The Subsidised Horizon
There is a kind of confidence that only borrowed time can sustain.
The AI market is not, in the economic sense most people intend, a market. It is an experiment in deferred accounting — a negotiation between balance sheets over who absorbs the loss, conducted in full public view, with enough velocity to look like growth. Two structural problems sit at the centre of this, and neither is obscure. They are simply not discussed in the register they deserve.
The first: the compute infrastructure that powers accessible AI is not, and has not been, priced at cost. What the end user pays — in API credits, subscription tiers, hardware access — reflects a deliberate choice by the major labs and cloud providers to operate at a loss in order to capture the market. In The Prime Effect, I described how modern businesses have moved away from product-level profitability toward something more architectural — subsidise the front door, build the habit, recover margin later across the system. AI compute is that strategy operating at infrastructure scale rather than consumer product scale. But it carries a particular problem that the consumer playbook did not. The infrastructure is not software. You cannot ship a cheaper version of a GPU cluster once user expectations have formed around the current one. The cost floor is physical. And the gap between what users pay and what the compute actually costs — that gap is someone's balance sheet problem. For now, it is venture capital and big-tech cross-subsidisation. Eventually it will be the user, or it will be a contraction in what AI actually delivers at accessible price points. Probably both, in sequence.
When does "eventually" arrive? This is the question the industry has developed a professional allergy to answering. But subsidies require willing subsidisers, and the appetite for indefinite subsidy is not bottomless. We are already seeing re-pricing at the margins — adjustments to free tiers, quiet reductions in context limits for lower-paying customers, metered throttling dressed as policy. These are not signs of healthy normalisation. They are the early signals of a floor beginning to rise.
The second problem is structural in a different, and in some ways more troubling, sense. What circulates through the AI market as investment capital is not, in all cases, making the directional journey that the word "investment" implies — capital moving outward, toward the creation of external value, taking on risk in exchange for future return. Some portion of it is rotating. Nvidia invests in OpenAI. OpenAI purchases more Nvidia GPUs. The capital makes a circuit and returns, having inflated both balance sheets and a number of share prices in transit. The money moves. Nothing external is built.
This is not uniquely an AI phenomenon. The pattern is older and more general. Large investment firms routinely hold equity stakes in companies whose shares they simultaneously package into investment vehicles and sell to retail clients. The fund benefits from appreciation in the underlying. The underlying benefits from capital inflows driven by the fund's own promotion. The retail investor, sitting at the end of this chain, is the one entity for whom the transaction is what it appears to be — a bet on future value — while everyone upstream is playing a different, more circular game. This is not conspiracy. It is structure. The incentives produce the geometry, and the geometry produces the outcome.
What makes the AI iteration of this particularly interesting is the speed of the loop. In traditional capital markets, these circuits are slow enough to be invisible to most participants. In AI, the loop is short and the actors are few and named and very publicly visible. Nvidia's market position is inseparable from the success of the labs it invests in. The labs' success is inseparable from Nvidia's continued dominance of the GPU stack. Each player's valuation is partly a function of the other's. What looks like a diversified capital structure is, under pressure, a single position held in several containers.
Both problems share an underlying shape. They are both about closed systems being mistaken for open ones. Subsidised compute mistakes a temporary cost deferral for a sustainable price. Rotating capital mistakes internal velocity for external growth. In The Prime Effect, the subsidy was funded by a genuine external profit engine — AWS bankrolling the consumer surface, Costco's volume bankrolling the hot dog. The compounding problem in AI is that the external capital funding the subsidy is not always fully external. When the subsidiser and the subsidised are investing in each other, the system is more closed than the balance sheets suggest. In thermodynamic terms: the energy recirculates, entropy increases, and no net work is extracted outward. The system looks active. It is not producing.
What happens when these two pressures resolve simultaneously — when the floor rises and the loop tightens? That is a question the industry is not, as yet, asking loudly enough.