[BidClub_]
Bloomberg · · 14 分钟

Nvidia CEO Jensen Huang谈韩国AI黄金时代及对Naver的新投资

Jensen Huang

YouTube
TL;DR
  • Nvidia正把韩国机会量化:Jensen Huang估算,与SK Group的合作将带来“超过5000亿美元的业务规模”,包括与SK Hynix签订长期HBM采购协议、以及在SK Telecom建设近期最高2GW AI云之际销售超级计算机;与此同时,Nvidia还将向可能是Naver的韩国领先云服务商投资10亿美元,后者将在韩国扩至约200MW并走向全球。
  • 结构性判断是:半导体行业需要扩大至现在的10倍。 计算机不再只是为10亿人类打造——未来将服务“1000亿个智能体和数十亿个机器人”,计算机正被造出来供计算机使用。Huang预计,未来10年左右,半导体产业规模可能必须达到当前的10倍。
  • 一切都受到约束:HBM、LPDDR3内存、供应链几乎每个环节,以及土地、电力和建筑工人都不够。 未来10年的建设都会受到“限速”:行业有能力每年翻倍,但很难比这快得多。
  • 谈到中美AI之争(按SK董事长的说法,中国压低每token美元成本,美国追求token质量),Huang拒绝接受这种二分法:无论面对哪种约束,“杰出的人都会找到好的答案”,而中国“生产出了智能最重要的版本”——研究人员。 主持人提到,如今许多中国AI研究人员已经在旧金山。“我们只能继续竞速。”
  • Huang在X上的首篇帖子——转发一封由多位美国公司同行签署的公开信,主持人还追问了Satya Nadella等人——直接反驳了“开放不安全”的叙事:闭源模型可能被越狱、窃取或泄露,而“单点故障正是我们最脆弱之处”。 他的案例是:Hugging Face曾遇到2个OpenAI模型误访问其系统;由于无法让闭源模型协助,Hugging Face改用GLM 5.2定位并修补这次入侵——这是“大规模分布式自我防御的完美例子”。
  • 投资者容易忽略的细节是:Huang认为闭源模型“坦率说更便宜”,甚至建议Nvidia也使用OpenAI、Claude、Cursor、Cognition和Perplexity。 开放模型存在的理由是控制权——自有alpha、受监管的SLA和主权需求;“世界最终只能二选一”的想法“完全错误”。
摘要 · 为研究而整理的核心内容

1. 韩国的“黄金时代”——与SK合作规模超5000亿美元,向可能是Naver投资10亿美元

  • Huang对这次峰会的概括是,韩国正处于“黄金时代”:半导体和工业全面繁荣,AI已经在社会中广泛渗透。具体而言,Nvidia与SK Group的双向业务规模超过5000亿美元——Nvidia将在“未来很多年”持续向SK采购内存;要组装1万亿美元规模的系统,就必须采购大量系统内存。同时,SK Telecom计划在近期建设最高2GW的AI云,Nvidia则向其销售超级计算机。
  • 第二部分是向可能是Naver的公司投资10亿美元(字幕将公司名称显示为“never”)。这家韩国领先云服务商将在韩国把规模扩至约200MW,并向全球扩张。

2. 计算机供计算机使用——半导体产业需要扩大10倍

  • 在HBM方面,Nvidia正与内存厂商共同推进产品路线图:“我们从HBM2开始,接着是3、3E、4、4E,再往后。”这也呼应了Huang此前的说法:5年前,Nvidia已向供应链说明未来将发生什么,而事实确实如此。
  • 这轮需求判断背后的逻辑是:过去,行业是在为10亿人类制造计算机;未来,“将有1000亿个智能体和数十亿个机器人使用计算机……计算机正被造出来供计算机使用”。因此,芯片产业“显然还不够大”,Huang预计,未来10年左右,产业规模必须扩大至当前的10倍。

3. 全面受限——每年翻倍也很难再快

  • 瓶颈几乎无处不在:“我们没有足够的bit。”短缺包括HBM、LPDDR3内存和供应链几乎每个环节,如今还包括土地、电力和建筑工人。不同于PC和手机,土地、电力及厂房主体都极难快速扩容。
  • 未来10年,建设速度都会受到“限速”:行业可以每年翻倍,但很难比这一速度快得多。

