Frontier AI capital spending for 2028 is projected to exceed the annual budget of France. That figure, cited by independent analyst Giovanni Cattani in a long post on X, is the jumping-off point for a more specific claim: a meaningful share of the revenue used to justify that spending may be generated by the same buildout it is meant to justify.
Cattani’s argument starts with a labor taxonomy. He splits AI tasks into bounded (tax prep, simple frontend code) and unbounded (AI research, trading strategy, software at scale), and separately into short-horizon and long-horizon work, borrowing the horizon framing from METR’s model-capability benchmark. Bounded tasks, he argues, will get commoditized by cheap, non-frontier models because the payoff is cost savings with a ceiling. Unbounded tasks behave differently: since there is no ceiling on the value of finishing research faster or trading with more alpha, buyers will pay a premium for the best available model regardless of price, which keeps demand concentrated on frontier labs.
His estimate is that three unbounded categories, AI research and development, software engineering at startups and AI-native companies, and quantitative trading, together account for roughly half of frontier labs’ inference revenue. He proposes rough splits of about 20 percent, 15 percent, and 15 percent respectively. Cattani is explicit that these are his own estimates, not disclosed figures from any lab.
The mechanism he flags is reflexivity: each of these three buyer groups converts token spend into revenue or capital gains quickly, then reinvests a share of that gain into more tokens. A lab spends on R&D compute, ships a stronger model, raises more money, and spends more on R&D. A startup ships more AI-assisted code, raises a bigger round, and buys more inference. A trading firm sees a strategy work, redeploys the profit into more AI-driven trading. None of this requires new external demand; it can be the buildout paying itself back.
Cattani’s own numbers illustrate how large the exposure could get. He notes that if Anthropic monetized its full expected 2026 compute capacity at the revenue-per-gigawatt rates discussed on Dwarkesh Patel’s podcast, the arithmetic lands at $500 billion annualized by the close of 2026, which on revenue alone would put it third globally. He also estimates that a contraction in any one of the three demand categories would strip 15 to 20 percent of frontier token spending straight out, rising toward half once the knock-on effects between labs, startups and trading firms are counted.
The skeptical read is straightforward: reflexivity arguments are close to unfalsifiable while the cycle is still running upward, since any revenue growth can be attributed to either genuine demand or self-reinforcing spend depending on the observer’s priors. Circular-revenue concerns have also trailed every major infrastructure buildout, from fiber in the late 1990s to cloud data centers in the 2010s, and some of those cycles ended in gluts while others became the backbone of decades of growth. Cattani’s post does not attempt to distinguish which case this is, and says so.
For an operator evaluating a vendor whose growth story leans on this dynamic, the useful question is not whether AI demand is reflexive in the aggregate. It is which of that vendor’s own customers, funders, or trading counterparties would still be paying at the same volume if frontier compute spending flattened for four consecutive quarters. That answer, not the aggregate revenue chart, determines whether a contract renewal is exposed to the loop Cattani describes.
Giovanni Cattani laid out this framework in a long-form post on X on September 1, 2026, presenting it as his own analysis and explicitly labeling the underlying figures as estimates.
