Data centers and frontier models sit in the same ecosystem, but they have very different economics.

This is the second article in a four-part weekly series. The series based on my new 12,000-word research paper, The Economics of AI Infrastructure, and the financial model used in that analysis. The first article AI may succeed spectacularly while infrastructure investments fail argued that AI demand growth and capital recovery are different questions. This article explains why the AI ecosystem should not be analyzed as a single business.

There are two distinct AI businesses aside from the supply of hardware such as GPUs and memory chips, power supplies and cooling systems. The latter are the picks and shovels in the AI gold rush.

The first is physical infrastructure construction and operation: data centers, power, cooling, grid interconnection, fiber connectivity, accelerators, servers, memory and high-speed networking. Its economics resemble other capital-intensive infrastructure sectors, including submarine cables, fiber networks, towers, cloud data centers and electric-power infrastructure.

The second is frontier-model and platform development: OpenAI, Anthropic, Google DeepMind, Meta, xAI, DeepSeek, Moonshot, Alibaba/Qwen and others competing to build increasingly capable models and ecosystems. Its economics, business models and competitive dynamics resemble software platforms more than fiber networks.

That distinction matters. Infrastructure returns depend on capex, utilization, electricity cost, financing conditions, obsolescence and commodity supply competition. Frontier models depend on scale economies, innovative differentiation, unique algorithms, proprietary or preferential data access, distribution, developer adoption, brand, trust, enterprise procurement and other ecosystem effects. 

However, there is some vertical integration with Big Tech companies developing AI models and becoming increasingly asset-heavy by also making large AI data center investments. For example, Google DeepMind’s main AI models — Gemini, PaLM, BERT, LaMDA — are all built in-house, owned, and developed by Alphabet’s Google. Alphabet is investing massively in AI data centers as well.

A gigawatt-scale AI data center has bounded physical capacity. Its commercial upside is limited by utilization and price. A successful frontier platform might be monetized across advertising, search, cloud, enterprise workflows, application software, devices, payments and commerce. Its upside can be far larger and is more speculative.

This is why simple comparisons between AI and the telecom boom can mislead. AI infrastructure may resemble fiber network economics. GPU sales are analogous to optical equipment sales. Frontier AI platforms better resemble Google Search, Microsoft Office, iOS, Android or NVIDIA CUDA ecosystems.

Some commentators argue that frontier AI models will become commodities. They may be partly right for many applications. Open-weight models, including strong Chinese models, already create pricing pressure. Many tasks will be well served by cheaper models that are good enough.

But complete commoditization is not my base case or expectation. Frontier-model development has scale economies and experience-curve effects. Some providers may secure superior access to training data. Some will integrate more effectively with tools, agents, enterprise systems and cloud platforms. Others will benefit from distribution through existing ecosystems such as Microsoft, Google, Meta or Apple.

Technology markets often remain concentrated even when the underlying technologies become widely available. Search engines, operating systems, office productivity software, smartphone platforms and GPU software ecosystems all show how developer communities, user habit, brand familiarity and switching costs can sustain differentiation and ensure handsome profits for many years. For example, while Android is open source, Google rules the roost with its Play Store and other ecosystem advantages in smartphone implementations.

Infrastructure is different. A GPU-hour from one well-connected, well-powered data center can be more substitutable than a frontier model embedded in a broader software ecosystem. Location, power and water access, sovereignty and latency matter, but physical compute is generally more fungible and easily switchable than a platform relationship.

That does not mean all frontier-model companies will prosper. Many may fail. Early leaders in technology markets often lose their positions. AltaVista, Netscape, Myspace, Lotus 1-2-3 and WordPerfect all appeared strong for a while.

The key point is that infrastructure and platforms should not be placed in the same analytical bucket. A model developer may pursue enormous strategic upside even while the data centers serving it face commodity-like pricing and inadequate returns on capital invested. The latter might be owned by the former or an entirely separate company and investors.

The third article in this series next week examines the most under-discussed variable in AI infrastructure economics: utilization. If a data center is built and operated well, the utilization rate will depend mostly on demand and competing supply, which are highly uncertain.

Keith Mallinson, founder of WiseHarbor, has more than 30 years of experience in the ICT sector as a technical, commercial and financial research analyst, consultant, and testifying expert witness.