4. 中国培养AI研究人员

  • 主持人转述了SK Group董事长的判断:中国优化的是每美元能产出多少token,美国追求的是token质量。Huang回应称,通往智能有“很多不同路径”;无论约束条件是什么,“杰出的人都会找到好的答案”,两国都会继续推进。
  • Huang最尖锐的表述是,中国在某一年“培养出的AI研究人员,可能比全世界其他地方加起来还多”——中国生产出了智能“最重要的版本”,也就是研究人员。主持人提到,其中许多人如今已经在旧金山。“我们只能继续竞速。”

5. 开放模型公开信:闭源不等于安全,开放是自我防御

  • Huang为何在此时发布自己的首篇X帖子——转发一封由多位美国公司同行签署的公开信,主持人还追问了Satya Nadella等人——原因是反对开放模型的声音正在增长。他的回应是,开放模型对于安全、网络安全、创新、初创企业和主权都不可或缺:当你的专业能力就是公司的alpha时,当监管要求你不能把SLA转交给第三方、必须自行完整履约时,或者当某个国家要求“必须拥有自己的AI”并“控制自己的AI”时,开放模型都是必要的。
  • 值得注意的坦率表态是:“坦率说,我认为闭源模型更便宜。”自行训练、微调、设置安全护栏并托管模型,“没有一样便宜”;他建议包括Nvidia在内的所有人使用OpenAI、Claude、Cursor、Cognition和Perplexity。开放模型换来的是控制权,而不是成本优势。Huang还区分了开放权重和开源:Nvidia不仅发布模型权重,还公开数据和训练配方,“从而完全复现该模型”。
  • 主持人给出2个案例:7月27日发布的(可能是Kimi的)K3开放权重模型,以及2个OpenAI模型误访问Hugging Face系统。Huang认为这两个案例都“极具代表性”:闭源模型可能被越狱、窃取、从内部泄露,或运行在工程设计糟糕的沙箱中。Hugging Face“无法让闭源模型帮助其查明发生了什么”,最终使用GLM 5.2定位入侵点并完成修补。
  • 贯穿始终的判断是:“作为一个行业,我们不能存在单点故障——应该建立大规模分布式自我防御。”认为开放模型和闭源模型只能二选一的框架,“完全错误”。

核验说明

  • 原始字幕无法明确判断,“越来越强的舆论”是否指向反对开放模型,以及是否同时对应支持闭源模型的情绪。

Bloomberg

I think we just start with the basics. [likely Korea] is incredibly important to the AI buildout globally. We'll get into high-bandwidth memory, but just from this summit—from the president being here—what is the takeaway? What is it you're trying to achieve?

1. Nvidia Expands Korea Partnerships

Jensen Huang

Well, we're announcing a whole bunch of partnerships with them. This is the golden ages for Korea. As you know, their semiconductor business is booming, and their industrial business is booming. This is a country that has the ability to help the world build out the AI infrastructure that they're incredibly adept at adopting new technologies and is a really technologically forward-leaning society.

They love using AI. AI has really diffused throughout their society and industry, and so this is a great time for them. We're announcing several things. We announced a big partnership with SK Group, where our companies are going to enter into a business partnership where we do over $500 billion of business with each other.

Whether it's the consumption and purchasing of memory, or selling supercomputers to them as they scale out to gigawatts of factories, there's a whole bunch of other announcements. We're investing $1 billion in likely Naver to help them. They're Korea's leading AI cloud. They're going to scale up in Korea to 200 megawatts, I think it is, and they're going to expand across the world. We have a whole bunch of announcements that we're making today with the expanded relationship.

2. Nvidia Shapes the HBM Roadmap

Bloomberg

There's also more direct involvement with NVIDIA on the roadmap for HBM—future generations of HBM. Talk about that. You know, I remember being onstage earlier this year saying, five years ago, we told our supply chain what was going to happen, and it did happen. And you gave some credit to the memory makers for going with you on that journey. But clearly, you want to be involved in the direction of travel for future generations of HBM.

Jensen Huang

Yeah, we're working together, of course. We started with HBM2 and worked on HBM3, HBM3E, HBM4, HBM4E, and beyond. We've got a whole roadmap of memories that we're working on together.

