The AI infrastructure boom is built on data nobody actually has

AI infrastructure spending has grown large enough to pose two opposite risks. If the buildout doesn’t pay off, the losses could ripple through the global economy. If it does pay off, and AI adoption outpaces even what the companies building this infrastructure are planning for, the resulting labor displacement could be just as disruptive.

Both scenarios hinge on metrics the industry can’t measure well.

Image credit: Ulla

How much does AI traffic (the data flowing between and within data centers as models are trained and queried) actually cost? And how much of it is there?

Solving that requires answers to more specific questions: How much do tokens cost under different scenarios? What are the profit margins across different use cases? How do traffic costs, volumes, and trajectories vary from provider to provider?

Financial filings from vendors and operators in this space as well as traffic studies from both Cisco and my firm MTN Consulting give some insight into these questions.

But overall, the AI boom is proceeding based largely on emotion. There’s a lack of hard, vetted industry data on AI traffic and its implications that makes the current boom precarious.

The measurement gap

The clearest finding, from our review of over 30 Optical Fiber Communications conference papers in 2026, is that no comprehensive public study of AI traffic volumes or patterns exists. Hyperscalers, like Amazon Web Services, Google, and Microsoft, do not share traffic data after all.

Cisco’s May 2026 report adds some insight since it measures live AI inference traffic across service-provider networks. AI inference refers to what happens every time an AI model processes an input and produces an answer.

See also: SEA’s data laws chase investments, but will they protect data?

However, Cisco frames itself as “the first in an annual series.” Data from one baseline year does not yet support reliable forecasting, certainly not enough for hyperscalers to underwrite over US$600 billion in annual capital expenditure (capex) with confidence.

The scale of that capex is itself contested. MTN Consulting projects 2026 hyperscale capex at just under US$700 billion, up from about US$500 billion in 2025. However, the CEO of GlobalFoundries said, at the company’s May 7 investor day, that more than US$700 billion comes from the top four hyperscalers alone.

Meanwhile, the Bank for International Settlements (BIS) put the top five hyperscalers’ combined 2025 and 2026 capex above US$1 trillion.

This is only the tip of the iceberg when it comes to measurement problems. There are many large private equity-backed data center builders who don’t report financials.

What makes AI traffic different

Follow the backlogs

Boom, bust, or both?

Opacity all the way down

What we actually know