Comparisons of US AI infrastructure spending to historic railroad or fiber build-outs miss a key structural difference. Dylan Patel explains that while historical infrastructure served domestic users, American AI data centers run inference for Europe, Asia, and global users. Consequently, US CapEx represents an outsized share of GDP—closer to 5% or 6% than historic 2% to 3% rates—because America finances physical compute for the entire world.
Hyperscalers and neoclouds cannot sustain triple-digit CapEx growth indefinitely. Jordan Nanos explains that organizations are hitting physical and organizational boundaries in mobilizing resources for gigawatt-scale sites, industrial equipment, chips, and cooling systems. Finding trillions of dollars in capital becomes increasingly difficult, meaning spending growth must inevitably decelerate even though underlying customer demand remains robust.
Google Cloud's commercial priorities have reduced internal compute allocations to DeepMind relative to frontier rivals. Dylan Patel notes that the share of TPU compute going to DeepMind has steadily declined, leaving DeepMind with less total compute than OpenAI and Anthropic by the end of the year. Google has opted to sell chips and rent compute externally rather than exclusively feeding proprietary frontier model development.
While critics view AI models as low-margin commodity businesses, Dylan Patel notes that Anthropic generates gross margins exceeding 75% on frontier model inference, rivaling mature software giants like Salesforce. Unlike SaaS companies burdened by customer acquisition costs reaching 30%, frontier AI labs face virtually zero customer acquisition costs. High workload stickiness and slower R&D growth relative to revenue have already pushed leading labs toward operating profitability.
Long-term capacity deals can backfire for GPU clouds during periods of rising compute prices. Dylan Patel points out that CoreWeave and Oracle signed massive AI contracts at fixed margins, preventing them from capturing price upside as GPU costs increased. In contrast, Nebius maintained shorter average contract durations, allowing them to raise rental rates and capitalize on intense market demand far more aggressively.
Nvidia supports an estimated $588 billion in off-balance sheet financial backstops, including revenue floors and lease guarantees, to assist neoclouds in securing investment-grade financing. Jordan Nanos emphasizes this strategy is pragmatic: Nvidia avoids total dependency on a handful of hyperscalers while nurturing a diversified buyer ecosystem across the entire hardware supply chain, without using meaningful internal GPU compute for self-dealing.
Severe security vulnerabilities in neoclouds leave multi-tenant GPU clusters open to lateral movement. Dylan Patel and Jordan Nanos report that open-source models can easily breach multiple neoclouds; SemiAnalysis itself accidentally viewed storage and Slurm workloads belonging to intelligence agencies of a top-10 GDP nation. Despite sovereign AI ambitions, single-facility hardware constraints force sensitive government workloads onto insecure, shared infrastructure.
Leading AI labs deliberately delay releasing cutting-edge models for months after training completes. Dylan Patel and Jordan Nanos explain that Anthropic and OpenAI keep models internal to build end-user applications and prevent competitors, particularly open-source Chinese labs, from immediately distilling their capabilities. Public releases represent minimum viable versions needed to maintain a market lead rather than their true technical frontier.