家芯片出口将削弱美国的技术领先地位。他还在敏感时期访问了中国政府机构。

黄仁勋在中美两国之间寻求解决方案,同时不断向美国政府强调,中国的人工智能能力正在迅速赶上。他警告称,对人工智能芯片的更广泛限制可能会严重损害美国芯片制造商,并加速中国的发展进程。

在最近与特朗普的联合露面中,黄仁勋甚至提到了被美国列入黑名单的华为,并指出该公司已进入人工智能芯片领域,并在计算和网络技术方面取得了优异的成绩。他重申,美国的政策应优先考虑增强国内竞争力,限制对中国等国的芯片出口将威胁到美国目前的技术领先地位。

引起担忧的不仅仅是黄仁勋。AMD 首席执行官苏姿丰最近也表示,中国对全球半导体和人工智能产业来说代表着“巨大的机遇”,并敦促在出口管制和人工智能发展之间采取平衡的做法。有传言称,特朗普政府正在考虑放松一些现有的人工智能芯片限制。

据供应链消息,NVIDIA 的黄仁勋正在积极参加公开活动,向特朗​​普政府发出警告,同时重申对“美国制造”倡议的支持。

在运营方面,NVIDIA 历来保持着强大的控制力。尽管关税和 AI 芯片限制存在不确定性,但该公司仍在努力实现稳定的盈利。据报道,许多产品的官方价格已上调,合作伙伴也纷纷效仿,纷纷提价。

以需求旺盛的RTX 50系列为例:RTX 5090今年早些时候全球上市后,立即以高价被抢购一空。尤其是中国的AI芯片买家,他们毫不犹豫地掏钱购买,愿意不惜任何代价。RTX 5090最初的售价约为8万新台币,但渠道价格迅速翻了一番。

人工智能芯片禁令颁布后,RTX 5090 的采购难度加大,价格也随之上涨。由于中国订单急剧下降,各大厂商纷纷推出捆绑销售以维持盈利能力并清理库存。华硕等厂商推出了包含电源或整机的套装。RTX 5090 的价格一夜之间从约 9 万新台币飙升至 10 万新台币。其他 RTX 50 系列显卡的价格也上涨了 5% 至 10%。

此外,据报道,NVIDIA已上调H200和B200芯片及模块的价格,以应对成本上涨和平台转型策略。此次涨价已产生连锁反应,服务器厂商也调整了对客户的报价,涨幅约为10%至15%。

供应链指出,将 Blackwell 芯片的生产转移到台积电的美国工厂,导致制造、材料和物流成本显著增加。与台积电本身一样,NVIDIA 拥有将这些成本转嫁给客户的定价权,因此价格上涨在意料之中。

尽管像H20这样的AI芯片被禁止销往中国,但中国买家仍然能够通过大陆以外的其他渠道购买。与此同时,来自美国云计算服务提供商(CSP)和其他国际客户的需求依然强劲。预计NVIDIA将在即将发布的财报中公布业绩,业绩将达到预期,并令人印象深刻。

NVIDIA 截至 2025 年 1 月 26 日的第四财季收益超出市场预期。

营收达到393.3亿美元,环比增长12%,同比增长78%。数据中心营收达到356亿美元,环比增长16%,同比增长93%。整体毛利率为73%。

展望2026财年第一季度,NVIDIA预计营收约为430亿美元,同比增长65%,但增幅低于去年同期262%的增幅。预计毛利率将达到70%。

NVIDIA计划在5月底发布财报。在此之前,黄仁勋将于5月12日赴台与供应链合作伙伴见面,17日与各大厂商聚餐,19日发表主题演讲,重燃这场被许多人称为“万亿美元盛宴”的盛宴,AI热潮蓄势待发。
 楼主| 发表于 前天 09:30 | 显示全部楼层
This was one of the most difficult calls I made and shared with AI Investor community in the past year: holding on to $NVDA.

On April 6, 2025, following Trump’s “Liberation Day” tariff war announcement, the market was in a state of extreme fear.

When I reviewed the tech valuation metrics, Nvidia’s numbers were absurd: a 0.78 PEG ratio for next year, a forward P/E of 20.96, and a next-year P/E of 16.5.

