Goldman Sachs: AI Investment Is Shifting from Hype to Data Centers and Infrastructure

Goldman Sachs believes the AI investment market is moving out of its phase of general frenzy and becoming noticeably more selective. The focus is shifting from high-profile AI applications to data centers, power supply, networks and other infrastructure without which models simply do not run.

Where capital is moving

According to the bank, investors increasingly look not at companies that merely added AI to a slide deck, but at players that control the physical foundation of the industry. These are the owners and operators of large data centers and the suppliers of compute capacity, chips and networking equipment. The logic is simple: whichever interface or AI service wins the market, training models and running inference still require hardware, communication links and a stable power supply.

During the first wave of generative AI, the market boosted the valuations of many companies for the mere mention of AI. That effect is now fading. Goldman Sachs describes the shift as a "flight to quality": money is going where there are infrastructure assets, clear revenue and a chance to earn from the long cycle of AI adoption, not just from hype around yet another app. For investors, this is a move from betting on promises to betting on the foundation of the AI economy that already works.

"Flight to quality" is how

Goldman Sachs describes the new stage of the AI market.

Why data centers matter more

Goldman Sachs Research expects that within the next two years, AI workloads may take up about 30% of total data center capacity. The reason lies in the nature of AI tasks themselves. Training large models requires thousands of chips running in parallel for weeks without stopping. Inference is not free either: once a service is launched for users or businesses, it needs constant compute capacity, not short peaks like some classic cloud workloads.

Load is growing through several channels at once:

• new GPU and CPU clusters for model training
• permanent capacity for inference in products and enterprise systems
• expanded network infrastructure between clusters, storage and the cloud
• new sites with cooling, backup power and physical security

This is why the largest cloud providers and AI developers are putting tens of billions of dollars into new data centers and computing hardware. For the market, this means a simple shift: the most durable layer is once again at the bottom of the stack. Applications can still appear and vanish quickly, but demand for compute, networks and hosting capacity remains a basic need in almost every scenario for AI's development. By this logic, both data center operators and makers of key components win.

The main limits to growth

The next stage of the AI race runs up against not only models but also utility reality. Goldman Sachs Research estimates that global data center electricity demand could grow by about 175% by 2030 compared with 2023 levels, with AI workloads as the main driver. The material compares this to adding another country from the top ten electricity consumers to the global power system. For governments, grid companies and the providers themselves, this is no longer theory but an investment task.

This already affects where new facilities are built. Large AI clusters need not only server halls but also stable energy sources, powerful communication lines, cooling systems and enough land. That is why some new sites are being considered in remote regions, where access to electricity and territory is easier. The researchers also note that energy and water consumption depend not only on chips but on geography and cooling architecture. In other words, the question of efficiency is increasingly a question of location.

There are also more prosaic barriers. Building large data centers takes years and depends on a long supply chain: securing land, connecting to power grids, agreeing long-term energy supply, buying electrical equipment and waiting for network infrastructure to expand. Component shortages and delays in grid upgrades slow projects down, so investors especially value companies that already have a large data center network today. Ready infrastructure becomes not just an asset but a high-value competitive advantage.

What this means

For the AI market, this is a sign of maturing. The era when it was enough to add the acronym AI to a pitch is gradually ending. The next round of competition will be won not only by whoever has the better model or interface, but by whoever can deliver compute, energy and reliable service at scale. Put simply, the future of AI depends more and more not only on algorithms but on concrete, cables and megawatts.