From Hype to Harvest: When Will AI Investments Deliver Returns?

Every major technology investment cycle follows a similar arc. The early phase is characterised by narrative and optionality — valuations reflect what might be, not what is. The middle phase brings capital abundance, competitive intensity, and the first signs of differentiation between winners and also-rans. The late phase is where returns are actually realised — or not. The AI investment cycle of the early 2020s is now entering that third phase, and the questions being asked in LP boardrooms are becoming sharper.

The challenge is that AI infrastructure and AI application companies have very different return profiles, and they are often conflated in portfolio reporting. Infrastructure assets — data centres, power, fibre — generate contracted, yield-like returns that are relatively predictable. The question is not whether they will deliver, but whether entry prices adequately reflect the risk of technology obsolescence and demand concentration. Application-layer companies, by contrast, are subject to the full force of competitive dynamics: the moat that looks wide today may be narrowed by the next model release or the next open-source breakthrough.

For family offices and long-duration investors, the most important discipline right now is vintage awareness. The funds and direct investments made in 2021 at peak valuations are facing a different reality than those made in 2023 and 2024, when the market had corrected and the infrastructure thesis had become more specific. GP-led secondaries are increasingly being used to manage the former — providing liquidity to LPs who need it while allowing patient capital to stay in assets that still have a credible path to value creation.

At Lensbridge, our view is that the harvest phase of AI investing will be uneven and extended. The infrastructure layer will deliver first and most reliably. The application layer will produce a small number of exceptional outcomes and a large number of disappointing ones. The skill, as always, is in the selection — and in having the patience to let the thesis play out on the right timeline.