$NVDA, like most other stocks, had a completely broken chart. I understood why both institutions and retail investors were selling. But if history has taught us anything, it's that having the conviction to hold on to a stock is key to achieving significant long-term gains.

This was a follow-up to the DeepSeek moment, NVDA investors have been through a tough few months.

翻译自 英语
这是过去一年中我做出并与 AI 投资者社区分享的最艰难的决定之一:坚持$NVDA 。

2025年4月6日,特朗普“解放日”关税战宣布后,市场陷入极度恐惧状态。

当我查看技术估值指标时,Nvidia 的数字很荒谬:明年的市盈率为 0.78,预期市盈率为 20.96,明年的市盈率为 16.5。

$NVDA和大多数其他股票一样,图表完全崩盘了。我理解为什么机构投资者和散户投资者都在抛售。但如果说历史教会了我们什么,那就是坚定持有股票的信念是实现长期大幅收益的关键。

这是 DeepSeek 事件的后续,NVDA 投资者经历了艰难的几个月。


McKinsey’s base case projects global capex of $5.2 trillion on continued AI momentum, estimated 124GW of incremental capacity will be needed through 2030, versus 205GW in the accelerated demand scenario.

$NVDA $AMD $AVGO $MSFT $AMZN $GOOG

翻译自 英语
麦肯锡的基本预测是,在人工智能持续发展的推动下,全球资本支出将达到 5.2 万亿美元,预计到 2030 年将需要 124GW 的增量容量,而在需求加速增长的情况下则需要 205GW。
 楼主| 发表于 前天 09:32 | 显示全部楼层
美中经贸高层会谈中方代表团记者会实录
中国国务院副总理何立峰:
5月10日至5月11日,我作为中美经贸中方牵头人与美国财政部长贝森特和贸易代表格里尔大使共同举行了中美经贸高层会谈。双方就彼此关心的经贸问题问题开展了深入交流,会谈的氛围是坦诚的、深入的、具有建设性,达成一系列重要共识,会谈取得实质性进展。
双方一致同意建立中美经贸磋商机制,明确双方牵头人,就各自关切的经贸问题展开进一步磋商。中美双方将尽快敲定有关的细节,并且将于5月12日发布会谈达成的联合声明。
我想借此机会特别感谢瑞士政府作为东道主对本次中美经贸高层会谈提供的大力支持和热情的接待。同时美方同事的专业勤勉也令人印象深。
三个多月来由美方挑起的贸易战,全球关注。这场贸易战,中方的态度始终是十分明确的,也是一贯的,也就是中方不愿意打贸易战,因为贸易战没有赢家。但是如果美方执意把贸易战强加于我们,为了维护我自身利益,中方也绝不惧怕,一定会奉陪到底。
在当前的形势下,本次会谈受到了国际社会的高度关注,经过中美双方的共同努力,会谈富有成效,迈出了双方通过平等对话、协商解决分歧的重要一步,也为进一步弥合分歧和深化合作打下的基础创造了条件。
当今世界正处于百年未有之大变局。中美经贸关系既对两国的利益重大,也对全球经济的稳定和发展有着重要的影响。
中美经贸关系的本质是互利共赢,合作中双方难免会出现一些分歧和摩擦。关键要本着相互尊重、和平共处、合作共赢的原则,通过平等对话协商,及时找到妥善解决问题的办法,以推动中美经贸关系稳定。
我们也愿意同美方一道,积极落实两国元首今年1月17日的通话共识,本着解决问题的务实态度,坦诚对话,协商管控分歧,深挖合作潜力,拉长合作清单,做大合作的蛋糕,推动中美经贸关系取得新的发展,为世界经济注入更多的确定性和稳定性。

中国财政部副部长廖岷:
接下来我们请今天(周日)的记者朋友们围绕今天的新闻发布会的主题进行提问。提问前,请大家先通报一下(所在机构)。

新华社记者:
何副总理,您好,我来自新华社。我想提问的问题是,美国白宫已经发布消息称,中美有可能达成的协议有助于解决美国1.2万亿美元的贸易赤字,请问中方对此有何评价?

