The Tale Of Two AI Economies

The first is the capital economy. The five largest hyperscalers are on track to spend between 600 and 690 billion dollars on AI infrastructure this year, close to double the roughly 400 billion they spent in 2025. Around the third quarter of 2026, their combined cash spending is projected to overtake their operating cash flow, so they are turning to debt and equity markets to close the gap. Moody’s flagged roughly 662 billion in signed data center leases that have not yet commenced, obligations that sit off balance sheet and never appear in the capex figures analysts scrutinize. JPMorgan now models 5.5 trillion in total AI capex through 2030. And of the 2.1 trillion in cloud revenue backlog, about half, 1.05 trillion, traces back to two companies still burning cash, OpenAI and Anthropic. A meaningful share of the revenue moving between the chipmakers, the labs, and the clouds is recycled capital. 

The second is the economy where the work actually happens. MIT’s NANDA study found that 95 percent of enterprise generative AI pilots deliver zero measurable impact to the income statement. Not slow returns. None. IBM put the share of initiatives hitting expected ROI at 25 percent, with 56 percent of CEOs reporting no significant financial benefit. Morgan Stanley found only 21 percent of S&P 500 companies could cite a measurable AI benefit at all. Forty-two percent of companies abandoned most of their AI initiatives last year, up from 17 percent the year before. Eighty-seven percent of executives claim an AI governance framework. Fewer than a quarter have actually operationalized one.

Here is the assessment I run.

First, baseline before you build. If you cannot state the current cost, cycle time, and error rate of the process in numbers, you have no scoreboard and no way to prove a return later.

Second, define the outcome and the measurement in the same breath. A named P&L or risk metric, an owner, and a holdout to compare against. If the only number you track is seats logged in, you are measuring theater.

Third, scope the data and the third party exposure before the demo, not after. What leaves your environment, which vendors touch it, what the model retains, and what your contracts and regulators actually allow.

Fourth, operationalize governance instead of framing it. A policy that lives on a slide is not a control. Access, logging, human approval on the steps that matter, and a shadow AI inventory that reflects what people are really using.

Fifth, set the kill criterion up front. Define the number and the date at which you shut it down. That discipline is what separates the 5 percent capturing value from the 42 percent quietly writing it off.

So before your next budget cycle: are you measuring what the tool actually returns and what it actually exposes, or are you buying the narrative?