In an era where financial markets react to headlines about AI infrastructure bottlenecks, a new analysis cuts through the noise with a starkly different conclusion: the widely cited claim that half of America's 2026 datacenter capacity is canceled is not just wrong, it is a statistical artifact. Dylan Patel, writing for SemiAnalysis, dismantles the narrative by contrasting vague, AI-generated forecasts against a proprietary model built on satellite imagery and granular construction tracking. For investors and industry observers who cannot afford to be misled by "vibe-coded" data, this piece offers a necessary correction based on physical reality rather than press release optimism.
The Illusion of the 50% Cancellation
The core of Patel's argument targets the methodology behind the alarming headlines. He traces the origin of the "half canceled" statistic to a Bloomberg report that relied on data from Sightline Climate, which he argues fundamentally misunderstands the pipeline. "We find these claims quite amusing," Patel writes, noting that while other outlets amplified the story, his team's internal forecast for 2026 North American hyperscaler self-build capacity has shifted by only about one percent over the last six months.
This discrepancy, Patel contends, stems from how different models define "capacity." The flawed models treat every public announcement as a guaranteed delivery date. In contrast, Patel's approach filters out speculative projects that lack financing or interconnection studies. He argues that the headline number is built on a "hugely flawed denominator, off by multiples," because it counts early-stage, unproven megaprojects alongside shovel-ready sites.
"Claude Code pulls press releases, views unfounded GW-scale announcements as ground truth, misunderstands construction timelines and grid complexities, and compiles a terribly inaccurate report."
Patel's critique of automated forecasting tools is particularly sharp. He suggests that the rise of AI-generated analysis has led to a market flooded with confident but baseless predictions. While this framing effectively highlights the danger of relying on surface-level data, critics might argue that dismissing all non-proprietary models as "Claude Coded" risks overlooking legitimate concerns about supply chain fragility that independent analysts have raised regarding Chinese electrical components.
The Anatomy of Real Delays
Patel does not deny that delays are happening; rather, he insists they are being mischaracterized. He categorizes real setbacks into three buckets: aggressive announcements by new players, overly optimistic construction timelines, and permitting hurdles. To illustrate the difference between a "canceled" project and one simply delayed, he points to specific case studies where physical evidence contradicts public claims.
Take the Nebius flagship campus in New Jersey, developed by DataOne. The developer initially targeted a four-month delivery for its first phase, a timeline Patel deemed unrealistic from the start. "Satellite imagery shows the shell going up remarkably fast, but setting up the MEP took far longer than the shell itself," he notes, pointing out that while the building structure was visible quickly, critical cooling equipment lagged significantly behind schedule.
Similarly, he examines Core Scientific's Denton campus, where a combination of weather, construction issues, and even an exploding transformer pushed timelines back. However, Patel emphasizes that his model successfully predicted these specific slippages while maintaining that the broader industry pipeline remains robust. "Our datacenter Model flagged the STACK Infrastructure / Oracle permitting impasse months before Bloom Energy stepped in," he writes, highlighting the predictive power of tracking regulatory hurdles rather than just construction milestones.
"The '50% cancelled or delayed' figure isn't really a statement about the US datacenter pipeline, but rather about the slice of the pipeline most prone to slipping."
This distinction is crucial for understanding the current market dynamic. The projects that are failing are often those announced by inexperienced developers with no track record, not the core infrastructure being built by established hyperscalers. As Patel puts it, "When we see a brand-new announcement in 2025 saying 2026 operational from a new 'unknown' developer, it should raise some red flags." This focus on developer experience adds necessary nuance to the conversation about capacity constraints.
The Grid and the Pipeline Reality
While Patel successfully debunks the "cancellation" myth, he acknowledges that structural bottlenecks are real. He points to the STACK Infrastructure/Oracle site, which has been pushed to 2029 not because of a lack of demand, but due to a "lack of gas pipeline and the ensuing burdensome regulation." This aligns with broader findings from related deep dives into PJM Interconnection, where the interconnection queue is clogged with over a terawatt of speculative load requests.
Patel argues that the market is currently suffering from an oversupply of early-stage projects that were never viable for 2026 delivery. "In December 2025, we flagged that over half a Terawatt is in the large load queue in the US," he writes, noting that these speculative announcements create a false sense of urgency and scarcity.
"We do not put low probability datacenters in our datacenter timelines."
This rigorous filtering process explains why his numbers remain stable while others fluctuate wildly. By excluding projects without financing or equipment orders, his model reflects the actual physical reality on the ground rather than the marketing aspirations of developers. However, this approach may underestimate the speed at which new entrants could eventually mobilize if capital conditions improve, a factor that remains uncertain.
Bottom Line
Dylan Patel's analysis provides a vital corrective to the panic-driven narrative surrounding AI infrastructure, grounding the debate in satellite imagery and construction logistics rather than press releases. His strongest contribution is exposing how automated models and media amplification have distorted the perception of risk by counting speculative announcements as guaranteed capacity. The most significant vulnerability in his argument lies in its potential underestimation of how quickly regulatory hurdles could compound across the entire industry, not just for new entrants, potentially creating a broader slowdown than even his conservative model predicts.