何立峰:
中美双方会谈的成效明天将有联合声明。想了解这些具体情况可以找李成钢副部长。

中国商务部副部长李成钢(国际贸易谈判代表):
谢谢记者的问题,中方在经贸谈判中从来追求的都是双赢的成果。所以,任何可能达成的成果也一定必须是符合中方自身发展利益。

中央广播电视总台央视记者:
请问中方如何看待本次会谈的氛围?

李成钢:
正如刚才何副总理评价的,这两天中美两国的经贸团队在日内瓦进行了坦诚、深入和建设性的会谈,达成了重要共识。我觉得这次的会谈体现了以下三个特点:
一是相互尊重。双方以两国元首共识为指引,从维护中美经贸关系大局出发,认真听取彼此的关注,充分考虑各自的国情、发展阶段和制度差异,拿出了诚意,推动会谈取得了实质性进展。
二是平等互惠。双方在会谈中照顾彼此的关切和发展利益。秉持理性、客观、务实的精神,相向而行,积极寻找最大公约数,为实现中美经贸关系健康、稳定、可持续发展不断累积条件。
三是专业高效。双方经贸团队本着专业精神,充分利用各自的专业知识,在有限的两天时间展开了密集磋商,对中美渠道领域的相关问进行了坦诚深入的讨论,为本次谈判达成共识提供了坚实基础。

彭博记者:
我的问题问李副部长,首先,请您对于这次会议这个决定建立的这个磋商机制,您能不能介绍一些相关的讯息?第二个是关于提到的明天双方要发的联合声明,我想问一下大概是什么时间能发?是不是在资本市场开盘之前? (众笑)

李成钢:
关于机制,刚才何副总理实际上也做了介绍。通过这次的会谈,双方同意建立中美经贸磋商的机制性安排。就中方而言,是何立峰副总理作为中美经贸谈判中方牵头人,与美方牵头人共同引导该会谈。
在双方牵头人领导下,双方工作团队将就中美经贸相关问题进行定期和不定期的沟通。未来沟通的时间和地点都将由双方进一步商定。而且现在通讯很发达,电话会、视讯会议都是沟通的方式。
对于你问题的第二部分,我想说的是,基于中美之间达成的谅解,今天我们不便过多地披露成果的细节。至于成果公布的时间,中国有句话:“好饭不怕晚”,我想无论什么时候发布,世界的反应都是积极的。

中国财政部副部长廖岷:
今天就到这里,有关情况的后续进展我们还会及时地发布,请大家持续关注,谢谢大家。
 楼主| 发表于 前天 10:15 | 显示全部楼层
WILL $NVDA BE THE FIRST $10 TRILLION COMPANY?

NVIDIA didn’t sneak up on the world -- the world simply didn’t realize what it had built. It wasn't just chips. It wasn’t just GPUs. It was a new substrate for value creation. A new rail system for intelligence itself.

Because we’re no longer living in an internet-first world. We’re living in an intelligence-first one. Where product design, logistics, diagnostics, creative tools, drug pipelines, legal services, national defense -- all of it -- are shifting from deterministic software to probabilistic intelligence. And that shift requires compute. Unimaginable amounts of it. Not once, but perpetually. Not centrally, but everywhere.

Blackwell isn’t a chip. It’s a control layer. It’s the thing hyperscalers and governments alike are racing to secure not because they want to run proofs of concept -- but because they know that without it, they don’t get to participate in the next economy. A single Blackwell rack can cost $3M -- and they’re backordered. This isn’t a bubble that all the bears try to spread. This is a global bidding war for the new oil.

And yet, the noise persists.

“Capex is peaking.”
“Big Tech is building their own silicon.”
“China has caught up.”
“CUDA is replaceable.”
“AI demand isn’t monetizing fast enough.”

But here’s the reality: none of it holds up under scrutiny.

$AMZN, $GOOGL, $META & $AAPL may be designing chips. But they’re still training their models on NVIDIA infrastructure. They still benchmark performance with CUDA. They still build in a development environment NVIDIA owns end to end. Owning inference doesn’t matter if you can’t get your model production-ready -- and CUDA is the only ecosystem that makes that seamless. It’s not just a moat. It’s gravity.

