How Long Can the GPU Shortage Last? A Deep Dive into AI Compute Economics
Cointime reports: The GPU shortage that has defined the AI boom may persist for another two to three years, but the warning signs of a turning point are already visible. Understanding the durability of demand requires distinguishing between two types of workloads: bounded and unbounded tasks.
Bounded vs. Unbounded Demand
Bounded tasks have a completion point. Once a simple process is automated well enough, buyers shift to the cheapest “good enough” model. Unbounded tasks, by contrast, continuously reward more intelligence. Software engineering can test more solutions and deliver faster; investment and trading teams can constantly hunt for alpha, with no diminishing marginal value from additional units of intelligence. AI labs pour compute into next-generation models in pursuit of AGI. Anthropic and OpenAI reportedly allocate about 60% of their compute to internal AI training.
These workloads underpin the scarcity of frontier compute because there is no saturation point for additional intelligence.
Frontier Demand Is Strong but Highly Concentrated
Frontier demand is concentrated in a handful of open-ended use cases: software engineering, investment and trading, and AI R&D itself. This concentration cuts both ways. On the bullish side, penetration is still far from uniform. On the risk side, today’s demand is reflexive and can reverse quickly.
Cheaper Intelligence Must Create More Demand Than It Destroys
Chinese open-source models and broader model competition are driving down the cost of intelligence—a structural and irreversible trend. For bounded tasks, this brings direct substitution risk. As smaller or open-source models catch up, buyers optimize costs, and the premium for frontier models narrows.
Lower prices also expand usage. The bet is that this “Jevons-style” reaction is strong enough that paid consumption grows faster than prices fall. If token prices drop 50% while paid usage rises 40%, revenue still falls to 70% of the starting point before accounting for new capacity. That’s why total token consumption growth alone is insufficient. Fortunately, current data aligns with a Jevons-style response: cheaper intelligence opens new uses, agents run more steps, and each workflow consumes more tokens.
The Cycle Is Reflexive in Both Directions
The upcycle is self-reinforcing: higher token spending improves models and products, attracting more revenue and capital, which funds more token spending and factory construction. The same loop can run in reverse, especially with highly concentrated demand. A demand or funding shock can cut token budgets, or simply capacity growth can cause utilization to fall. Weaker revenue and valuations then restrict financing, prompting AI companies to budget more cautiously, slash internal training spend, and push more compute onto the market. Falling utilization may trigger commitments from chip sellers and hyperscalers, dragging down their market caps. Given their size and importance, the impact could spread to broader markets, further weakening demand and financing capacity.
The first pressure may come from leveraged new cloud providers with concentrated customers—they own the factories and carry the debt.
Early Warning Signs
The base case is that demand will outpace supply for another two to three years. But the key question is which early signs will herald a turning point. Start with the oldest GPUs: when supply catches up, they should be the first to lose overflow demand. Next, watch the interest expense-to-revenue ratio of leveraged new cloud providers. Rising burdens mean financing costs are growing faster than revenue. Finally, track the speed at which these two trends deteriorate. Faster erosion in old GPU economics or a steeper rise in interest burdens could signal a turning point before overall utilization data does. Can new cloud providers raise compute prices faster than debt rolls over at higher rates?
