[BidClub_]
The a16z Show · · 18 分钟

AI正在成为一场区域竞赛

Anjney Midha

YouTube
TL;DR
  • AI主权已经从“是否采用”的争论,转向“自建还是外购”的决策。 Anjney Midha认为,已有数十亿人使用AI,各国政府实际上已经没有太多选择,无法再讨论是否拥抱AI。主持人将未来24个月各民族国家的相关支出称为可能“单笔最大的采购决策”。
  • 较小国家不需要掌握完整的AI全栈,但必须选择值得信赖的合作伙伴。 Midha将世界划分为前沿“超算中心”和“算力荒漠”;通过合资,国家可以获得基础设施独立性,而不必实现全面自有。由于训练数据会嵌入文化规范,对齐也意味着要判断“谁的价值观与你更接近”。
  • Midha认为,真正重要的只有算力、能源、数据和监管这4项要素。 资源丰富的国家可以用自身比较优势换取缺失能力,例如中东的能源可以吸引模型实验室和技术人才。共同训练的模型,可能让各国“摆脱自己不认同的价值体系,同时实现联合独立”。
  • 真正的主权意味着控制关键依赖,而不是复制每一层能力。 ASML约2亿美元的光刻机说明了这一限制:美国重建这项能力可能需要“10多年”,而训练本土前沿模型可能只需数月到几个季度——前提是一个国家能找到全球少数具备能力的研究团队。
  • Midha认为,美国的数据、能源和推理责任政策是关键风险,同时表示私人算力市场目前表现不错。 他估计,2024年美国各州层面出现了700多项AI立法,形成不断变化的拼图,让企业甚至无法判断“自己到底该遵守什么”。他还认为,核能开发受限,以及让模型开发者为下游滥用承担责任的提案,可能把创业公司推向海外,并进一步巩固大科技公司的地位。
  • 最强的信号来自主权GPU订单和具备技术可信度的创始人。 各国政府在交付前12—36个月就开始订购NVIDIA GPU,因为数据中心正成为“主权新的原子单位”。Midha正在关注那些拥有前沿模型经验、愿意为政府建设基础设施的深度技术型创始人,例如创办Mistral AI的Arthur Mensch,以及在Meta主导最初Llama系列的Guillaume Lample;这一角色的影响力可能“具有相当长远的一代人意义”。
摘要 · 为研究而整理的核心内容

1. AI普及后,主权问题变成自建还是外购

  • Midha将AI视为大约20—22项通用技术之一,与电力和印刷术并列:它是一种横向放大器,而异常迅速的扩散意味着“我们早已过了‘要不要拥抱它’的阶段”。
  • 主持人用企业作类比,进一步明确了这一决策:企业可以自建、租用或购买基础设施,超过190个国家面对的也是同一个选择。主持人认为,未来24个月各民族国家的采购可能成为“单笔最大的采购决策”。

2. 较小国家可以与超算中心结盟,而不是继续留在算力荒漠

  • Midha把能够训练、构建和托管前沿模型的国家称为“超算中心”,把缺乏有意义的在役算力的国家称为“算力荒漠”。他的历史参照是电气化早期:当时较小国家通过与前沿伙伴成立合资企业来推进发展。
  • 与电力不同,AI会编码价值观,因为训练数据携带着当地规范:在美国互联网数据上训练的模型会“总体上带有美国色彩”,而法国数据则会产生更细微、反映法国规范的价值取向。因此,AI可能会重演互联网在中国与世界其他地区之间的分化。
  • Midha还把货币当作先例:缺乏本地资源、无法维持货币与美元挂钩的国家,曾围绕现代货币体系展开合作;新加坡、爱尔兰、卢森堡和苏黎世等较小经济体,则通过与权力中心对齐,成为金融“流量节点”。较小国家在选择兼容的超算中心后,也可以在AI领域扮演同样的角色。