China’s Huawei may have engineered a chip that mimics the H20 -- but let’s not lose the plot. H20 was designed under export restriction. It’s decades behind. Matching that performance is like building a $TSLA that performs like a 2015 Prius. And no -- DeepSeek didn’t train with less compute because China suddenly became efficient. It trained with less because China’s domestic supply is starving. The famine isn’t metaphorical. It’s real. And it’s why sovereigns are now stockpiling NVIDIA systems like strategic reserves.

As for capex? Of course hyperscaler spend will normalize. But that’s like saying road construction slows down once the highways are built. Inference is what comes next. And inference is exponential. Training happens in stages. Inference happens every second of every day, across every app, every device, every industry. And guess who’s built the only full-stack system that can scale that?

The thing people miss is this: NVIDIA doesn’t just sell to cloud providers. It sells to every company trying to escape dependence on them. They all need compute. And not just hardware. They need software, firmware, drivers, simulators, developer tools, vertical models, reference architectures. They need a system. NVIDIA is the only company that offers one.

So when people ask: Is NVIDIA the next $10 trillion company? -- they’re asking the wrong question.

They’re still thinking in quarters. Still thinking in cycles. Still asking if the numbers are sustainable, if margins will hold, if unit volumes will surprise to the upside.

But this isn’t a product story. It’s a platform shift. It’s a new industrial base.

Because NVIDIA isn’t just a chip company anymore -- it’s the compute layer of the modern world. What Amazon was to commerce. What Google was to knowledge. What Microsoft was to productivity. NVIDIA is becoming to intelligence.

And if the 21st-century economy runs on AI, then NVIDIA doesn’t just supply the picks and shovels.

It owns the mine. 由   翻译自 英语
 楼主| 发表于 前天 10:15 | 显示全部楼层
WILL $NVDA BE THE FIRST $10 TRILLION COMPANY?

NVIDIA didn’t sneak up on the world -- the world simply didn’t realize what it had built. It wasn't just chips. It wasn’t just GPUs. It was a new substrate for value creation. A new rail system for intelligence itself.

Because we’re no longer living in an internet-first world. We’re living in an intelligence-first one. Where product design, logistics, diagnostics, creative tools, drug pipelines, legal services, national defense -- all of it -- are shifting from deterministic software to probabilistic intelligence. And that shift requires compute. Unimaginable amounts of it. Not once, but perpetually. Not centrally, but everywhere.

Blackwell isn’t a chip. It’s a control layer. It’s the thing hyperscalers and governments alike are racing to secure not because they want to run proofs of concept -- but because they know that without it, they don’t get to participate in the next economy. A single Blackwell rack can cost $3M -- and they’re backordered. This isn’t a bubble that all the bears try to spread. This is a global bidding war for the new oil.

And yet, the noise persists.

“Capex is peaking.”
“Big Tech is building their own silicon.”
“China has caught up.”
“CUDA is replaceable.”
“AI demand isn’t monetizing fast enough.”

But here’s the reality: none of it holds up under scrutiny.

$AMZN, $GOOGL, $META & $AAPL may be designing chips. But they’re still training their models on NVIDIA infrastructure. They still benchmark performance with CUDA. They still build in a development environment NVIDIA owns end to end. Owning inference doesn’t matter if you can’t get your model production-ready -- and CUDA is the only ecosystem that makes that seamless. It’s not just a moat. It’s gravity.

China’s Huawei may have engineered a chip that mimics the H20 -- but let’s not lose the plot. H20 was designed under export restriction. It’s decades behind. Matching that performance is like building a $TSLA that performs like a 2015 Prius. And no -- DeepSeek didn’t train with less compute because China suddenly became efficient. It trained with less because China’s domestic supply is starving. The famine isn’t metaphorical. It’s real. And it’s why sovereigns are now stockpiling NVIDIA systems like strategic reserves.

As for capex? Of course hyperscaler spend will normalize. But that’s like saying road construction slows down once the highways are built. Inference is what comes next. And inference is exponential. Training happens in stages. Inference happens every second of every day, across every app, every device, every industry. And guess who’s built the only full-stack system that can scale that?

