All GPUs Are Running, but the Bills Don’t Add Up
The view reshared by Harry Stebbings says data centers have no idle GPUs and capex is entirely justified. The replies pull the question back to another spreadsheet: utilization is not profit, and who pays for the next gigawatt?
Harry Stebbings reshared a take on X about AI capex. The core claim is blunt: this buildout is entirely justified and the market won’t crash; GPUs powered on in data centers are fully utilized, with no dark GPUs; if there is enough compute, inference still has orders of magnitude of thinking left to do.
Let’s define a few terms first. Capex means capital expenditure: building data centers, buying GPUs, and laying power infrastructure all count. Dark GPU refers to cards that were bought but sit idle. Dark fibre refers to the old overbuild of fiber around 2000: much of it was laid and never used, companies went bankrupt, and the assets changed hands. The main post tries to cut off this analogy: GPUs aren’t dark, the machines are running, so this cycle is different.
The replies don’t stop at “are the machines running?” Straggler Liu splits the question into two spreadsheets: demand for compute, and the return on building compute. Internal workloads can fill a cluster, but being full doesn’t mean outside customers are willing to pay. More thinking at inference time may be very valuable, but the question is whether the incremental value covers the incremental compute cost. For an infrastructure owner, the question is whether this generation of hardware can earn back its cost before new chips push rents down. Full utilization matters, and so does the price at which that fully utilized capacity is sold.
One commenter compressed it further: Busy and profitable aren't the same GPU. A GPU being busy and a GPU making money are two different things. A restaurant can be full and still have every table using a discount voucher.
Dominick Pandolfo and kevindurantsburner press the same point: who pays for the next gigawatt? The question isn’t new, but it’s hard to avoid. The demand curve and willingness to pay are not the same line. Someone needing compute does not mean someone is willing to buy compute at a price that covers the cost of capital.
Rolf points to another kind of idle compute: GPUs in gamers’ hands. SETI@home once mobilized idle computer cycles; why couldn’t something similar be done for GPUs? The idea is vivid, and it reminds us of something: compute supply doesn’t exist only in data centers. Edge devices have always been there.
Roger Smith uses solar as a reference point. AI may be useful and still be a bubble. Solar had a bubble in the late 2000s; many companies went bankrupt or lost 90% of their market value. Twenty years later, global solar panel production is up 20x, yet no trillion-dollar solar company emerged. The analogy isn’t a verdict; it shows only one thing: technology diffusion and capital returns can diverge. The production curve can go up while the shareholder return curve does not follow.
Marcus Castro is more optimistic. He thinks many people are still thinking in SaaS-era frameworks, and that demand for compute, inference, data, and research will approach infinity. He could be right. One caveat is needed: demand can approach infinity while willingness to pay remains finite. Free demand is still demand, but the bill doesn’t care.
Shawn Wen adds a concrete scenario: in voice, the constraint is per-user speed, not fleet size. Someone on a call won’t wait, and the thinking available to each turn of conversation is limited. That detail matters. Inference compute isn’t released evenly, and product form factors will cut off some of the theoretical demand.
Roger Smith’s second reply puts the pressure on phones: what stops super AI from being free on a phone? A 2026 phone AI model beats the frontier models running in data centers in 2022—only a four-year gap. Apple, Samsung, and Arm could take 90% of the market. The claim is debatable, but it points to the cost curve. Centralized buildout being real does not mean returns accrue only at the centralized-buildout end.
Put the original post and the replies together, and the debate is actually clear. The original post offers a testable fact: powered-on GPUs are highly utilized. The replies raise a harder-to-test question: have those GPUs earned back their cost of capital over their lifetime? Dark GPUs are easy to count; dark capex is hard to audit. The machines are running, and so are the bills. Where the money for the next gigawatt comes from is closer to the answer than whether there are empty cards in the racks.
发布时间: 2026-11-16 18:26