The Hidden Threat in the Anthropic Prospectus: Diseconomies of Scale

Potentially the largest IPO in history Anthropic PBC takes new paths. His brochure lists 88 risk factors, which alone cover 93 pages. From artificial intelligence posing an existential threat to humanity to the possibility of AI blackmailing people, the company lists a whole host of risks that should make any potential Anthropic investor wary.

However, a potentially damaging risk factor does not appear explicitly in the prospectus: Economies of scale.

Economies of scale occur when a firm’s average cost of production for each unit of output increases as the firm grows.

Given the time and effort Anthropic has put into completing their list of risk factors, you would think they would have included this list. But there’s a reason they missed it, even if economies of scale are the biggest problem.

We have already been down this path with cloud computing. The fundamental economic driver of hyperscaler clouds was Economies of scale: As they grew larger, the cost per unit of production decreased, leading to massive growth and profitability over time.

Since the AI ​​boom, like cloud computing, requires a massive expansion of data centers, one can assume that AI will benefit from the same economic benefits of massive growth as cloud computing.

However, this assumption may be wrong. To understand why, we need to look at how software and cloud companies 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.

For physical goods, from cars to chicken nuggets, each unit has raw material and production costs. The more widgets you create, the higher the cost.

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

This negligible marginal cost allows software providers to become hugely profitable as they scale their business.

The economics of scaling the cloud

In contrast, computer hardware is nothing more than more widgets – the more you put in, the higher your material and production costs.

Cloud computing combines both software and hardware in its offerings. While the marginal cost of the software is negligible, every cloud provider has to spend money on every physical server.

To address this cost challenge, cloud providers are taking advantage Economies of scale.

The hyperscalers built their clouds by deploying large numbers of standard servers. As each company expanded its business, the cost of each unit of production fell – for example, the cost of deploying a server instance or a cloud storage unit.

As a result, the cost of servers fell as they scaled, and cloud computing companies became more profitable as they grew, even though their business models relied on both hardware and software.

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

Back to Anthropic and the rest of the major language model providers: AI is supposedly just a type of software, and Anthropic runs its software in hyperscale data centers that are essentially clouds. This underlying assumption is contained in the company’s prospectus.

However, this assumption is flawed. Frontier vendors need graphics processors and other high-end, expensive hardware from day one – and their specialty processor requirements continue to increase as they rapidly evolve their software.

This increasing need for specialized processors has not influenced the rise of the cloud. While each cloud provider eventually added different specialized hardware instance types as it grew, these instances did not change the fundamental economics of the cloud business.

In contrast, Anthropic doesn’t have the luxury of using off-the-shelf hardware to achieve economies of scale. Instead, expensive servers must be built, which will only become more expensive as LLMs’ processor requirements continue to increase.

Adding people to the economic argument

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

The hyperscalers undoubtedly need an experienced workforce with in-depth expertise to deploy and keep their clouds running. But as they grew, they were able to use that staff more efficiently. An administrator managing a thousand physical servers could easily manage 10,000 or a million as the underlying cloud platforms mature.

However, unlike cloud providers, Anthropic and the other frontier model providers must continually expand the expertise of their teams. New LLMs are not simply scaled versions of older ones. Each requires new expertise from experienced people.

Anthropic is competing with its competitors to innovate quickly and introduce ever more powerful models in unprecedented time frames. This results in frontier LLM vendor teams experiencing diseconomies of scale that the hyperscalers never faced.

Software and hardware and people, oh my!

Anthropic and the rest of the frontier LLM vendors are facing a fundamentally new economic reality as they must continue to expand their hardware capabilities while upskilling their workforces so they can competitively innovate their software.

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

The end result is a diseconomies of scale the tech 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 not prepared for the economies of scale that frontier LLM providers like Anthropic must contend with.

The Intellyx version

While the market for Frontier models is still fairly new, we are already seeing the impact of its inherent diseconomies of scale as token costs become an increasingly significant cost driver for AI consumers.

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

Anthropic and the other LLM model providers may attempt to respond by increasing the efficiency of their more powerful LLMs or introducing less powerful and therefore more cost-effective models – assuming their customers accept these steps.

We may also reach a plateau at some point where the models are sufficiently powerful and the hardware sufficiently powerful that more familiar economies of scale can take over. However, such a plateau is not in sight.

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

Jason Bloomberg is the founder and CEO of Intellyx, which advises business leaders and technology providers on their digital transformation strategies. He wrote this article for SiliconANGLE. All IntellyxThe content is 100% human-generated, including this article. Anthropic is not an Intellyx customer.

Picture: Crayon