3. 比较优势比全面拥有更重要

  • Midha说,真正重要的有4项要素:算力、充足且低成本的能源、高质量数据Token,以及监管。它们分布不均,反而为联盟创造了机会;一个中东国家可能没有大型数据中心,却可以利用能源储备吸引顶尖团队、企业和基础模型实验室。
  • 他看好盟友关系,包括与其他国家的私人企业合作,并表示共同训练的模型可以让国家在AI栈的某一部分建立优势,同时摆脱自己不信任的价值体系。
  • Midha反对把主权定义为“对全栈每一部分拥有100%的所有权”。ASML约2亿美元的精密光刻机说明了原因:美国复现EUV能力可能需要10多年,而建设并训练一个本土前沿模型可能只需数月到几个季度——前提是能找到全球少数具备能力的研究团队。独立性的真正含义,是避免依赖由不信任对象提供的关键环节。

4. 政府能否接入私人AI,仍是战略分水岭

  • 主持人询问,美国主权是否可以直接交给Anthropic或OpenAI。Midha表示,在一些国家,政府与私人部门的边界非常清晰,在另一些国家则相当模糊;他将美国与中国2017年《国家情报法》作对比,称该法要求中国个人和实体支持国家情报工作。
  • 在美国及大多数盟友国家,私人技术通常受到保护,但双重用途技术或涉密国防技术等受管制领域除外,尤其是由国防项目出资开发的技术。Midha表示,Five Eyes通常采用联合框架对这类基础设施进行分类,但AI模型总体上尚未被归类为双重用途技术,也未被纳入国家安全保护范围。
  • 主持人认为,减少官僚流程拖延、释放一个国家最优秀的人才,历来都是保持领先的关键。

5. 美国的政策瓶颈可能浪费其算力优势

  • Midha认为,美国私人市场正在满足算力需求,但数据政策“异常棘手”。他把拜登行政命令描述为发令枪:命令发出后,各州开始自行行动;他表示,2024年美国各州层面出现了700多项AI立法,但联邦层面仍没有统一的训练数据框架。
  • 他认为,只要规则清晰,前沿模型创始人会接受符合监管要求的美国制度;他们不愿面对50套不断变化、且有时根本无法执行的体系。他还指出,现实中的数据壁垒,以及政府对跨境合作支持不足,都限制了盟友地区之间的数据流通。
  • 能源是另一个弱点:法国20年前拥抱核能,使其具备建设极高能效数据中心的条件,而美国“亲手让自己失去了优势”。如果要求开发者为他人滥用推理输出承担责任,也可能把创业公司推向其他地区,削弱它们对抗大科技公司的能力,并进一步巩固既有巨头。

6. 主权GPU订单和技术型创始人是领先指标

  • Midha称数据中心是“主权新的原子单位”,AI供应链的第一公里就从这里开始。他指出,政府资产负债表正在产生前所未有规模的NVIDIA采购订单,各国把订单排到了交付前12—36个月:“如果排不到队伍前面,一切就结束了。”
  • 他还关注那些拥有学术或深度研究背景、已经主导过前沿模型开发的深度技术型创始人,其中许多人此前在超大规模云厂商的实验室工作。代表人物包括创办Mistral AI的Arthur Mensch,以及在Meta主导最初Llama系列的Guillaume Lample。
  • Midha认为,一类新的创始人正在出现:他们愿意解决大型国家和地区面临的复杂基础设施问题;只要具备必要的“胆量”,他们对人类的影响可能“具有相当长远的一代人意义”。
Speaker 1

Are we going to welcome this technology, or are we going to be hostile to its development?

Anjney Midha

AI has already percolated throughout society at one of the fastest diffusion rates of any general-purpose technology. Now the question everybody’s asking is: Do we build or buy?

My big idea is infrastructure independence, and it’s the idea that a lot of countries and regions are starting to realize that modern AI—you know, deep-learning-based AI, generative models—is a form of what’s been called general-purpose technologies. In the history of humanity, we’ve only had maybe 20 or 22 or so general-purpose technologies, like electricity and the printing press, that have very broad-based applications in society. They end up being largely horizontal economic multipliers and progress multipliers across a whole set of pillars and domains in society.

