2025: The Year AI Infrastructure Grew Up

Twelve months ago, the AI infrastructure thesis was well-established among specialist investors but still regarded with scepticism by generalist allocators. By December 2025, that scepticism has largely dissolved. Data centre M&A hit a record US$61 billion globally through November. Hyperscaler capital expenditure commitments crossed US$700 billion in aggregate lease obligations. Nuclear power agreements — once the preserve of science fiction — became a standard line item in data centre development plans. The year, in short, was one of institutionalisation.

The defining event of 2025 was Project Stargate — the US$500 billion joint venture announced in January between OpenAI, SoftBank, and Oracle, with the explicit backing of the incoming US administration. Whatever one thinks of the timeline or the financing structure, Stargate did something important: it made AI infrastructure a matter of national industrial policy. That shift in framing — from technology investment to strategic infrastructure — opened the door for a new class of capital, including sovereign wealth funds, pension funds, and government-backed development vehicles, to participate at scale.

The DeepSeek moment in late January provided a useful counterpoint. The release of a highly capable open-source model trained at a fraction of the cost of frontier US models briefly rattled markets and prompted a genuine debate about whether the capex supercycle was justified. The answer, as the year unfolded, was nuanced: efficiency gains in model training do not reduce infrastructure demand — they expand it, by lowering the cost of inference and broadening the universe of applications that become economically viable. The data centres being built today are not for training GPT-5; they are for running ten thousand derivative applications at scale.

For private investors, the lesson of 2025 is that the infrastructure layer of AI is now a mature, institutionally recognised asset class. The window for early-mover advantage has not closed — but it has narrowed considerably. In 2026, the differentiation will come not from identifying the theme, but from identifying the right assets, the right geographies, and the right operators within it. That is precisely the work we do at Lensbridge.