Rules of relevance

Everyone's heard of confirmation bias. Literary theorist Paul Fry takes things a step further: he argues that everything we perceive is preselected by our background, training, and analytical habits. Bias implies distortion, spinning data we've already seen. Fry describes an earlier operation, in which our minds build "rules of relevance" that render some information functionally invisible before we ever get the chance to interpret it.

The AI investment boom illustrates the problem. An engineer at an AI lab views the explosive growth in compute capex as a rational response to a binding technical constraint; a statistician sees an accelerating series with no precedent in the sample, and a distribution of outcomes that widens by the day. Neither is wrong. They were simply trained to notice different things.

In Rules of Relevance, our newest whitepaper, we explore how such “learned blindness” can become a major source of market risk in periods of rapid technological and geopolitical change.  It may not be changes in the data that trigger market volatility ahead, but a change in the lens through which investors interpret that data.

Key takeaways

What is the "rules of relevance" thesis behind Carlyle's new paper on AI and markets? 
Disagreements about AI capex, credit, and Treasury yields increasingly turn on which facts count as evidence, not on competing interpretations of the same facts. When market participants share an interpretive framework, uncertainty attaches to the next data point. When the framework itself is in dispute, asset prices become prone to sudden jumps as the same phenomenon gets viewed through a new lens.

Why haven't credit spreads widened along with the AI capex boom? 
Markets are currently pricing AI capex the way a scaling engineer would, as a rational response to a technical constraint, rather than the way a statistician would, as an unprecedented structural break with a widening range of outcomes. Required returns and credit spreads have not moved to reflect the scale or duration of the buildout; market-wide risk premia have actually declined since it began.

How has depreciation accounting flattered AI-era tech earnings?
AI compute capacity is typically depreciated over a roughly six-year useful life, so every $100 of semiconductor revenue has been offset by just a $17 depreciation charge in the year of sale. Combined with an estimated $150 billion in reported profits from paper gains on stakes in frontier labs, this accounting treatment has made reported tech earnings growth look stronger than the underlying cash economics.

What's actually backing the surge in AI-related debt financing?
More than 80% of anticipated AI capex through 2030 is expected to require external financing¹. In many of the resulting off-balance-sheet structures, debt service is secured against chip residual values, but credit enhancements — residual-value guarantees and senior-tranche wrappers — mean lenders are effectively underwriting a handful of investment-grade counterparties rather than the economics of the hardware itself.

How much economic value is really in AI providers' contracted backlogs?
By Q2 2026, top AI compute providers had disclosed more than $2.5 trillion in contracted backlog and remaining performance obligations². Once the present value of the capital required to deliver that backlog is netted out, the paper estimates the net economic value at roughly $375 billion, about 15% of the nominal contract figure².

Is AI borrowing the main driver of higher Treasury yields?
Not on its own. The bigger shift is the reversal of the corporate savings glut that kept rates low from 2009 to 2019: US corporate net savings are down 86% relative to the same point in the last cycle³, as hyperscalers have moved from being the economy's largest suppliers of liquidity to its largest consumers of it.

What does a reversing globalization mean for US deficit financing?
Cumulative US borrowing from the rest of the world has reached 118% of GDP since 2000, financed at an effective cost below zero thanks to persistently strong foreign demand for Treasuries⁴. As trade volumes stagnate and reverse, that implicit borrowing subsidy is fading, and foreign capital is increasingly directed toward private-sector assets instead.

What should asset allocators do given this uncertainty?
The paper offers five specific implications we believe allocators should consider: most portfolios likely already carry sufficient direct AI exposure; returns from past technological revolutions have tended to accrue to users of the technology more than its builders; fiscal strain is more likely to show up as inflation than as credit risk in Treasuries; credit risk currently offers better compensation than duration risk; and we believe that defense and energy are the clearest "deglobalization" plays. 

^{1 J.P. Morgan Global Research, midyear outlook, June 2026
2 Carlyle analysis; Company filings as of September 2, 2026
3 Carlyle analysis; Bloomberg, IMF WEO Database. September 2026
4 This represents the summed total of current account deficits, net of primary income, since 2000}