The post under review is short and confident: a user posting as Joel wrote on X on October 28, 2026 that GPU prices "will crash next year" because "half of data center buildouts are already cancelled or stalled" and because neoclouds are "stuck in 5 year deals with no confirmed demand." It is a forecast about 2027 built on two factual assertions about the present, and those assertions are what can actually be checked. Neither survives contact with the available reporting, and the current direction of GPU rental prices runs against the prediction.
Start with prices, the thing the post says will collapse. A GPU price index compiled from 19 providers and published on October 7, 2026 recorded a median on-demand H100 SXM rate of $4.20 per GPU-hour, ranging from $3.20 to $12.29. More telling than the level is the direction. The same index documents that the SemiAnalysis H100 one-year contract index rose 38% between October 2025 and March 2026, climbing from $1.70 to $2.35 per GPU-hour, and that Nebius raised its on-demand H100 price from $3.85 to $4.50 on October 1, 2026 — increases of 17% to 21%. That follows a genuine decline in 2025, when AWS cut on-demand H100 pricing by about 44% in June. The market has been volatile, and secondary indexes built from executed trades show lower numbers — around $2.81 to $2.86 per GPU-hour on October 7 — but the documented trend through 2026 is upward, not a collapse.
The data center premise is the most distorted part. The "half of buildouts are cancelled" formulation circulates widely, and SemiAnalysis examined it directly in June 2026, tracing it to an April 2026 Bloomberg report about dependence on Chinese electrical equipment and to sharper downstream versions that hardened a supply-chain warning into a claim that half of US data centers were cancelled. The firm found that its end-2026 North American hyperscaler self-build forecast had moved only about 1% in six months, with colocation under 5%, and argued the headline number rested on a denominator that undercounted capacity under construction by multiples while counting speculative, pre-construction announcements as though they were 2026 projects. Delays do exist — the firm itself flagged the Nebius New Jersey campus and a STACK Infrastructure site pushed to 2029 — but scattered delays are not half of a national pipeline.
There is a real and serious story the post gestures at. Heatmap reported on October 6, 2026 that at least $260 billion of data center investment had been cancelled in 2026 after sustained local opposition, with about $130 billion of that in the third quarter alone, roughly $1 trillion facing some form of organized pushback, and about half of projects that met backlash ultimately killed. That is significant. But it measures dollar value of announced projects facing community opposition, and the outlet itself said the figures exclude projects stalled by state-level moratoriums in Texas and New York. It does not establish that half of data center buildouts are cancelled or stalled, and it says nothing about GPU rental rates next year.
The final premise — neoclouds holding contracts with no confirmed demand — is the weakest. An analysis of neocloud unit economics published in March 2026 described CoreWeave's model as multi-year take-or-pay contracts, typically two to five years, in which customers pay whether or not they use the capacity, and put its contracted backlog at $66.8 billion at the end of 2025, roughly 13 times annual revenue. Industry reporting on the sector describes surging demand, with Synergy Research Group sizing the neocloud market at $9 billion in the fourth quarter of 2025, up 223% year over year, and a Nebius executive saying the company has four to five customers lined up for every GPU cluster it brings online. That undated reporting should be treated cautiously, but it points the same way as the contracted-backlog figures.
The genuine caveats deserve airing. The neocloud model does carry real risk: CoreWeave paid about $1.2 billion in interest in 2025 on $21.4 billion of debt, relies on depreciation schedules of four to six years, faces hyperscaler competition and needs costly hardware refreshes, and on-demand rates for older chips have swung sharply over three years. Cancellations driven by local opposition are accelerating. A 2027 price decline is therefore not impossible, and this report does not claim otherwise. What the evidence does not support is the post's specific reasoning: the "half cancelled" figure is a disputed and largely debunked framing, and the claim of no confirmed neocloud demand is contradicted by large contracted backlogs — which is why the post, as written, misleads readers about the state of the market rather than accurately describing it.
