Anthropic前景中隐藏的威胁：规模不经济

Anthropic前景中隐藏的威胁：规模不经济

Anthropic PBC作为历史上可能最大的IPO，正在开拓新的领域。 报告列出了仅涉及93页的88个风险因素。 从对人类构成生存威胁的人工智能到AI可能敲诈他人，公司列举了一连串风险，这些风险应使任何潜在的Anthropic投资者警惕。

然而，在计划书中没有明确出现一个潜在的连锁风险因素：规模不经济。

当公司生产每一产出单位的平均成本随着公司的增长而增加时，就会出现规模上的不经济。

你们会认为，鉴于Anthropic为确保其风险因素清单而投入的时间和精力是完整的，它们本会包括这一个因素。 然而，尽管房间里的大象不经济，他们还是错过了这个机会。

我们以前一直用云计算法。 超规模云的基本经济驱动力是规模经济：其规模越大，每个产出单位的成本就越低，导致随着时间的推移的大规模增长和盈利。

由于AI繁荣需要像云计算一样大量建立数据中心，因此可以合理地假设AI会经历与云计算一样的大规模增长所带来的经济效益。

然而，这种假设很可能是错误的。 以及AI为何如此不同。

缩放软件的经济学

在整个企业计算历史中，软件从其微不足道的边际成本中受益。

有了实物商品，从汽车到鸡肉，每个单位都有原料和生产成本。 你做的部件越多成本就越多

软件根本不同。 一旦一个供应商花钱建造一个应用程序，它就可以将该应用程序出售给它能找到的尽可能多的客户，在那里，与开发该应用程序的原始成本相比，发送该应用程序的额外副本的费用微不足道。

这一可忽略不计的边际成本使软件供应商能够因其业务规模而获得巨大利润。

缩小云层的经济学

相反，计算机硬件只不过是部件 — — 你部署得越多，你的材料和生产成本就越高。

云计算在其提供品中兼具软件和硬件。 软件可能有可忽略的边际成本，但每个云提供商必须为每个物理服务器抛出钱。

为应对这一成本挑战，云供应商利用规模经济。

超标者通过部署大量商品服务器来构建云层。 随着每个公司业务的扩展，每个产出单位的费用，例如提供服务器实例或云存储单位的费用，都减少了。

因此，服务器的成本随着规模的缩小而下降，云计算公司也随着增长而变得日益有利可图，尽管其业务模式依赖于硬件和软件。

AI的经济学与软件和云有什么不同

回到Anthropic和其他边疆大语言模型提供商：Ostensibly,AI只是一类软件，Anthropic正在其软件运行于基本上是云层的超尺度数据中心。 这一基本假设隐含在该公司的意向书中。

然而，这一假设是有缺陷的。 前沿模式的供应商需要图形处理装置和其他价格昂贵的硬件，而他们的专门处理器要求随着其软件的快速创新而继续加快。

这一先进的专门处理器要求并没有影响云层的上升。 事实上，每个云提供商最终会随着增长而增加各种专门的硬件实例类型，但这些实例并没有改变云业的基本经济学。

Anthropic并没有奢侈地部署商品硬件来实现规模经济。 相反，它必须建造昂贵的服务器，这些服务器只有在LLMs的处理器要求不断推进时才会变得更贵。

让人类加入经济争论

AI繁荣的人类侧面与云层的生长形成对比，也存在规模化的不经济。

超规模者当然需要一支经验丰富的、具有丰富专业知识的队伍来部署其云层并维持其运行。 但随着他们的成长，他们能够更有效地利用这些人员。 一个管理着一千个实体服务器的管理员，随着基础云平台的成熟，可以很容易地管理一万个或一百万个。

然而，与云供应商不同，Anthropic和其他前沿模式供应商必须不断扩大其团队的专门知识。 新的LLMs不仅仅是老版本的缩放。 每一份都要求人类专家提供新的专门知识。

Anthropic正与竞争对手进行斗争，以快速创新，并在前所未有的时间范围内推出日益强大的模型。 因此，前沿LLM供应商团队遭遇了超规模商从来不需要面对的规模不经济。

软件和硬件和人，哦，我的天！

Anthropic和其余前沿的LLM小贩都遭遇了全新的经济现实，因为他们必须在提高员工技能的同时继续提高硬件能力，从而能够以竞争方式创新软件。

他们不能简单地在商品硬件上运行越来越强大的模型 — — 甚至连他们运行的前一代模型上的硬件都一样。 它们必须以跟上软件革新的速度增加对硬件的投资。

最终的结果是技术行业从未见过的规模不经济。 我们习惯于软件的微不足道的边际成本和云层的巨大经济规模，以至于我们完全没有准备好应对像Anthropic这样的边境LLM供应商必须谈判的规模不经济。

