The $25 Billion AI Backlog Nobody's Talking About

30-second take

Andrew Feldman (Cerebrus CEO) argues the AI infrastructure buildout has created a $25B+ backlog with multi-year memory shortages ahead, driven by demand that outstrips data center construction capacity. His core thesis: NVIDIA has deliberately over-allocated chips to "neoclouds" (smaller cloud providers like CoreWeave) to create competitive pressure on hyperscalers (AWS, Azure, Google), a strategy Feldman calls "probably not healthy." The conversation ties chip economics, geopolitics (US-China), and energy constraints to the AI boom, framed around Cerebrus's record $5.5B semiconductor IPO.

Key takes

• $25B backlog claim: Data centers cannot be built fast enough to absorb current AI chip demand, creating a sustained supply-demand imbalance lasting "at least the next several years" on memory specifically.
• NVIDIA's neo-cloud strategy: Feldman asserts NVIDIA intentionally funded and over-allocated chips to newer cloud providers (CoreWeave, Lambda, etc.) to force traditional hyperscalers into competitive chip-buying behavior, creating unhealthy dependencies.
• No speed ceiling for hard problems: The chip industry's historical trend is radical cost-per-compute reduction, but for frontier AI workloads there's "no upper bound" on desired performance—demand will absorb whatever capacity exists.
• Cerebrus IPO as validation: The $185→$311 price jump and record semiconductor IPO ($5.5B raised) signal market belief that alternatives to NVIDIA's architecture have commercial viability at scale.

Useful details

• Cerebrus went public last week at $185/share, jumped to $311—largest semiconductor IPO ever.
• The backlog is explicitly tied to memory shortages, not just chips generally.
• "Neoclouds" = newer GPU cloud providers that NVIDIA allegedly bankrolled to compete with AWS/Azure/GCP.
• Energy and data center construction are framed as the real bottlenecks, not chip design.

Caveats / counterpoints

• Feldman is Cerebrus CEO—his characterization of NVIDIA's strategy is competitive framing, not neutral analysis.
• No specifics on which memory types are bottlenecked (HBM? DDR5?) or how the $25B figure was calculated.
• The transcript doesn't include counterarguments to the "unhealthy dependence" claim or NVIDIA's perspective.
• Doesn't explain how Cerebrus's architecture solves the memory/energy problem or whether their IPO success translates to technical/market dominance.

Ken relevance

High relevance for Ken's AI ops and investing lens. The memory shortage thesis affects agent system deployment costs and timelines—if inference/training capacity stays constrained for years, building around efficiency (smaller models, edge inference, better caching) becomes critical. NVIDIA's neo-cloud strategy exposes a structural risk in GPU supply chains for startups reliant on CoreWeave-type providers. The "no upper bound on speed" claim justifies continued investment in AI infra but also signals price competition won't ease soon. For GTM/content: this framing of NVIDIA as creating competitive cloud dynamics is a spicy take Ken could explore.

Watch verdict

Watch fully. Feldman offers a CEO-level view of chip supply dynamics with specific strategic claims (NVIDIA's neo-cloud play) that aren't widely discussed. The $25B backlog figure and multi-year memory shortage timeline are decision-relevant for anyone building or investing in AI infrastructure.