The thing people miss is this: NVIDIA doesn’t just sell to cloud providers. It sells to every company trying to escape dependence on them. They all need compute. And not just hardware. They need software, firmware, drivers, simulators, developer tools, vertical models, reference architectures. They need a system. NVIDIA is the only company that offers one.

So when people ask: Is NVIDIA the next $10 trillion company? -- they’re asking the wrong question.

They’re still thinking in quarters. Still thinking in cycles. Still asking if the numbers are sustainable, if margins will hold, if unit volumes will surprise to the upside.

But this isn’t a product story. It’s a platform shift. It’s a new industrial base.

Because NVIDIA isn’t just a chip company anymore -- it’s the compute layer of the modern world. What Amazon was to commerce. What Google was to knowledge. What Microsoft was to productivity. NVIDIA is becoming to intelligence.

And if the 21st-century economy runs on AI, then NVIDIA doesn’t just supply the picks and shovels.

It owns the mine.

翻译自 英语
$NVDA会成为第一家市值 10 万亿美元的公司吗?

NVIDIA 并非悄无声息地占领了世界--世界只是没有意识到它创造了什么。它不仅仅是芯片,也不仅仅是 GPU。它是价值创造的新基石,是通往智能本身的全新轨道系统。

因为我们不再生活在互联网优先的世界,而是生活在一个智能优先的世界。产品设计、物流、诊断、创意工具、药物研发、法律服务、国防--以及所有这些--都在从确定性软件转向概率智能。而这种转变需要计算。其规模之大难以想象。计算不是一次性的,而是持续性的。计算不是集中式的,而是无处不在的。

Blackwell 不是芯片,而是一个控制层。超大规模数据中心运营商和政府都在竞相抢占它,并非因为他们想进行概念验证--而是因为他们知道,没有它,他们就无法参与到下一个经济时代。一个 Blackwell 机架的价格高达 300 万美元-- ,而且目前订单积压。这并非所有看跌者都试图制造的泡沫,而是一场争夺新石油的全球竞价战。

然而,噪音仍然存在。

“资本支出正在达到顶峰。”
“大型科技公司正在打造自己的硅片。”
“中国已经赶上来了。”
“CUDA是可替代的。”
“人工智能需求货币化速度不够快。”

但现实是:这些都经不起推敲。

$AMZN 、 $GOOGL 、 $META和$AAPL可能正在设计芯片。但他们仍在 NVIDIA 基础架构上训练模型。他们仍在使用 CUDA 进行性能基准测试。他们仍在 NVIDIA 拥有的端到端开发环境中构建。如果你的模型无法投入生产-- ,那么拥有推理能力就毫无意义,而 CUDA 是唯一能够无缝衔接的生态系统。这不仅仅是一条护城河,更是引力。

中国华为或许已经设计出一款模仿 H20 的芯片-- ,但我们别忘了重点。H20 是在出口限制下设计的,落后了几十年。要达到那样的性能,就像打造一辆性能堪比 2015 款普锐斯的$TSLA 。而且, -- DeepSeek 训练时减少计算量,并非因为中国突然变得高效。它减少训练量是因为中国国内供应短缺。饥荒并非比喻,而是真实存在的。这也是为什么各国现在像储备战略物资一样囤积 NVIDIA 系统。

至于资本支出?当然,超大规模数据中心的支出会恢复正常。但这就像说高速公路建成后,道路建设的速度就会减慢一样。接下来是推理。推理是指数级增长的。训练是分阶段进行的。推理每时每刻都在发生,跨越每个应用程序、每个设备、每个行业。猜猜是谁构建了唯一一个能够扩展的全栈系统?

人们忽略了一点:NVIDIA 不仅仅向云提供商销售产品。它面向所有试图摆脱对云提供商依赖的公司。这些公司都需要计算,而不仅仅是硬件。他们需要软件、固件、驱动程序、模拟器、开发工具、垂直模型、参考架构。他们需要一个系统。NVIDIA 是唯一一家提供这种系统的公司。

