Tom Tunguz documented Nvidia Blackwell GPU rental prices rising 48% to $4.08 per hour within two months, CoreWeave implementing 20% price hikes and three-year minimum contracts, and OpenAI CFO Sarah Friar stating the company is making "tough trades" on unpursued projects due to insufficient compute (Tunguz, 2024). Anthropic has restricted its latest model to approximately 40 organizations, illustrating gated access based on strategic relationships and security (Tunguz, 2024). These developments confirm the end of the post-2000s era of effectively unlimited AI supply chains.

Original coverage identified five hallmarks—relationship-based selling, highest-bidder access, variable speed, inflationary pressure, and forced diversification into smaller models or on-premise setups—but missed the primary energy bottleneck now constraining new data center builds. Goldman Sachs projected AI data centers could consume 8% of total U.S. power by 2030, while regions such as Northern Virginia face multi-year delays in grid upgrades as utilities cannot match hyperscaler demand (Goldman Sachs, 2024). Epoch AI's compute trends analysis shows training runs for frontier models have scaled 10,000x since 2010, with energy requirements now outpacing chip efficiency gains and creating physical infrastructure limits measured in gigawatts rather than just GPUs (Epoch AI, 2023).

Patterns from the 2021-2022 global chip shortage, combined with current U.S. export controls on advanced AI silicon to China, indicate this scarcity will concentrate capability among well-capitalized incumbents and sovereign-backed entities while intensifying competition for power purchase agreements, including restarts of nuclear facilities and deals for natural gas. Primary sources show buildout timelines of 3-5 years for new substations and generation capacity, suggesting the compute-constrained regime will persist through at least 2028 and force procurement, margin management, and efficiency optimizations as core disciplines across AI developers.