There are usually 2 moments in the adoption of a general-purpose technology when countries, nations, and states start asking, first, “Are we going to welcome this technology, or are we going to be hostile to its development?” That’s the first step that becomes pretty important in a country’s progression: Do we allow this in, regardless of whether we own it? Do we embrace it or not?

And then the second is: Do we build or buy? Can we trust somebody else to provide it for us? We’re well past the stage of asking whether we embrace it or not. Billions of people around the world have already embraced it, so the government doesn’t really have a choice, so to speak. In that sense, AI has already percolated throughout society at one of the fastest diffusion rates of any general-purpose technology, and now the question everybody’s asking is: Do we build or buy?

Speaker 1

It’s probably the single largest purchasing decision that’s going to happen in the next 24 months. Do nation-states start buying? Do they build or buy?

I love the parallel of companies, because there are many companies that choose to build, but also companies that choose to rent or buy. As you think about that, there are large nations around the world, like the United States, which are clearly building. But talk about the argument for smaller nations—the 190-plus countries around the world.

Anjney Midha

The good news here is that we’ve got hundreds of years of human history to look at for clues about what happens next. If you were a small country in the early 1900s, watching the modern electrification of the developed world, and you chart what happened with many of those countries, many of them decided to enter into what were called joint-venture agreements.

It starts with a joint venture with a country that’s at the frontier. These countries at the frontier of AI are what I call hypercenters. These are countries that have the ability to develop, train, build, and host their own frontier models. I call them hypercenters mostly as an homage to the word hyperscaler, because there have been a handful of companies that have had the compute and the talent to actually build frontier AI.

Now what we’re seeing is a shift from just those companies driving a bunch of frontier AI to countries and regions driving it. If you’re a small country and you’re saying, “We certainly believe that it’s important to have our own AI infrastructure. We want to be independent, but we don’t have all the compute required to train these models, or we don’t have all the talent locally,” then what you enter into is a joint venture with a country or an overseas partner that matches your values.

This is the really important thing about AI models and how they’re different from infrastructure like electricity: There’s a fundamental encoding of human values in AI models, because they’re trained on data. The data has local norms and cultural values embedded in it. If you happen to train a model on a bunch of internet data collected in the U.S., the models are generally American. They’re encoded with that.

If you train a model on data in France, the models subtly have a bunch of different values encoded in them that reflect those cultural norms. I think step 1, if you’re a small country, is actually being crystal clear about which value systems you align with most among the hypercenters.

The way the internet worked out was that there essentially ended up being 2 internets: the Chinese internet and the rest of the world. AI may not end up looking that different. If you’re a small country, what you really have to figure out is whose values align more with yours.

A good historical precedent to look at here is the technology of money. Money is a pretty general-purpose technology, and what happened in the early 1900s with the modernization of finance is that a number of countries started to ask the same question: Do we build or buy our own currency? Do we rely on the dollar, or do we have our own currency?

That led to the modern-day currency regime, where the dollar is a single global reserve currency. It happened through a bunch of allied cooperation, where a number of countries realized they did not have the local resources required to hold the peg of gold to the dollar.

You have the U.S., China, and India, and then you have a number of smaller countries that decided they wanted to be flow points. You have Singapore and Ireland, and you have Luxembourg and Zurich, which became massive global leaders in modern finance because they decided they wanted to ally with one of those power centers.

So, if you think about the AI world, let’s call the regions at the frontier hypercenters. Then we have compute deserts—places that have literally no installed base of compute capacity to even be relevant. All the smaller countries have to figure out which of the hypercenters they want to align with, and how to become a modern-day Singapore, Ireland, Luxembourg, or Zurich for the world of AI infrastructure.

It starts with deciding whether you want to be a compute desert or not. If you’re not, and you’re going to embrace AI infrastructure as a government, then I think you’ve got to figure out which hypercenter you want to align with most. Then it becomes quite easy to reason about how to be a valuable ally.