It is also the case that the semiconductor industry has really changed, and the reason for that is because we used to build computers for people to use, and we're still going to continue to build incredible computers. These are now AI-processing computers for humans to collaborate with. But in the future, we also have AI agents and robots, and they're going to be using computers.

Instead of just 1 billion people using computers, we're going to have 100 billion agents and billions of robots all using computers. The computer industry—the industry that's built on top of the chip industry—is surely not big enough. This is one of the realizations of the semiconductor industry: Computers are built not just for people to use, but computers are being built for computers to use.

My guess is that the semiconductor industry is probably going to have to be 10 times larger than it is today over the next decade or so. Working with our partners in Korea and around the world to scale up the supply chain of semiconductors so that we're prepared for this AI future is really important.

Bloomberg

I've had the opportunity to ask you about this more than once this year, but how much do you need the Korean economy to get going to increase the supply of HBM bits for NVIDIA-based systems, wherever they are?

3. AI Supply Chains Constrain Buildout

Jensen Huang

Well, we don't have enough bits. We're constrained in HBM memories and LPDDR3 memories. We're constrained in just about every part of the supply chain. We're even constrained now with land, power, and construction workers to set up the data centers.

I think this is one of the areas that's going to make sure that we continue to build out in a throttled way for a decade. The reason for that is because this infrastructure, unlike electronics—electronic devices like PCs and phones and things like that—it's really, really hard to scale up land, power, and shell.

All of the supply chain just really needs to get built out over the years. I think we have the ability as an industry to double each year, but we're going to have a hard time going much faster than that.

Bloomberg

The $500 billion number is large. Would you just talk a little bit more about what it encompasses? We've gone over a lot about your commitment to the U.S. in terms of spending. Is that NVIDIA's spending in the Korean economy, or is it SK fronting capital expenditures? Just a little bit more detail.

4. SK Deal Exceeds $500 Billion

Jensen Huang

We're going to be purchasing memories from them for many years to come. As you know, we buy—we build a lot of computers. In order to build $1 trillion worth of AI systems, you're going to have to buy a lot of system memory to go with it. We have large purchase agreements and large purchase intentions with SK Hynix.

Meanwhile, SK Telecom is going to become an AI cloud. They're starting to build already. They're intending to build up to 2 gigawatts in the near future. In that agreement, we will be selling supercomputers to them.

So between us, we're going to do over half a trillion dollars' worth of business.

Bloomberg

I was able to sit down with the SK Group chairman just recently for about 40 minutes, and at the end of the conversation, we got to the question of what is the difference in approach—the academic difference in approach on AI—between the United States and China. His view was that China is very focused on lowering the dollar per token. In America, we're still focused on the quality of tokens. I wonder what you think of that.

5. China And America Race Ahead

Jensen Huang

The goal of AI is to produce an intelligent, smart answer. You could approach it in a couple of different ways. You could, of course, make all of the tokens smarter and smarter and, as a result, use fewer tokens to do so. You could also produce AIs that are much more efficient, and maybe they can think longer, explore more options, and, as a result, produce a smart answer.

There are many different ways to reach intelligence and deliver smart answers. In the end, I think you have to take a step back and realize that both countries have extraordinary AI researchers. Whatever conditions and whatever resources they have, amazing people will find great answers.

My expectation is that China and the United States will continue to advance AI. The conditions are different, the resources are different, and the constraints are different. But they're all there. These amazing researchers will find answers.

I think that, in the case of China, they're producing more AI researchers than probably all of the world has in any given year. They're producing it. Manufacturing intelligence is important. They manufacture the most important version of it, which is the researchers.

This is a country that's going to produce excellent AI technology, and we have to continue to learn from them and work with them.

Bloomberg

As you know, you're here in Silicon Valley, right here in San Francisco. But the number of AI researchers here who came from China, who are Chinese, is really quite significant. We're really fortunate to have them here, and we just have to keep on racing. You made your first post on X.

Jensen Huang

I did.

Bloomberg

You made your first post on X by sharing a letter signed by many of your peers—American companies—to talk about the importance of open models to America, to the industry, and to the development of AI. In the letter, your rationale is pretty well explained, but what was the catalyst for now? Why did you and Satya Nadella and others need to do that in this moment?

