VIP Deep Dive: What Are Companies Actually Getting for All That AI Spending?

Amazon is raising its 2026 AI infrastructure spending to $220 billion, the FT reported. Apple is warning that memory prices will hurt margins. Anthropic just raised at a valuation that implies AI labs are bigger than the GDPs of mid-sized countries. The New York Times asked a simple question this week: what are companies actually getting for all that AI spending?

The Tokenomics Frame

A new field of "tokenomics" has emerged to measure the return on the money flowing into AI. The basic unit is the token: how much does it cost to produce one, and how much value does it create? Early tokenomics work suggests the gap between the two is wider than boards have been willing to admit.

"We are buying the most expensive compute in the history of information technology, and we have very little idea what the marginal output is worth," one CFO told the NYT.

Three Things the Numbers Say

Three patterns are emerging from the data: training-capex is up sharply, inference revenue per dollar of capex is falling, and the share of AI spend going to the same handful of vendors is rising. None of these are comfortable for hyperscale operators. None are comfortable for the AI labs either. The next phase of the AI economy, the FT argues, will be defined by which players can close the gap.