Anthropic PBC’s impending initial public offering, which could rank as the largest in history, is accompanied by a prospectus that devotes 93 pages to 88 risk factors. Those disclosures range from artificial intelligence posing an existential threat to humanity to the possibility that AI will blackmail people. One risk that could prove company-sinking does not appear explicitly in the document: diseconomies of scale, which occur when a company’s average cost of producing each unit of output rises as the company grows.
The omission stands out because the economics of cloud computing might suggest the opposite outcome. The fundamental driver of the hyperscaler clouds was economies of scale, in which getting larger lowered the cost of each unit of output and produced massive growth and profitability over time. Because the AI boom requires massive data center buildouts just as cloud computing does, it might be reasonable to assume AI will enjoy the same benefits. That assumption is likely to be false.
Software has historically benefited from negligible marginal cost. Once a vendor has spent the money to build an application, it can sell that application to as many customers as it can find, and the cost of delivering additional copies is negligible compared with the original development cost. Hardware, by contrast, is nothing more than more widgets, with materials and production costs that add up with every unit deployed. Cloud computing combines both, and providers addressed the hardware cost challenge through economies of scale, deploying massive numbers of commodity servers so that the cost of each unit of output fell as they expanded.
Anthropic and the other frontier large language model providers ostensibly run software in hyperscaled data centers that are essentially clouds, an assumption implicit in the prospectus. But 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. Anthropic cannot deploy commodity hardware to achieve economies of scale; instead it must build out expensive servers that only get more expensive as LLM processor requirements advance.
The human side of the AI boom presents a similar contrast with the cloud. Hyperscalers required an experienced workforce, but as they grew they leveraged such personnel more efficiently, and an admin who managed a thousand physical servers could easily manage 10,000 or a million as the underlying cloud platforms matured. Anthropic and the other frontier model vendors must continually scale up the expertise of their teams, because new LLMs are not simply scaled-out versions of older ones and each requires new expertise from expert humans.
Anthropic is competing to innovate quickly and roll out increasingly powerful models on unprecedented timeframes. As a result, its frontier LLM vendor teams experience diseconomies of scale that the hyperscalers never had to face.
More AI news from TechManNews.
