Circularity and Concentration Risk

The most important thing to understand about the current neocloud boom is that a large share of its money moves in circles. Tracing those circles is the single most valuable analytical habit you can build, because circular demand and genuine end demand look identical on an income statement and behave completely differently under stress.

Start with the canonical loop. A chip vendor with enormous profits takes an equity stake in a neocloud. The neocloud uses capital, some of it from that very investment, much of it borrowed, to buy the vendor chips. The vendor books the chip sale as revenue and marks up the value of its stake as the neocloud valuation rises. The neocloud, now equipped, signs a capacity contract with a large customer, often a hyperscaler or a lab that the vendor is also close to. Each participant records a transaction that flatters its own numbers, and the same underlying dollars appear as revenue at more than one stop around the ring.

None of this is necessarily improper, and much of it reflects real activity, chips are shipped, clusters are built, models are trained. The danger is interpretive. When money circulates among a few entangled parties, aggregate demand looks larger, more diversified, and more independent than it actually is. An outside observer sees many companies all growing and concludes the market is broad and deep. In truth a handful of actors may be passing capital around a small loop, and the apparent breadth is an illusion created by counting the same activity multiple times.

Concentration is the reason this matters so much. The loop is not made of thousands of independent participants; it is made of a few very large ones. A single chip vendor dominates supply. A small number of hyperscalers and labs dominate demand. A modest set of lenders dominates the financing. Because so few nodes carry so much of the system, the retreat of any one of them propagates everywhere at once. If the vendor slows its strategic investments, several neoclouds lose capital and validation simultaneously. If one hyperscaler cancels or fails to renew a block of capacity, the shock is not absorbed by a diverse market, it lands directly on the specific operators exposed to that name.

This is what distinguishes systemic risk from an ordinary business risk. In a diversified market, one customer retreat is a rounding error. In a concentrated, circular one, it is a chain reaction. The contracted revenue that made a neocloud financeable was underwritten on the assumption that the anchor would keep paying; the anchor willingness to keep paying may itself depend on the same demand conditions that, if they soften, soften for everyone at once. Correlations that looked independent in good times turn out to be the same bet wearing different labels.

There is also a valuation feedback loop worth naming. As neocloud valuations rise, the vendor stakes rise, strengthening the vendor balance sheet and its capacity to invest in still more neoclouds, which buy still more chips, which lift valuations further. On the way up this is a virtuous spiral that mints paper wealth at every turn. But spirals run in both directions. If sentiment turns and valuations fall, the same linkages transmit the decline: the vendor stakes shrink, financing tightens, contracts get renegotiated, and the neoclouds most dependent on the loop find every source of support contracting simultaneously.

The practical discipline is to trace demand to its ultimate source. For any dollar of a neocloud revenue, ask who is really paying and why. Is the customer buying compute to sell a product to independent end users who pay real money for intelligence? Or is the customer buying compute because it is entangled with the neocloud survival, or because it, too, is spending investor capital in anticipation of demand that has not yet materialized? End demand, someone outside the loop paying for a service they value, is the only kind that survives a downturn. Circular demand evaporates the moment the capital stops circulating.

None of this means the boom is hollow or that every neocloud is a house of cards. Real models are being trained, real products are being sold, and some operators sit on genuinely independent, diversified demand. The point is analytical humility: in a concentrated, circular system, headline growth is the least reliable signal, and you cannot take aggregate numbers at face value. You have to disaggregate the loop, identify which nodes carry the load, and size the shock if the biggest of them steps back. The neoclouds that survive that exercise are resting on end demand. The ones that fail it are resting on a circle, and circles do not hold weight when the music stops.