Speaker 1

That’s such a good parallel, because a lot of people think about resources in terms of the farmland that you have and the people who are working in that economy. But what you’re pointing out is that countries have, for a long time, offered value or a resource in other ways.

As we think about AI, there are a few things that you’ve pointed out that countries can invest in, whether it’s the compute capacity they have, the energy resources to power AI, or forward-thinking policy. Maybe we can break down each of those. How do you think about each of those blocks, and how countries should be maneuvering or investing in those things?

Anjney Midha

The good news is that there are only 3 or 4 ingredients here that really matter. The first is compute, which we’ve talked about. The second is abundant and low-cost energy, which powers the data centers. The third is data—the availability of really high-quality tokens for these models to learn on. The fourth is regulation.

The bad news is that the world is pretty unevenly split up. Some countries have dramatically more compute than others. Others have dramatically more energy than others because of their natural reserves.

If you’re in the Middle East, you may not have massive data centers yet, but what you do have is vast reserves of oil. How you translate that into becoming a hypercenter is the law of comparative advantage: You’ve got energy, so you should use that to attract the world’s best teams, companies, foundation-model labs, and so on, by trading what you have with what they have.

I’m quite bullish on allied ties between countries that recognize what their strengths are and then partner with other countries to fill that gap. By countries, I mean private companies, too, from other countries. One of the things we may end up seeing is jointly trained models between countries.

For most countries, it’s impossible to have total infrastructure independence across all parts of the stack. What is much more feasible is to be great at one part of the stack and then collaborate with another sovereign, country, or region to achieve joint independence from a value system that you don’t subscribe to.

In the long term, you might be able to build things out, but infrastructure of this kind can often take years, if not a decade, to scale. As an example, lower down in the stack from the model layer, you have the chip layer, and even below that you have the lithography layer.

There’s a company in Holland called ASML that builds literally the world’s most important machines. How many machines do they make per year? It’s some very small number. Each machine costs about $200 million, and they’re the only company that can do lithography at this level of precision.

Is it feasible for the U.S. to say, “We’re going to build our own ASML tomorrow”? No. It’s just going to take 10-plus years. EUV lithography takes a really long time.

On the other hand, is it feasible for a smaller country to say, “We’re going to build and train our own local models at the frontier”? That’s a little bit easier to do over a months-to-quarters timescale if you’ve got a leading research team—and that’s a big if. There are only a handful of research teams globally that are capable of this.

To answer your question, I don’t think sovereign AI or infrastructure independence means you have 100% ownership over every part of the stack. That’s infeasible over the short term. It means that you don’t rely on somebody for a critical part that you don’t trust.

Speaker 1

Can we talk about private companies for a second? You’ve brought them up a few times. How do you think about that dynamic where, as a nation-state, you’re saying, “We need this sovereignty,” but at the same time, can you rely on that sovereignty through the companies that exist within your nation?

Using America as an example, does the government really need to be involved, or can it just let Anthropic or OpenAI command that part of the stack? How do you think about the difference between government and private enterprise?

Anjney Midha

The line is pretty stark in a few countries and more blurry in others. In China, the line is very clear. There’s a law called the PRC’s 2017 National Intelligence Law that says Chinese individuals and entities are required to support PRC national intelligence work by law. That means if there’s any technology that a PRC company has access to, it’s automatically obliged to make that available to the government.

That’s not the case in the United States. There are some covered types of technology, like dual-use technology or classified defense technology, where, if you’re developing it—particularly if you’re funded under a defense program—then you’re required to make that available to the government, because the government is paying for the development of that technology.

But by and large, the private sector in the United States and most other allied countries is, by default, protected from having to make its technology available to the government. That’s not the case in the CCP.

I think the question for most countries becomes: Where on that spectrum do you want to exist? Every country approaches it slightly differently, but the Five Eyes—the U.S., Canada, the U.K., Australia, and New Zealand—generally have a joint approach or framework for categorizing this infrastructure.

By and large, AI models have not been categorized as dual-use or protected under national security.

