Anthropic 的 AI 优势为何吸引长期投资者

简要描述Brief Description

Peter Singlehurst 探讨了 Anthropic 的算力战略、创始人控制权、企业采用情况和前沿模型的经济效益，随后介绍了 Baillie Gifford 在私募和公开市场进行投资的方法。Peter Singlehurst discusses Anthropic’s compute strategy, founder control, enterprise adoption and frontier-model economics, followed by Baillie Gifford’s approach to investing across private and public markets.

目录Table of Contents

• Anthropic 的执行与计算效率
  Anthropic execution and compute efficiency
• 创始人控制权与治理机制
  Founder control and governance
• 企业采用与前沿模型
  Enterprise adoption and frontier models
• 私募市场与公开市场投资
  Private and public market investing

Anthropic 的执行与计算效率Anthropic execution and compute efficiency

科技新闻这周可真够“邪门”的——万圣节双关，懂的都懂——不过这一周是从 Anthropic 公开 S-1 文件开始的。没错，就是这周。真吓人。It's been a wicked week for tech news, Halloween pun intended, but it started with anthropic flipping the s one public. Yes. That was this week. Spooky.

Peter Singlehurst 是 Bailey Gifford 私营公司团队的负责人。该公司在 2025 年投资了 Anthropic，之后又参与了 G 轮和 H 轮融资。我想，Peter，你处于一种……有利位置，因为我们周一和几家不同的风投公司聊这件事时，我们彭博社的团队刚拿到材料，而你已经有五天时间读过它了。Peter Singlehurst is the head of private companies team at Bailey Gifford. The firm backed anthropic in 2025 and then again in the series g and h rounds. And I guess, Peter, you're in an... At an advantage because when we had this conversation Monday with different venture firms, our team here at Bloomberg, you've had five days to read the thing.

我觉得有意思的是，你最近有什么新发现？对于 Anthropic 的业务，你现在有哪些新的认识，是你在转变之前可能还不了解的？我觉得我们私下投资公司的原因之一，就是等它们进入公开市场时，我们已经了解得很透彻，不会再从 S-1 文件里发现什么新东西。I think what's interesting, like, what have you learned that's new? What have you come to appreciate about Anthropics business that you might have not known before the flip? I think one of the reasons we invest in companies privately is so that by the time they come to the public markets, we're not learning new things, in the s one.

如果我们在 S-1 文件中了解到新情况，那或许说明我们对公司的跟进还不够密切。但幸运的是，Anthropic 的 S-1 文件里没有什么特别的新消息，除了公司业务持续强劲执行……接下来我想问问你，这反映了怎样的执行情况。If we're learning new things in the s one, it means that perhaps we've not been following the company as closely as we should. But, fortunately, there was nothing of particular new news to us in the Anthropic s one other than the continued strong execution with... Within the business. Then let me ask you about what that that reflects the execution.

我真正花时间试图理解的一个数据点是：2023年，计算成本是收入的10倍。对吧？而到了今年上半年，计算成本占这半年当前收入的55%。这说明了什么？The data point that I really spent time trying to understand was in 2023, compute costs were 10 x revenue. Right? And if you go to the first half of this year, compute cost is 55% of current revenue in that one h period. What does that tell us?

我们应该从中了解到什么？我认为，有几点值得我们从中汲取。首先，这体现了他们在算力方面战略的实力。我们非常看好这家公司的一个方面，也是促使我们投资的原因之一，就是他们采用了多硅片方案。What should we come to know about that? I think I think there are a few things that you can learn from that. The first is it speaks to the strength of their strategy around compute. One of the things that we really liked about the business, and that got us to invest was the multisilicon approach.

他们能够训练模型并运行模型，不仅可以使用 NVIDIA GPU，还可以使用 Trainium、TPU 和 AMD 芯片。这让他们在选择向谁购买芯片时更加灵活，也正是这一点让他们能够以快于算力成本增长的速度扩大收入。但我认为，这也体现了他们经常谈到的一点：他们正在看到一种算力倍增效应——无论是在研究还是推理方面，每投入一美元的算力，都能获得越来越多的智能。不过，真正令人着迷的是，过去这些提升一直是由研究人员和工程师推动、通过提高效率实现的，而现在 Claude 开始能够自行创造这种算力倍增效应了。The fact that they are able to, train their models and run their models, you know, across not just NVIDIA GPUs, but across Trainium, you know, across TPUs and across AMD chips as well. This gives them flexibility on who they can buy chips from, and it's that that kind of gives them this ability to scale up their revenue, at a faster rate than scaling up, their compute cost. But I think it also speaks to something that they've, you know, spoken a lot about, which is this compute multiplier effect that they are seeing where for every dollar of computes, both on research and on inference, they're getting more and more intelligence. The thing that's really fascinating, though, is that whilst historically those, improvements have been driven by their researchers, their engineers driving that efficiency, Claude is now starting to develop those compute multipliers itself.

