GPT-4 and the Infrastructure Reckoning

On 14 March 2023, OpenAI released GPT-4 — a multimodal language model that could process both text and images, passed the bar exam in the 90th percentile, and demonstrated reasoning capabilities that its predecessor had not come close to matching. The release came four months after ChatGPT had introduced the world to conversational AI, and it confirmed what many in the technology industry had suspected: the pace of AI capability improvement was not slowing down. If anything, it was accelerating.

For investors in digital infrastructure, GPT-4 was a clarifying moment. The question of whether AI would drive significant demand for compute infrastructure — a question that had been debated in specialist circles for several years — was effectively settled. The question that replaced it was more interesting and more actionable: how much infrastructure would be required, where would it need to be located, and who would build and own it? The answers to these questions would determine the investment opportunity for the decade ahead.

The scale of the infrastructure requirement was becoming clearer by the week. Training GPT-4 reportedly required tens of thousands of Nvidia A100 GPUs running for months — a compute investment that dwarfed anything that had come before in the commercial AI space. Serving GPT-4 to millions of users required a different but equally substantial infrastructure: inference clusters distributed across multiple geographies, connected by high-bandwidth networks, cooled by industrial-scale systems, and powered by electricity at a rate that was beginning to concern utility operators. The infrastructure reckoning had arrived.

At Lensbridge, we had been building our digital infrastructure thesis for several years before GPT-4 — our investment in Cisco's digital infrastructure partnership, our early positions in Asian data centre markets, and our focus on GP-led secondaries in the technology sector had all been expressions of our conviction that the infrastructure layer of the digital economy was a compelling investment opportunity. GPT-4 validated that conviction in the most public way possible. The work now was to translate that conviction into specific investments at the right prices in the right markets — and to do so before the rest of the institutional investment world arrived at the same conclusion.