英特莱克斯拿

尽管前沿模型市场仍然相当新颖，我们已经看到其内在规模不经济的影响

LLMs越来越强大，所需硬件越来越昂贵

Anthropic和其他LLM模式的销售商可以尝试通过提高它们更强大的LLMs的效率来应对，或者推出的功率较小，从而降低成本的模型 — 假设它们的客户购买进入这些动作。

我们还可能最终到达一个高原，即模型足够强大，硬件也足够发挥作用，从而使更熟悉的规模经济能够接管。 然而，没有看到这种高原。

因此，潜在的Anthropic投资者必须问，公司在LLM市场的领导地位是强还是弱？在其供品太贵之前，它能创新和成长到什么程度？然后它会做什么？

Jason Bloomberg)是Intellyx的创始人兼总经理，该公司为商业领袖和技术供应商提供数字化转型策略的建议。 他为SiliconANGLE写了这篇文章。 Intellyx的所有内容都是100%的人造，包括这篇文章。 Anthropic不是Intellyx客户。

图片：克雷扬

SiliconANGLE联合创始人John Furrier的留言：

支持我们的使命，通过与CUBE社区接触，保持内容的开放和自由。 加入CUBE的校友信托网络，其中技术领导者相互连接，分享情报，创造机会。

- 15M+观看CUBE视频，为AI、云层、网络安全等各种对话提供动力

- 11.4k+CUBE校友——通过独特的信任网络与11400多名技术和企业领袖建立联系，塑造未来

你是AWS客户吗？在财务上支持SiliconANGLE，从我们的市场门户页面和链接购买您的AWS服务：https://siliconangle.com/aws-marketplace/.

关于 SiliconANGLE 媒体

SiliconANGLE Media由技术视觉家约翰·弗瑞尔(John Furrier)和戴夫·韦兰特(Dave Vellante)创立，构建了产业领先数字媒体品牌的动态生态系统，达到1500+百万精英技术专业人士。 利用CUBEAi.com神经网络，帮助科技公司以数据为导向做出决策

本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。

The hidden threat in Anthropic’s prospectus: diseconomies of scale

As potentially the largest IPO in history, Anthropic PBC is breaking new ground. Its prospectus lists 88 risk factors covering 93 pages alone. From artificial intelligence posing an existential threat to humanity to the possibility AI will blackmail people, the company lists a litany of risks that should make any potential Anthropic investor wary.

One potentially company-sinking risk factor, however, does not explicitly appear in the prospectus: diseconomies of scale.

Diseconomies of scale occur when a company’s average cost of producing each unit of output increases as the company grows.

You would think that given the time and effort that Anthropic devoted to ensuring its list of risk factors was complete that they would have included this one. There is a reason, however, that they missed it, despite the fact that diseconomies of scale is the elephant in the room.

We’ve been down this road before with cloud computing. The fundamental economic driver of the hyperscaler clouds was economies of scale: the larger they became, the lower their cost of each unit of output, leading to massive growth and profitability over time.

Since the AI boom requires massive data center buildouts just as cloud computing does, it might be reasonable to assume that AI will experience the same economic benefits of massive growth that cloud computing enjoyed.

This assumption, however, is likely to be false. To understand why, we must take a look at how the software and cloud businesses scale – and why AI is so different.

The economics of scaling software

Throughout the history of enterprise computing, software has benefited from its negligible marginal cost.

With physical goods, from automobiles to chicken nuggets, every unit has a raw materials and production cost. The more widgets you make, the more those costs add up.

Software is fundamentally different. Once a vendor has spent the money to build an application, it can sell that application to as many customers as it can find, where the cost of delivering additional copies of that application is negligible compared to the original cost of developing it.

This negligible marginal cost enables software vendors to become immensely profitable as their business scales.

The economics of scaling the cloud

Computer hardware, in contrast, is nothing more than more widgets – the more you deploy, the greater your materials and production costs.

Cloud computing combines both software and hardware in its offerings. The software may have negligible marginal cost, but each cloud provider must shell out money for each physical server.

To address this cost challenge, cloud providers take advantage of economies of scale.

The hyperscalers built their clouds by deploying massive numbers of commodity servers. As each firm expanded its business, the cost of each unit of output – say, the cost of providing a server instance or a unit of cloud storage – decreased.

As a result, the cost of servers went down as they scaled, and the cloud computing firms became increasingly profitable as they grew, despite the fact that their business models depended on hardware as well as software.