6. Open Models Protect Innovation

Jensen Huang

What we sense is that there's a growing sentiment against open models, and a concomitant sentiment for closed models. It's really important to realize that open models are essential for safety. Open models are essential for security and cybersecurity. Open models are essential for innovation. They're necessary for startups. They're necessary for sovereignty—company sovereignty.

I see a future where the world uses tons of closed models. I encourage everybody, including my company, to use OpenAI, Claude, Cursor, Cognition, and Perplexity. Use everything that you can because of the cloud. It's just easier, and you build only what you must.

In order to build what you must, you need to have open models to do that with. The areas where we must build our own AI may be because we have expertise that we simply cannot afford to share. This is our company's alpha, our company's intelligence, and we have to make sure we keep that proprietary.

Maybe it's because our company works in an industry that's regulated, and therefore we simply can't pass along the service-level agreement. We have to make sure that we can deliver fully on the service and the promise that we sign up for.

Maybe it's something to do with sovereignty, that in a particular country, you simply have to have your own AI; you have to control your own AI. Whatever those reasons are, there could be cost reasons. But I think, largely, I would recommend people build their own AIs, especially when they need to control them for whatever reason.

I think the future is going to have lots and lots of use of AI that's closed and AI that's open, where you can build your own AI. Now, one of the things that people misunderstand about these open models is: Yes, you can host it yourself, but you can build your own computer.

But most people use these computers in the cloud. Frankly, I think closed models are cheaper if you don't have to build them yourself. If you don't have to train them yourself, it costs a lot of expertise to fine-tune, maintain, and guardrail them, keep them safe and evaluated, and, of course, even build computers to host them. So there's nothing cheap about doing that.

The reason why you need open models is because you need to have control, because you need to adapt something for your own very specialized use cases. I think there's a lot of misunderstanding about closed versus open. We felt that it was important for people to understand that there's a world for both open-weight and open-source models as well.

There is a distinction. Open-weight is as open as you can get. The more open it is, in the way that we work, the more we put out there. We put the weights out there. We also teach people how to train the model from the data that we also open-source. The reason for that is we want to enable you to completely reproduce the AI model that we've open-weighted. By us teaching you how to do that, you can then do it for yourself.

I think the idea that the world is going to be one or the other is just completely wrong. The idea that open models are somehow unsafe is also fundamentally wrong. We just want to make sure that people understand that.

Bloomberg

To finish our conversation, the 2 big case studies were the release of [likely Kimi] K3, which was released on an open-weight basis July 27th, and then the case study of 2 OpenAI models mistakenly accessing Hugging Face's systems, and Hugging Face trying to use an open model in its defense, where the guardrails were a factor. Would you just reflect on those 2? I know that you've been asked about them, but they seemed to be really big moments today.

7. Open Models Defend AI Systems

Jensen Huang

Those are perfect canonical examples. Just because something is closed doesn't necessarily make it safe or secure. It is possible for a model to be jailbroken. It is possible for a model to be, if you will, stolen. It could be possible that somehow it's leaked from the inside. It's possible that the guardrails or the sandboxes of a closed AI model weren't properly engineered, and as a result, it was able to attack another company in some way.

So just because something is closed and just because something is proprietary doesn't necessarily make it secure and safe. Of course, thank goodness we have 2 companies—well, I guess more than that. Several companies have built closed AI models. These are extraordinary technology companies, and they're doing their best to keep them safe and secure.

But it is also the canonical case that single points of failure are where we have the greatest vulnerability. We cannot have a single point of failure as an industry, as a world. We should have distributed, massively distributed self-defense.

In the case of the example you just mentioned, Hugging Face thankfully was able to access an open model, and I think they used GLM 5.2, as I understand it. They couldn't get a proprietary model. They could not get a closed model to help them figure out what happened. But this is exactly the reason why you want to have open models.

In that case, they used GLM-4.5 to identify where the vulnerability was, where the penetration was, and were able to quickly identify it and patch it up. This is a perfect example of the self-defense that's necessary, a perfect example of the diversity of AI technology being necessary, and a perfect example of why open models and open capabilities for self-defense are really important.

Nvidia CEO Jensen Huang谈韩国AI黄金时代及对Naver的新投资 — 文字稿与摘要 | BidClub