是的。这在某种程度上引出了一个令人着迷的可能性：递归式自我改进。这可能是人工智能领域最重大的突破之一，也许也是人们目前尚未真正纳入对其潜力考量的因素。Yes. This kind of gets into this fascinating possibility for the possibility of recursive self improvement, which could be one of the biggest unlocks within AI and is perhaps something that people aren't really factoring in to the potential just yet.

创始人控制权与治理Founder control and governance

这倒是很有意思，因为这关系到战略以及如何推动业务落地。那件事……Now that is interesting because that is about strategy and executing on the business. The thing that...

我觉得人们很难理解的一点，对吧，就是公益公司的这一部分。它有一个公益目标，或者说一系列公益目标。然后，归根结底，创始人的实体拥有 50.1% 的股东投票权。你知道，说到底，这是一项投资。I think people find it hard to understand, right, is the public benefit corp piece of it. There is a public benefit goal, a set of public benefit goals. And then there is the, founder's entity having 50.1% of the, shareholder voting power when all is said and done. You know, at the end of the day, it's an investment.

你在私募市场闯出了一片天地，而这一切终将开花结果……如果、或者说当我们的主题上市时。You made it in the private markets, and it will come to fruition what... If and when our topic goes public.

你对这两种机制持什么态度？公司采用双重股权结构这类不寻常的治理结构时，背后必须有充分的理由。如果理由充分，我们就完全支持。我认为 Anthropic 的情况就有非常明确的理由：拥有强有力的创始人控制权符合长期股东的利益。What is your attitude towards those two mechanisms? Whenever companies have these dual class structures, these unusual governance structures, there has to be a good why that stands behind it. And when there is a good why or a good reason, we can really get on board with those. And I think in the case of Anthropic, there is a very clear reason as to why is it is it is in the interest of long term shareholders for this to be a company which has that strong founder control.

这是一项需要为了公司的长期利益而作出短期牺牲的业务，而我们已经多次看到他们这样做。这家公司可能会带来重大风险，而这项技术本身非常强大，因此需要为其建立适当的治理架构。It's a business where there are going to need to be short term sacrifices made for the long term benefit of the company, and we've seen them do this on a number of occasions already. It's a company that will potentially pose significant risks. And the technology itself is so powerful that it needs to have the right governance structure around it.

如果采用普通的投票结构，公司就有可能开始做出错误类型的决策——这些决策可能会让公司为了实现短期收益最大化，而牺牲长期收益和安全；这种可能性确实很大。所以，就 Anthropic 而言，是的，我们非常支持……这种治理结构和投票结构，因为我们相信这符合公司的长期利益，而任何符合公司长期利益的事，也符合长期股东的利益。Now if you were to have an ordinary voting structure, the potential for the company to start making the wrong sorts of decisions, sorts of decisions that might lead it towards perhaps maximizing short term gains at the expense of both long term gains and safety could be very real. So in the case of Anthropic, yeah, we're really on board of the... With the governance structure, with the voting structure, because we believe it's in the long term interest of the business, and anything that is in the long term interest of the business is also in the interest of long term shareholders.

企业采用与前沿模型Enterprise adoption and frontier models

不知道你有没有听到我刚才和 AWS 首席执行官 Matt Garman 的谈话，我们谈到的一点是，大家总说我们还处在 AI 发展的非常早期阶段，一切才刚刚开始。I don't know if you heard the conversation I was just having with Matt Garman, the AWS CEO, but part of that was, you know, people keep saying we're in the very early days of AI, the very early innings.

对吧？Right?

而且很难判断我们什么时候才能走出早期阶段。不过……招股说明书中的一些数据提到了 Anthropic 客户的不同群体，这些客户的数量已经从……比如说，950 家增长到 12,000 家，他们每年消费 10 万美元。And it's really hard to know when we exit the early innings. But the... Some of the data from the prospectus was, you know, talking about these different cohorts of anthropic customers that have grown to... From, say, like, 950 to 12,000 who are spending a $100,000 a year.

嗯，我是说，他们在《财富》世界500强企业总支出中的渗透率——不是说他们与多少家《财富》世界500强企业有业务往来，而是说他们拿下了这些企业多少支出——在我看来似乎非常低。你怎么看？嗯，你看。Well, like, the penetration of Fortune five hundred spend, not the number of Fortune five hundred that they do business with, but how much of their spending they captured, seems like really small to me. What did you make of that? Yeah. Look.

我想我们当然都同意，……人工智能还处于非常早期的阶段。这让我想起亚马逊的股东信，杰夫·贝佐斯年复一年地都会说，你看。它……I think we certainly agree that it's... We're still in the very early days for AI. I'm reminded actually of the Amazon shareholder letters where for years and years and years, Jeff Bezos would still say, look. It...

电子商务还处于起步阶段。我认为，任何技术变革都是以几十年而非几年为尺度发生的。It's still day one for ecommerce. And I think with any technological change, these things happen over decades, not years.

