From Hype Cycle to Infrastructure Cycle
The first wave of generative AI investment was characterized by promiscuous enthusiasm — companies saw valuations rise simply by including AI in a press release. Goldman Sachs Research argues that phase is ending, replaced by something more selective and grounded in physical realities.
The firm's analysts describe a flight-to-quality dynamic in which sophisticated investors are shifting attention toward the foundational layer of the AI economy: data centers and the computing hardware that fills them. The reasoning is straightforward. Model capabilities will continue to improve, applications will come and go, but the physical infrastructure required to train and serve those models is an absolute necessity whose supply is constrained by energy availability and construction timelines that cannot be compressed by software breakthroughs alone.
The Numbers Behind the Shift
Goldman Sachs estimates that AI workloads will account for approximately 30 percent of total data center capacity within two years, based on announced hyperscaler capital expenditure plans representing hundreds of billions of dollars in new construction.
The energy dimension is even more striking. The firm estimates global data center power demand could rise roughly 175 percent by 2030 compared to 2023 levels, driven predominantly by the energy intensity of AI training and inference. That increase alone would be roughly equivalent to adding the electricity consumption of an entire top-ten global economy to the grid. This is not a background consideration for AI strategy — it is a primary constraint already shaping where and how fast development can proceed.
Infrastructure Constraints Reshaping Strategy
Building a large-scale AI data center is not simply a matter of capital. Land must be acquired and zoned near reliable energy. Grid connections must be negotiated with utilities that may require multi-year lead times to expand transmission capacity. Large power transformers have become a genuine bottleneck; lead times have extended to two or more years in some markets, constrained by limited manufacturing capacity and competing demand from the renewable energy buildout.
Site selection has consequently become a strategic function at major AI companies. Remote locations with access to hydroelectric or geothermal power, cooler ambient temperatures, and existing high-capacity fiber are now genuinely scarce assets. The geographic concentration of AI compute reflects the clustering of favorable infrastructure conditions.
The Investment Implication
For investors, Goldman Sachs's analysis points toward a pattern from previous computing cycles. During the internet build-out, companies that owned physical cables and data centers captured stable revenue while application-layer companies experienced volatile cycles. A similar dynamic may be forming in AI.
Data center operators, utility companies serving AI campuses, cooling technology specialists, and networking hardware manufacturers sit closer to the infrastructure base than most AI software companies. The firm notes that hyperscale cloud providers, despite their enormous market capitalizations, are primarily infrastructure businesses when analyzed by where their capital is actually deployed.
The Energy Wildcard
Goldman identifies energy as the variable most likely to become binding before compute capacity. Existing power infrastructure was not built to accommodate the growth rates projected for AI. Utilities are investing in grid expansion, but regulatory approval and construction timelines mean new generation and transmission capacity lags demand by years.
This is already driving AI companies to explore unconventional solutions: nuclear power purchase agreements, dedicated natural gas generation, and long-duration battery storage collocated with data centers. The energy question is no longer peripheral to AI strategy — it may well be the determining factor in which companies can scale and which cannot.
This article is based on reporting by AI News. Read the original article.
