GPU Shortage to Persist Until 2028 as Token Demand Drives $2 Trillion Data Center Build-Out
Key Takeaways
- ▸GPU prices expected to remain elevated through 2027-2028 before reverting toward cash operating costs when supply-demand balance is achieved
- ▸Token demand growing 6-10x per year, sustained by continuous model improvements via reinforcement learning unlocking new AI use cases
- ▸Data center financing increasingly relies on capacity deals with hyperscalers to secure debt lending; lower-certainty revenue structures face higher costs and greater risk
Summary
A comprehensive analysis of GPU supply and demand dynamics reveals that graphics processing units will remain in critically tight supply until approximately 2028, with prices holding steady until demand growth converges with supply growth. Token demand is currently expanding at 6-10x annually, driven by advances in model capabilities through reinforcement learning and expanding use cases, significantly outpacing GPU supply expansion. This mismatch is catalyzing an estimated $2 trillion in global data center capital expenditure by 2028, increasingly financed through debt structures backed by revenue commitments from hyperscalers like Meta and Google. The analysis reveals complex financing arrangements where data center operators like CoreWeave structure capacity deals to secure lending, but face substantial residual value risks on GPU assets as technology evolves and potential oversupply materializes after 2028.
- $2 trillion estimated global data center capex through 2028, with mounting reliance on debt financing backed by assets (GPUs) with declining residual value
- Potential for oversupply and price compression when supply catches up, creating financial stress for over-leveraged data center operators
Editorial Opinion
The GPU supply crunch represents the single most critical bottleneck constraining AI infrastructure deployment today. While this analysis provides a realistic timeline for supply normalization, the genuine risk lies in the financial engineering layered atop hardware scarcity—banks and investors are lending heavily against revenue contracts and GPU assets whose residual value could evaporate if model efficiency improvements, new GPU competitors, or demand softening arrive sooner than expected. The 2028 convergence point is not a certainty but a directional estimate; strategists should monitor model improvement velocity and watch for signs of supply elasticity.
