The pitch decks that were so popular two years ago resembled each other in the waiting area of a venture fund’s Menlo Park office. A writing tool using generative AI. AI customer support representative. AI helper for coding. AI reviewer of legal documents. The basic architecture remained the same despite the different category labels: a revenue model that assumed sufficient switching costs to support the value, a prompt engineering layer that the team dubbed their defensible IP, and a wrapper developed on top of GPT, Claude, or Gemini. At their height, the partners met with a number of these each week. A few received funding. The majority didn’t. And those who did are learning what those who didn’t already suspected: pricing is the foundation of competition when all competitors are built on the same model with the same API, and competing on price with someone who has more runway is a strategy with a predictable outcome.
Over the past eighteen months, Silicon Valley venture capital has seen a shift that is somewhat a correction from the app layer froth and partially something more significant—a realization that the limitations preventing AI progress are not at all in the application layer. They are tangible. Neither software nor market demand are limiting the businesses that are unable to obtain adequate GPUs. Power and silicon are limiting them. Furthermore, with venture capital, limits are typically where the profits are.
Of the two, the computational infrastructure argument is the more well-established. Since 2023, GPU constraints have dominated the AI supply chain. While smaller businesses had to wait months for access, NVIDIA’s H100 and H200 allocations went to hyperscalers and major enterprise clients. Packaging capacity and enhanced memory manufacture, not chip design, are the fundamental limitations, and they won’t be solved anytime soon.
Infrastructure funds and later-stage venture capital firms have been investing in the data center supply chain, which includes cooling technology companies, power-contracted campus developers, and colocation providers, on the grounds that all AI companies require these resources and that the companies supplying them are not subject to the same competitive dynamics as app-layer companies. The advent of a less expensive alternative does not allow you to leave the data center where your GPU cluster is located.
In certain structural aspects, the energy argument is more recent and intriguing. A single large training cluster can demand 100 megawatts or more of electricity continuously, which is equal to the residential load of a sizable city. A huge AI training run uses electricity at a rate that would have seemed impossible to describe to a power utility ten years ago. AI-driven data center expansion, especially in hyperscale locations like Northern Virginia, Texas, and Arizona, has been hampered by the grid’s incapacity to provide committed power at the timeliness required by data center operators rather than by permits, real estate, or technology. For significant new interconnection requests, utilities have lineups that last five to seven years. In order to create the power, transmission infrastructure, generation capacity, and grid upgrades must be completed, all of which take years.
Venture capital has responded by treating energy as a technology sector issue rather than a utility sector issue. In order to meet its data center requirements, Microsoft reopened a nuclear facility at Three Mile Island. Similar partnerships and investments in nuclear energy companies have been made by Google and Amazon. After spending the majority of the last ten years attempting to secure institutional funding for what appeared to be a long-term, low-probability venture, small modular reactor developers are now closing funding rounds with venture capitalists who had not previously invested in anything requiring a construction crew. Funds that established their reputations on SaaS multiples are showing interest in grid-scale storage firms, long-duration energy storage entrepreneurs, and transmission technology providers.
Venture capitalists in this field frequently point to the internet infrastructure development of the 1990s as a historical analogy. The internet was routed by Cisco. The data was carried by the ISPs. The cable was installed by fiber optic firms. The infrastructure suppliers produced long-lasting, structural profits because every online company required their services and couldn’t develop the internet without them, but none of them were as thrilling as the businesses creating websites. The current capital deployment is betting on whether AI infrastructure generates similar long-term returns or whether the technology commoditizes in ways that reduce infrastructure margins, as happened to fiber optics after the buildout exceeded demand in the early 2000s.
