Goldman Sachs says heavy assets could thrive in AI era

The AI boom may be rewarding the wrong layer of the stack

The AI trade may be less about the next app and more about the steel, copper, and concrete underneath it. Goldman Sachs Research is using its HALO framework to push clients past the lazy assumption that the biggest winners in the AI era will be software names alone, and toward the physical assets that keep the buildout running.

HALO stands for Heavy Assets, Low Obsolescence. Goldman’s point is simple but useful: in a market where technology changes fast, the businesses that own long-lived productive assets may be less vulnerable to getting overwritten by the next code cycle. That means investors need to look harder at power, grid equipment, data-center infrastructure, industrial machinery, transport, and materials, not just the digital layer that sits on top.

Why low obsolescence suddenly matters more

The HALO report argues that a long stretch of underinvestment, especially in Europe, has already shifted the market’s center of gravity back toward tangible assets. Goldman ties that shift to higher real yields, geopolitical fragmentation, and supply-chain rewiring, all of which make it harder to ignore capacity, ownership, and physical bottlenecks.

That matters because “AI winners” is too narrow a frame if the real constraint is not model quality but the availability of electricity, land, hardware, and industrial capacity. A business can be less flashy and still be more durable if its assets continue to earn through multiple technology cycles. That is the logic behind low obsolescence, and it is the key reason HALO is more than a catchy acronym.

For Goldman employees, the message cuts across the franchise. Equity research has to think beyond semis and software. Coverage bankers need to see where clients will need financing for real assets, not just digital strategy. And corporate advisory teams that talk about capital allocation now have a cleaner way to explain why the physical side of the AI economy could be just as investable as the model layer.

The bottleneck is power, not just code

Goldman’s own research on data centers makes the bottleneck visible. In one report, the firm said global power demand from data centers is forecast to rise 165% by 2030 versus 2023 levels. In that baseline, AI is projected to account for 28% of the data-center market, which is a reminder that the AI story quickly becomes an electricity story.

Goldman has also argued for more than a year that the load growth is big enough to force major new investment. Its 2024 research said data-center power demand was poised to more than double by 2030 and estimated that utilities would need about $50 billion in new power-generation capacity to support that growth. In other words, the AI buildout is not just creating demand for chips and cloud services; it is creating demand for generation, transmission, and everything that connects the two.

A later Goldman framework sharpened the point with six constraints, the so-called 6 Ps: pervasiveness of AI, productivity of servers and compute, prices of electricity, policy, parts availability, and people availability. Goldman said that setup could push data-center power demand up 175% by 2030 versus 2023 levels. That is the kind of number that changes how investors think about utilities, grid operators, and the industrial companies that sell the equipment behind the wall.

The AI story has already been moving in this direction

HALO is not a one-off theme. It fits into a broader Goldman research arc that has been building for years around power demand, industrial capacity, and the physical consequences of AI. Goldman’s 2025 data-center work said the explosion in generative AI had created an arms race for high-density data centers and much more electricity, while also flagging concern about returns on AI investment after DeepSeek.

That matters because it shifts the debate from “how big can AI get?” to “what has to be built for it to work?” If the returns on software spend are uncertain, the businesses that supply power and infrastructure may look more attractive on a risk-adjusted basis than the headlines suggest. This is exactly where heavy, harder-to-replace assets can start to look like the real bottleneck, and the real pricing power, in the AI economy.

For Goldman’s clients, that can change the way sectors get framed in pitch books and strategy sessions. A utilities, industrials, or logistics conversation is no longer just defensive. It can become a growth discussion tied directly to AI capex, data-center expansion, and the need to keep physical systems ahead of digital demand.

Europe is part of the reason this matters now

Goldman’s HALO argument also leans on Europe’s weaker backdrop. In January 2025, Goldman forecast euro-area growth of 0.8% for the year, below Bloomberg consensus of 1%, citing manufacturing headwinds, fiscal drag, and trade tensions. Its 2026 euro-area outlook says Europe still faces structural headwinds and only a small improvement in growth.

That slower-growth environment helps explain why low-obsolescence assets may be getting more attention. When demand is not broad-based and growth is uneven, the market tends to reward businesses that own scarce, durable capacity rather than those that depend on a clean macro backdrop. Add in years of underinvestment and you get a setting where the old economy can look newly relevant, especially when it is wired into the AI supply chain.

The practical implication for readers inside Goldman is that the AI conversation has widened. The winners may still include software and semiconductors, but HALO says the harder-to-obsolete layer underneath them may be where the real bottlenecks, and some of the best opportunities, sit. For a bank that lives on reading capital flows before they become consensus, that is not a side note. It is the trade.