The chart is one of the clearest rebuttals to the claim that AI compute is already being overbuilt. GPU rental prices are not collapsing as new capacity comes online. They are rising across almost every generation. B200 rates have moved toward roughly $5.50 to $5.80 per GPU-hour, H200 rates are approaching $3, while even older H100 and A100 capacity continues to reprice upward. The most important signal is not the absolute premium on B200. It is that older chips are also becoming more expensive. If this were merely temporary excitement around the newest architecture, B200 prices could remain high while H100 and A100 rates declined as customers substituted toward newer hardware. Instead, demand appears broad enough to absorb new systems without materially weakening pricing for the installed base. That suggests the industry is constrained by total usable compute, not simply by access to the latest chip. This supports the argument that hyperscalers and neoclouds are currently under-earning on older contracts. Customers that secured GPU capacity in 2024 or 2025 at fixed rates are benefiting from pricing materially below today’s spot market. As those contracts expire, providers should have room to reprice capacity upward, improving cloud revenue and operating cash flow without requiring an equivalent increase in physical infrastructure. It also strengthens the Jevons paradox thesis. Chips and models are becoming more efficient, but lower cost per task is expanding the number of economically viable workloads. Enterprises are deploying more coding agents, research agents, video models, longer-context inference and autonomous workflows. Efficiency per token is improving, yet total compute consumption is rising faster. The investment implications remain constructive for $NVDA, HBM, networking, optical connectivity, cooling and power infrastructure. Rising rental prices indicate that the marginal value of energized compute remains high, which encourages operators to keep ordering equipment and accelerating data-centre deployment. However, investors should not assume that high rental prices automatically produce high returns for every neocloud. Realized economics still depend on utilization, financing costs, electricity, networking quality, customer concentration and hardware depreciation. A provider renting B200s at $5.60 per hour can still destroy equity if the machines sit idle, were financed expensively or require rapid replacement. The chart covers non-hyperscaler cloud providers, where spot pricing can be more volatile and scarcity premiums greater than in long-term enterprise contracts. It should therefore not be interpreted as the average rate across the entire market. But the direction is difficult to dismiss. The AI overcapacity thesis requires supply to begin outrunning demand. For now, the opposite appears to be happening. More GPUs are entering the market, yet the price of renting them continues to rise. The industry is not struggling to find workloads for its compute. It is still struggling to find enough compute for its workloads.