Artificial intelligence investment is entering a phase in which algorithms are no longer enough. Goldman Sachs sees attention shifting toward the physical infrastructure required to train models and serve them continuously.
The change is described as a “flight to quality.” Investors are scrutinising companies that own data centres, computing capacity and the networks connecting them, while narrow or experimental AI software attracts less automatic enthusiasm.
This is not simply a rotation between technology stocks. It reflects a hard constraint: every successful AI application eventually needs chips, electricity, cooling, land and fibre.
AI workloads redraw the data centre
Goldman Sachs Research estimates that AI could account for about 30% of total data centre capacity within two years. Training a large model can require thousands of accelerators working together for long periods. Inference also consumes steady capacity once a service reaches regular users.
Hyperscale cloud providers are therefore spending tens of billions of dollars each year on facilities and hardware. High-performance networking expands with the compute because chips running in parallel must exchange data quickly enough to function as one system.
Unlike an application that can be released rapidly, a large data centre may take years to plan and build. Developers need land, grid connections, electrical equipment, fibre access and long-term energy agreements. Delays in any one layer can slow the entire project.
That favours operators that already control suitable sites and large networks. Existing capacity becomes strategically valuable when demand rises faster than new infrastructure can be connected.
Electricity becomes part of AI strategy
Goldman Sachs estimates global data centre power demand could grow around 175% by 2030 compared with 2023, driven largely by AI. The additional consumption would be comparable to adding another top-ten electricity-consuming country to the world’s grids.
Power availability now influences where AI clusters are built. Some facilities move toward remote areas where land and stable electricity are easier to secure. High-capacity fibre remains essential, so location is a balance between energy, connectivity and access.
Cooling adds another geographic variable. Climate, cooling design and local water conditions can affect a facility’s environmental footprint as much as chip efficiency. Choosing a site therefore shapes operating cost, resilience and resource use over many years.
Utilities and governments become participants in AI expansion because grid upgrades, generation and transmission determine which projects can proceed. A company’s model roadmap may look like a software strategy, but its practical ceiling can be a transformer, substation or delayed connection.
Investors move closer to the foundation
During the first generative AI wave, companies could attract attention by associating a product with AI. The market is becoming more selective about revenue, infrastructure access and the ability to support long-term deployment.
Data centre operators and chipmakers sit near the base of the ecosystem. Their capacity is needed regardless of which individual applications succeed. Previous computing cycles often rewarded infrastructure providers with steadier demand while software winners changed more rapidly.
That does not make every infrastructure investment safe. Facilities require enormous capital, forecasts can overshoot and energy constraints can leave projects waiting. Environmental scrutiny may also raise costs or restrict development.
Goldman’s analysis nevertheless captures a fundamental transition. AI is moving from an era dominated by model announcements into one constrained by industrial execution. Compute has to be installed, powered, cooled and connected before software can turn it into revenue.
For technology leaders, infrastructure access should now sit beside model selection in planning. For investors, the question becomes not only which company has the most compelling AI story, but which one controls a scarce layer that many stories depend upon.
The next phase of the AI economy may be decided as much in planning offices, power markets and construction sites as in research laboratories. Intelligence remains digital; the bottlenecks beneath it are unmistakably physical.