因此,当人们问:NVIDIA 会是下一个价值 10 万亿美元的公司吗? --时,他们问错了问题。

他们仍然以季度为单位思考,仍然以周期为单位思考,仍然在问这些数字是否可持续,利润率是否能保持,销量是否会意外上升。

但这并非产品故事,而是平台转变,是一个新的工业基础。

因为 NVIDIA 不再只是一家芯片公司--它是现代世界的计算层。就像亚马逊之于商业,谷歌之于知识,微软之于生产力。NVIDIA 正在成为智能领域的领导者。

如果 21 世纪的经济依靠人工智能运转,那么 NVIDIA 提供的就不仅仅是镐和铲子了。

该矿由该公司所有。
翻译得准确吗?请提供反馈,以便我们加以改进:  
下午8:53 · 2025年5月10日
·
11.7万
查看
 楼主| 发表于 前天 10:16 | 显示全部楼层
network 发表于 2025-5-12 10:15
WILL $NVDA BE THE FIRST $10 TRILLION COMPANY?

NVIDIA didn’t sneak up on the world -- the world s ...

WILL $NVDA BE THE FIRST $10 TRILLION COMPANY?NVIDIA didn’t sneak up on the world -- the world simply didn’t realize what it had built. It wasn't just chips. It wasn’t just GPUs. It was a new substrate for value creation. A new rail system for intelligence itself.Because we’re no longer living in an internet-first world. We’re living in an intelligence-first one. Where product design, logistics, diagnostics, creative tools, drug pipelines, legal services, national defense -- all of it -- are shifting from deterministic software to probabilistic intelligence. And that shift requires compute. Unimaginable amounts of it. Not once, but perpetually. Not centrally, but everywhere.Blackwell isn’t a chip. It’s a control layer. It’s the thing hyperscalers and governments alike are racing to secure not because they want to run proofs of concept -- but because they know that without it, they don’t get to participate in the next economy. A single Blackwell rack can cost $3M -- and they’re backordered. This isn’t a bubble that all the bears try to spread. This is a global bidding war for the new oil.And yet, the noise persists.“Capex is peaking.”“Big Tech is building their own silicon.”“China has caught up.”“CUDA is replaceable.”“AI demand isn’t monetizing fast enough.”But here’s the reality: none of it holds up under scrutiny.$AMZN, $GOOGL, $META & $AAPL may be designing chips. But they’re still training their models on NVIDIA infrastructure. They still benchmark performance with CUDA. They still build in a development environment NVIDIA owns end to end. Owning inference doesn’t matter if you can’t get your model production-ready -- and CUDA is the only ecosystem that makes that seamless. It’s not just a moat. It’s gravity. China’s Huawei may have engineered a chip that mimics the H20 -- but let’s not lose the plot. H20 was designed under export restriction. It’s decades behind. Matching that performance is like building a $TSLA that performs like a 2015 Prius. And no -- DeepSeek didn’t train with less compute because China suddenly became efficient. It trained with less because China’s domestic supply is starving. The famine isn’t metaphorical. It’s real. And it’s why sovereigns are now stockpiling NVIDIA systems like strategic reserves.As for capex? Of course hyperscaler spend will normalize. But that’s like saying road construction slows down once the highways are built. Inference is what comes next. And inference is exponential. Training happens in stages. Inference happens every second of every day, across every app, every device, every industry. And guess who’s built the only full-stack system that can scale that?The thing people miss is this: NVIDIA doesn’t just sell to cloud providers. It sells to every company trying to escape dependence on them. They all need compute. And not just hardware. They need software, firmware, drivers, simulators, developer tools, vertical models, reference architectures. They need a system. NVIDIA is the only company that offers one.So when people ask: Is NVIDIA the next $10 trillion company? -- they’re asking the wrong question.They’re still thinking in quarters. Still thinking in cycles. Still asking if the numbers are sustainable, if margins will hold, if unit volumes will surprise to the upside.But this isn’t a product story. It’s a platform shift. It’s a new industrial base.Because NVIDIA isn’t just a chip company anymore -- it’s the compute layer of the modern world. What Amazon was to commerce. What Google was to knowledge. What Microsoft was to productivity. NVIDIA is becoming to intelligence.And if the 21st-century economy runs on AI, then NVIDIA doesn’t just supply the picks and shovels.It owns the mine.
翻译自 英语