How the economics of AI differs from both software and the cloud

Back to Anthropic and the rest of the frontier large language model providers: Ostensibly, AI is just a type of software, and Anthropic is running its software in hyperscaled data centers that are essentially clouds. This underlying assumption is implicit in the company’s prospectus.

This assumption, however, is flawed. Frontier model vendors require graphics processing units and other premium, expensive hardware from day one – and their specialized processor requirements continue to accelerate as they rapidly innovate their software.

This advancing specialized processor requirement didn’t impact the rise of the cloud. True, each cloud provider eventually added various specialized hardware instance types as they grew, but those instances didn’t change the fundamental economics of the cloud business.

Anthropic, in contrast, doesn’t have the luxury of deploying commodity hardware to achieve economies of scale. Instead, it must build out expensive servers that only get more expensive as LLMs’ processor requirements continue to advance.

Adding humans to the economic argument

There are also diseconomies of scaling the human side of the AI boom that contrast with the growth of the cloud.

The hyperscalers certainly require an experienced workforce with deep expertise to deploy their clouds and keep them running. But as they grew, they were able to leverage such personnel more efficiently. An admin who managed a thousand physical servers could easily manage 10,000 or a million as the underlying cloud platforms matured.

Unlike the cloud providers, however, Anthropic and the other frontier model vendors must continually scale up the expertise of their teams. New LLMs aren’t simply scaled-out versions of older ones. Each one requires new expertise from expert humans.

Anthropic is battling its competitors to innovate quickly and roll out increasingly powerful models on unprecedented timeframes. As a result, frontier LLM vendor teams experience diseconomies of scale that the hyperscalers never had to face.

Software and hardware and people, oh my!

Anthropic and the rest of the frontier LLM vendors suffer from a fundamentally novel economic reality, because they must continue to increase their hardware capabilities while upskilling their workforce so they can competitively innovate their software.

They can’t simply run increasingly powerful models on commodity hardware – or even the same hardware they ran the previous generation model on. They must increase their investments in hardware at a pace that keeps up with the pace of software innovation.

The end result is a diseconomy of scale the technology industry has never seen before. We are so used to the negligible marginal cost of software and the massive economies of scale of the cloud, that we are simply unprepared for the diseconomies of scale that frontier LLM vendors like Anthropic must negotiate.

The Intellyx take

Even though the frontier model marketplace is still quite new, we’re already seeing the impact of its inherent diseconomies of scale as token costs become an increasingly significant expense for consumers of AI.

As LLMs become more powerful and the hardware they require becomes an increasingly expensive bottleneck, it’s just a matter of time until token costs become prohibitively high for a range of AI-based applications.

Anthropic and the other LLM model vendors can try to respond by increasing the efficiency of their more powerful LLMs or rolling out less powerful, and thus less expensive models – assuming their customers buy into these moves.

We may also eventually reach a plateau where the models are sufficiently powerful and the hardware is sufficiently performant, thus enabling more familiar economies of scale to take over. However, no such plateau is in sight.

The question potential Anthropic investors must ask, therefore, is whether the company’s leadership position in the LLM market is a strength or a weakness? Just how far can it innovate and grow before its offerings are too expensive? And what does it do then?

Jason Bloomberg is founder and managing director of Intellyx, which advises business leaders and technology vendors on their digital transformation strategies. He wrote this article for SiliconANGLE. All of Intellyx’s content is 100% human generated, including this article. Anthropic is not an Intellyx customer.

Image: Craiyon

A message from John Furrier, co-founder of SiliconANGLE:

Support our mission to keep content open and free by engaging with theCUBE community. Join theCUBE’s Alumni Trust Network, where technology leaders connect, share intelligence and create opportunities.

- 15M+ viewers of theCUBE videos, powering conversations across AI, cloud, cybersecurity and more

- 11.4k+ theCUBE alumni — Connect with more than 11,400 tech and business leaders shaping the future through a unique trusted-based network

Are you an AWS customer? Support SiliconANGLE financially by buying your AWS services from our Marketplace portal page and links: https://siliconangle.com/aws-marketplace/

About SiliconANGLE Media

Founded by tech visionaries John Furrier and Dave Vellante, SiliconANGLE Media has built a dynamic ecosystem of industry-leading digital media brands that reach 15+ million elite tech professionals. Our new proprietary theCUBE AI Video Cloud is breaking ground in audience interaction, leveraging theCUBEai.com neural network to help technology companies make data-driven decisions and stay at the forefront of industry conversations.

Auto-fetched and translated by AI Atlas. All rights belong to the original author.