Investors pivot to AI infrastructure

AI investing is getting more physical. That’s the real shift. For the last two years, the market treated AI like a software race — better models in, bigger valuations out. But the bottleneck turned out to be concrete stuff: power, substations, cooling loops, advanced packaging, and the software used to design the chips themselves. That is why money is moving down the stack. ### Why are investors talking about infrastructure now? Because the easy story — “the best model wins” — stopped being enough. The large model companies still matter, but running frontier AI now depends on access to GPU clusters, power-dense data centers, and the supply chain around them. Big Tech is reinforcing that point with spending, not slogans: Microsoft, Alphabet, Amazon and Meta together plan to spend more than $300 billion in 2025, mostly tied to AI infrastructure. (ig.ft.com) ### What changed in the market’s mental model? The market started to see AI less like an app launch and more like an industrial buildout. A model can improve in months. A substation, transmission upgrade, or new high-density data hall takes much longer. That changes who has leverage. If compute is scarce, then the owners of power, land, cooling systems, networking gear, and fab capacity can capture a lot of the upside t(ig.ft.com 1)(ig.ft.com 2) ### Why is power suddenly such a big deal? Because AI racks are power-hungry in a way ordinary cloud servers were not. One estimate puts new U.S. data-center demand at 44GW by 2028, with only about 25GW of power likely available on current timelines. That leaves a 19GW gap. Basically, the constraint is no longer just chips. It is whether you can actually energize the building that holds them. (ig.ft.com)so much? Higher rack density means more heat in less space. Air cooling starts to break down as a simple answer, so liquid cooling moves from niche to necessity. You can see capital following that pressure point. Carrier Ventures expanded its investment in ZutaCore on April 29 to scale direct-to-chip, waterless liquid cooling for high-density AI data centers. That is a pretty clean signal that t(ig.ft.com) infrastructure, not an accessory. (markets.ft.com) ### Where does Cadence fit into this? Cadence sits in a less obvious but very valuable layer — electronic design automation, or EDA. These are the tools chip companies use to design, verify, and tape out advanced silicon. If the AI race requires more custom chips and faster iteration at leading-edge nodes, EDA vendors become toll collectors. Cadence’s latest (markets.ft.com)s that use AI to improve power, performance and area faster. (cadence.com) ### Why are “agent” tools part of infrastructure? Because once AI agents start doing real work, they need payments, permissions, data access, monitoring and orchestration. That is not model magic. It is plumbing. Startups are getting funded to build exactly that layer — like Sapiom’s $15 million seed round for soft(cadence.com)techcrunch.com) ### Does this mean model companies matter less? Not less — just less alone. The best model still helps, but it no longer guarantees the best business. If rivals can rent similar intelligence, then durable advantage comes from controlling the scarce complements around it. Think of the model as the engine and infrastructure as the fuel, transmission and roa(techcrunch.com)ing to price AI like a capacity business. That means the next winners may be the companies that can secure megawatts, cool dense racks, tape out advanced chips, and turn agents into reliable systems — not just the ones with the flashiest chatbot.