你知道，如果你想想摩尔定律，它从二十世纪四五十年代起就一直在发挥作用。互联网已经存在三十年了，但它仍在不断发展。移动技术也是如此，这些技术的发展都需要很长时间。我觉得，如果人工智能技术的机遇和潜力在短短几年内就被充分挖掘，那简直太不可思议了。You know, if you think of Moore's Law, it's played out since the nineteen forties, the nineteen fifties. The Internet is continuing to play out even though it's been around now for three decades. Mobile, like, these things play out over very long periods of time. I think it would be utterly bizarre if the opportunities and the potential of the technology in AI was maxed out over a few short years.

所以，是的，我完全同意。我认为我们仍处于早期阶段。编程领域的应用迅速普及，因为，嘿。你知道，事实证明，程序员很擅长适应……So, yes, I fully agree. I think we are still in the early innings. We've seen rapid adoption in coding because, hey. You know, it turns out that coders adapt...

技术的采用速度非常快。我们才刚刚开始看到网络安全、法律、医疗保健和金融领域采用这项技术。人工智能现在仍然很像处于第一天。澄清一下我的数据点：在已经每年向 Anthropic 支出至少 10 万美元的《财富》500 强企业中，它们在 Anthropic 上的支出平均不到其企业总开支的 0.1%。Adopt technologies very quickly. We're only just starting to see adoption in cybersecurity, in law, in health care, in finance. This very much still feels like day one for artificial intelligence. Just to clarify my data point, it was that among Fortune 500 companies already spending at least 100,000 a year with Anthropic, their Anthropic spending averaged less than 0.1% of their total corporate expenses.

那……我的意思是，如果他们在企业领域的竞争中领先，这场竞争还有很长的路要走。你刚才谈到了专业模型。他们确实把专注于专业模型的公司可能对其模型构成威胁，列为一种风险因素。That's... My point is like, if they're ahead in that game of enterprise, there's a lot of the game left to play. You just talked about specialist models. They did note as a sort of risk factor that that those focused on specialist models might represent a risk to their models.

他们真的必须说这个吗？这可能确实是你必须说的。我是说，……基础模型技术未来可能会有很多种版本。Is that something that they'd have to say, really? It probably is something that you have to say. I mean, there are lots of potential future versions for the... For foundational model technology.

我们认为，最有可能的情况是，前沿实验室将推动前沿技术的整体提升和进步，而智能随后会体现在不同的工具框架和应用层中，这些框架和应用层背后会有开源模型，或者说有足够智能的模型作为支撑。We believe that the most likely one is that the frontier labs will drive the overall gains and the overall progress in frontier technology, and that intelligence will then manifest in different harnesses, different application layers that you will have open source models behind or kind of intelligent enough models that sit behind.

但问题与其说是规模，不如说是价值。我们对 Anthropic 的投资，很大程度上基于这样的信念：大部分价值将归于前沿智能。这个……我是说，你之前在节目里也非常有说服力地阐述过这一点。But the question is not so much of one of volume, it's one of value. And our investments in Anthropic is very much predicated on the belief that the majority of value will accrue to frontier intelligence. That that... I mean, you've you've you've made the case with conviction about that on the on the show before.

私募和公开市场投资Private and public market investing

稍微岔开一下 Anthropic 的话题。我知道这听起来可能有点像是在问：你为什么会想知道这个？不过，你是在贝蒂·吉福德负责私营公司团队，对吧？你能不能简单解释一下，一家大型 IPO 对团队来说意味着什么？如果你们的投资组合只涵盖私营公司，之后又会如何把这家公司交接给公司里的其他团队？Kind of away from Anthropic, you know, I know this might seem sort of like, well, why would you wanna know this? But you run the private companies team at Betty Gifford. Right? Could you just, like, explain what a big IPO would represent for the team and then how it kinda gets handed off to the rest of the firm if your portfolio is just the private side?

因此，在我负责非上市公司业务的同时，我们也将公开市场和非上市市场的投资活动深度整合。So whilst I run the private companies business, we keep our public and private investing activities deeply integrated.

所以，如果你来到爱丁堡这里的投资办公区走一走，就会看到私募公司团队紧挨着我们的北美股票团队，而北美股票团队旁边就是我们的长期全球增长团队，他们投资的是上市公司。当一家公司的股票从我们的私募投资组合进入公开市场后，其中一些投资组合实际上可以长期持有这家公司的股票；另一些投资组合在公开市场上的持有期限则比较有限。但作为私募团队，我们努力确保公募团队能够了解私营公司的管理团队并与他们建立关系，这样一来，当这些企业进入公开市场时，他们就能基于对这些公司的深入了解和已建立的关系，为公募市场客户作出尽可能好的决策。So if you were to come and walk around the investment floor here in Edinburgh, you'd see the private companies team sitting right next to our North American equities team, which is right next to our long term global growth team, and they're investing in public companies. When a company goes public from our private portfolios, some of those portfolios actually can own for the long term into the public markets. Other portfolios have a more limited ownership into, the public markets. But what we try to do as a private team is make sure that our public teams are able to build an understanding, build relationships with private company management teams so that as and when those businesses do come to the public markets, they are able to make the best possible decisions for their public market clients, you know, based on that accrued understanding and the accrued relationships with these companies.