This article analyses the implications of the massive increase in AI model training driven by surging demand for GPUs.⁵⁶ The GPU and data center overbuilding will likely waste capital in the short term but lower costs long-term, enabling more experimentation. However, startups must shift focus from infrastructure to concrete use cases.

What's going on here?

There is a disconnect between the AI hype fueling infrastructure overbuilding and the actual end-customer value creation needed to justify it.

What does this mean?

Based on a conservative estimate of Nvidia guiding $50B in GPU revenue, $100B will be spent on data centres annually. For investors to earn a return, $200B in value must be created by AI products leveraging these GPUs. But current known revenue from major tech companies utilizing AI is only around $75B, leaving a $125B+ gap.

Why should I care?

The infrastructure will most likely get cheap with time. Startups need to understand the implications of GPU overinvestment versus end-user value creation and find a balance that works for them and investors. One path forward is to identify specific customer pain points and build narrowly focused AI solutions, rather than building general-purpose LLMs.