Speaker 1

The history of technology has largely shown that, if you’d like to win, unlocking the best talent of a country with as few bureaucratic slowdowns as possible usually ends up winning.

If we think about wanting to keep America at the frontier, and we think about the different layers or ingredients that we talked about earlier, are there any high-risk areas where you think we’re falling behind? It could be in the energy department, at the model layer, or in the fact that we’re dependent on ASML. Going down the line, is there anywhere where you think we’re at high risk of not keeping our edge?

Anjney Midha

I think we go back to the 4 ingredients we talked about earlier at the frontier of AI: compute, data, energy, and laws.

On the compute front, I think the private market in the United States is doing a pretty good job. It’s pretty responsive to market demand, and I think there’s no coincidence that the largest infrastructure businesses in the United States are chip companies and computing companies. We’ve generally done a pretty good job of letting the market feed that demand.

On the data side, things are extraordinarily tough. The Biden executive order last year was a starting gun that said, “AI is important. Please do something about it,” and left it to the states to figure it out. The states have all taken a complete patchwork of approaches to data regulation.

In 2024 alone, I think there were more than 700 pieces of state-level legislation that were AI-specific. A bunch of those laws, if you look at them, are really well-intentioned but atrociously implemented ideas for data regulation—impossible to adhere to, basically.

I think one area where we’re handicapping ourselves is that there’s no unified framework in the United States at the federal level yet for data, especially around training. We needed that yesterday.

Overseas, in a number of countries where the rule of law, especially on copyright and IP, is just less stringent, those labs are happy to race ahead, whereas our companies here are trying to figure out what they should even comply with. That hurts you more than a laissez-faire approach.

I think our companies would be totally fine with the United States being compliant. The best founders at the frontier would be fine with that. They just want to know what to comply with, rather than having 50 different states with different regulations that are changing, unclear, and in some cases impossible.

There’s also a fundamental scientific problem: There are very real data walls that these models run into. One of the things that hurts frontier research in the United States and allied countries is a lack of government support for collaborating across borders to make more data available to allied regions.

On energy, I think we’ve obviously hamstrung ourselves in the United States with nuclear. France’s embrace of nuclear 20 years ago has positioned it to have extraordinarily efficient data centers today, whereas in the United States, I think we’ve basically shot ourselves in the foot around that.

Lastly, around inference regulation, I think what we’re not doing enough of is making it clear where liability rests. I’ve seen a number of proposals ahead of legislative sessions next year that want to hold model developers liable for the outputs of inference, even if the misuse is being done by somebody else.

What does that do? It drives those very important developers elsewhere, essentially forces most startups to lose much-needed ground to big tech companies, and entrenches incumbents even more.

Speaker 1

This idea is truly global. What are you looking out for—or what should a legislator or the head of a nation be thinking about in some of these decisions? Are you looking for countries that are buying GPUs or building out new energy centers? What are you paying attention to as indicators?

Anjney Midha

Definitely compute. If you think about the AI supply chain, the first mile starts at the data center. That’s the new atomic unit of sovereignty, I would say, which is a new thing. We’ve never actually had nation-states think about an AI data center as an atomic unit that countries should be purchasing.

About 24 months ago, we started seeing nations reason about that first mile as being important. We’ve seen an enormous amount of NVIDIA purchase orders come from the balance sheets of governments—unprecedented demand from nation-states realizing that they want to be hypercenters.

That starts with placing orders 12 to 36 months in advance to take delivery of GPUs, because if you don’t get in front of that line, it’s over. You’re getting them after everybody else.

The second thing I look for is founders who are deeply technical, often come from deep research backgrounds, and have already led frontier-model development, frequently inside large hyperscaler labs. An example is Arthur Mensch, who started Mistral AI; he worked at DeepMind. Or Guillaume Lample, who led the initial Llama family at Meta.

There’s a new class of founder who’s primarily technical, has their training in academia, and is motivated to solve all the really hard problems that come with delivering infrastructure for large nation-states and regions. If you’re a founder like that who has the guts, then your impact on humanity ends up being quite generational.