Compute overhang

Definition

A situation where available compute capacity grows faster than the algorithmic improvements needed to use it, creating a stockpile of unused potential. A sudden algorithmic breakthrough can then unlock rapid capability jumps.

Why it matters

Compute overhang is why capability jumps can appear sudden even though compute growth is gradual. If the hardware to run GPT-5-class models exists today but the algorithms are not ready yet, an algorithmic breakthrough could produce a dramatic capability leap almost overnight. This matters for forecasting and risk assessment: linear extrapolation of AI progress misses the possibility of step-function improvements driven by overhang release. For investors, compute overhang in cloud infrastructure suggests that current GPU investments could yield outsized returns when the next algorithmic breakthrough arrives. For safety teams, overhang is a scenario to plan for, not assume away.

In practice

The current AI boom partially reflects a compute overhang from decades of GPU development for gaming. When the transformer architecture arrived in 2017, the hardware to train it already existed, enabling rapid scaling. NVIDIA's pivot from gaming to AI accelerators unlocked massive latent compute. By 2025, hyperscalers had invested $200B+ in AI infrastructure, creating a potential new overhang: if a more efficient architecture than transformers emerges, these data centers could power capabilities far beyond current models. DeepSeek's R1 model, which achieved frontier performance at a fraction of typical training cost, demonstrated how algorithmic efficiency gains can exploit existing compute more effectively.