$NVDA会成为第一家市值 10 万亿美元的公司吗?NVIDIA 并非悄无声息地占领了世界--世界只是没有意识到它创造了什么。它不仅仅是芯片,也不仅仅是 GPU。它是价值创造的新基石,是通往智能本身的全新轨道系统。因为我们不再生活在互联网优先的世界,而是生活在一个智能优先的世界。产品设计、物流、诊断、创意工具、药物研发、法律服务、国防--以及所有这些--都在从确定性软件转向概率智能。而这种转变需要计算。其规模之大难以想象。计算不是一次性的,而是持续性的。计算不是集中式的,而是无处不在的。Blackwell 不是芯片,而是一个控制层。超大规模数据中心运营商和政府都在竞相抢占它,并非因为他们想进行概念验证--而是因为他们知道,没有它,他们就无法参与到下一个经济时代。一个 Blackwell 机架的价格高达 300 万美元-- ,而且目前订单积压。这并非所有看跌者都试图制造的泡沫,而是一场争夺新石油的全球竞价战。然而,噪音仍然存在。“资本支出正在达到顶峰。”“大型科技公司正在打造自己的硅片。”“中国已经赶上来了。”“CUDA是可替代的。”“人工智能需求货币化速度不够快。”但现实是:这些都经不起推敲。$AMZN$GOOGL$META$AAPL可能正在设计芯片。但他们仍在 NVIDIA 基础架构上训练模型。他们仍在使用 CUDA 进行性能基准测试。他们仍在 NVIDIA 拥有的端到端开发环境中构建。如果你的模型无法投入生产-- ,那么拥有推理能力就毫无意义,而 CUDA 是唯一能够无缝衔接的生态系统。这不仅仅是一条护城河,更是引力。中国华为或许已经设计出一款模仿 H20 的芯片-- ,但我们别忘了重点。H20 是在出口限制下设计的,落后了几十年。要达到那样的性能,就像打造一辆性能堪比 2015 款普锐斯的$TSLA 。而且, -- DeepSeek 训练时减少计算量,并非因为中国突然变得高效。它减少训练量是因为中国国内供应短缺。饥荒并非比喻,而是真实存在的。这也是为什么各国现在像储备战略物资一样囤积 NVIDIA 系统。至于资本支出?当然,超大规模数据中心的支出会恢复正常。但这就像说高速公路建成后,道路建设的速度就会减慢一样。接下来是推理。推理是指数级增长的。训练是分阶段进行的。推理每时每刻都在发生,跨越每个应用程序、每个设备、每个行业。猜猜是谁构建了唯一一个能够扩展的全栈系统?人们忽略了一点:NVIDIA 不仅仅向云提供商销售产品。它面向所有试图摆脱对云提供商依赖的公司。这些公司都需要计算,而不仅仅是硬件。他们需要软件、固件、驱动程序、模拟器、开发工具、垂直模型、参考架构。他们需要一个系统。NVIDIA 是唯一一家提供这种系统的公司。因此,当人们问:NVIDIA 会是下一个价值 10 万亿美元的公司吗? --时,他们问错了问题。他们仍然以季度为单位思考,仍然以周期为单位思考,仍然在问这些数字是否可持续,利润率是否能保持,销量是否会意外上升。但这并非产品故事,而是平台转变,是一个新的工业基础。因为 NVIDIA 不再只是一家芯片公司--它是现代世界的计算层。就像亚马逊之于商业,谷歌之于知识,微软之于生产力。NVIDIA 正在成为智能领域的领导者。如果 21 世纪的经济依靠人工智能运转,那么 NVIDIA 提供的就不仅仅是镐和铲子了。该矿由该公司所有。

翻译得准确吗?请提供反馈,以便我们加以改进:


[backcolor=rgba(29, 155, 240, 0.1)]













下午8:53 · 2025年5月10日
·
11.7万
查看





您需要登录后才可以回帖 登录 | 注册

本版积分规则

小黑屋|手机版|Archiver|boway Inc. ( 冀ICP备10011147号 )

GMT+8, 2025-5-14 04:41 , Processed in 0.108871 second(s), 16 queries .

Powered by Discuz! X3.4

Copyright © 2001-2021, Tencent Cloud.

快速回复 返回顶部 返回列表