ChipsMay 11, 2026via The Information
Capital, Not Compute, is the Real AI Bottleneck
Why it matters
As GPU demand outpaces financing timelines, the AI buildout has shifted from a compute problem to a capital and execution choreography problem. Infrastructure providers are now racing to secure long-term customer contracts (especially hyperscaler deals) to unlock financing for multi-year data center builds.
Key signals
- Single gigawatt capacity estimated at $50B (Nvidia CEO Jensen Huang)
- Data center demand projected to reach 156 GW by 2030 (McKinsey)
- Total infrastructure investment required could approach $7 trillion by 2030
- GPU depreciation lifecycle: ~6 years (vs. cable networks lasting decades)
- Lambda reports 10,000+ customers across public cloud business (roughly 1/3 of revenue)
- Nebius signed multibillion-dollar deal with Meta including demand backstop guarantee
- Older GPUs (V100, A100) continue generating strong returns beyond expected lifespan (CoreWeave)
- Data centers can take years to build while customer demand materializes much faster
- Demand extends beyond hyperscalers to AI startups (Cursor, Harvey) and enterprise adoption
The hook
$7 trillion. That's what AI infrastructure will cost by 2030—and capital, not compute, is now the real constraint.
Satisfying AI’s ravenous appetite for power and compute has kicked off a once-in-a-generation infrastructure expansion that is also one of the most expensive in modern history. Nvidia CEO Jensen Huang has estimated a single gigawatt of capacity can cost as much as $50 billion. With McKinsey projecti…