The Number: $2.8 trillion in aggregate market cap shifts toward AI infrastructure companies since January 2026. The top 5 chip manufacturers and cloud providers have added $420B in combined value.

The Lead: GPU allocation wars are reaching a tipping point. OpenAI, Anthropic, Google DeepMind, Mistral, and Alibaba are collectively spending >$8B per month on inference and training hardware. Supply remains constrained through Q3 2026 despite record fab expansions.

What's Happening:

The Math: A single frontier model training run (GPT-5 scale) requires ~$800M in hardware + $200M in energy costs. Amortized over 2 years, that's ~$500M annually per lab. At 5+ active labs per company, major players are burning $2.5B-5B/year just on training infrastructure.

Market Signals:

What This Means: The GPU shortage is structural, not cyclical. Until fabs 2-3 years out deliver, whoever controls compute controls model deployment speed. Expect M&A around compute infrastructure this summer.

The Bottom Line: If you're building AI applications, secure compute commitments now. Spot market prices will get worse before they improve. Infrastructure-adjacent businesses (cooling, power delivery, data center real estate) are the 2026 winners.