The load-bearing word is identical. Same silicon and same generation means the move Baker reports is not a mix effect or a spec upgrade being read as inflation [2]. For most hardware generations, the rental price of a fixed configuration decays as supply lands and the successor part is announced. Here it went the other way.
Annualise it, because that is the form a budget needs. Fifty percent over seven months compounds to 1.5^(12/7), about 100 percent a year. Sixty percent over six months is 1.6 squared, about 156 percent [1]. Both figures rest on one investor's channel checks in GPU rental markets, as relayed by theneuron.ai [2]. For that rate to land on your invoice, you have to be renewing short-term or spot capacity in the same hardware class; a multi-year offtake signed in the spring makes the reprice somebody else's renewal quote.
The per-megawatt numbers explain why the bid keeps climbing. Dylan Patel puts base AI compute at roughly $10M to $15M per megawatt against as much as $50M of revenue per megawatt that Anthropic has generated at times [6]. That is a ratio of 3.3x to 5x [2]. Whoever converts a scarce megawatt into the most revenue can outbid everyone else for the next one, which is the mechanism under the claim that two labs could absorb roughly half the world's new AI compute [7].
Then there is the $11T through 2029 [5]. Divide it by the same $10M to $15M per megawatt and you get 733 GW to 1,100 GW of base compute [3]. Nobody is energising a terawatt this decade, so most of that total has to be power, buildings, memory and financing rather than accelerators. Two numbers in the same argument rarely share a denominator.
Utilization is an occupancy metric. A busy rack tells you the machine is rented, not that the end customer paid a price covering the GPU, the power, the data center, the financing and the training run [3]. Baker's price series and Zitron's objection are measuring different quantities, which is why a shortage does not settle the question: subscriptions, investor capital, cloud financing and debt all sit between the meter and the person who eventually pays [4].
That is also the number no dashboard produces. In my context it argues for pricing renewals at today's level or above rather than putting hardware deflation into a 2026 plan, and for reading the offtake terms in anything funded out of the more than $500B Nvidia has lined up with outside investors, since the question those structures answer is who holds the asset when the clearing price finally gets quoted [8][11].
