The GPU Shortage Nobody Saw Coming

In the weeks following the release of GPT-4 in March 2023, a peculiar problem emerged in the AI industry: there were not enough GPUs to go around. Nvidia's H100 — the most powerful AI training chip available — had waitlists stretching to six months or more. Cloud providers were rationing access to GPU instances. AI startups that had raised tens of millions of dollars found themselves unable to train their models because the compute they needed was simply not available. The GPU shortage of 2023 was the first major supply constraint in the AI buildout — and it revealed something important about the structure of the AI infrastructure market.

The shortage was, in retrospect, entirely predictable. The demand for AI compute had been growing exponentially for several years, driven by the scaling laws that showed that larger models trained on more data consistently outperformed smaller ones. Nvidia had been the dominant supplier of AI training chips since the deep learning revolution of the early 2010s, and its H100 — built on TSMC's most advanced process node — had a production capacity that was constrained by the physical limits of semiconductor manufacturing. When ChatGPT demonstrated that AI had crossed a threshold of commercial viability, the demand for H100s went vertical in a way that no supply chain could have anticipated.

The investment implications were significant. Nvidia's stock, which had already been rising on AI enthusiasm, accelerated dramatically — it would go on to become one of the best-performing large-cap stocks in history. But the more interesting opportunities were in the second and third-order effects of the shortage. Data centre operators who had secured GPU allocations gained a competitive advantage that translated directly into pricing power. Cloud providers that could offer GPU access commanded premium rates. And the companies building the networking, cooling, and power infrastructure that GPU clusters require found themselves with order books that stretched years into the future.

At Lensbridge, the GPU shortage of 2023 reinforced our conviction that the infrastructure layer of AI — not the chips themselves, but the facilities, networks, and power systems that house and connect them — is the most durable and accessible part of the AI investment opportunity for private capital. The chips will become more abundant over time; the infrastructure that supports them will remain scarce.