A partner at a mid-sized venture firm opens a spreadsheet he would prefer not to make public on the fourteenth floor of an office building in midtown Manhattan. It displays the company’s AI-related portfolio positioning in relation to anticipated timetables for enterprise adoption. Eighteen months ago, the lines were meant to align, but they don’t. The technology advanced more quickly than anticipated. The business clients moved more slowly. The majority of the worry in the VC community in New York today resides in the space between those two facts.
Compared to nearly every other technological area in recent memory, generative AI received more financing more fast. In retrospect, the wagers made by the companies that were early adopters of foundation model infrastructure and AI tooling appear clear, and this is the version of history that is frequently recounted. The fact that many of those same organizations are now covertly sponsoring businesses that exist specifically because the initial wave of generative AI failed to fully meet the needs of the enterprise sector is something that is less often recognized. tools for integration. layers of reliability. vertical-specific applications designed for sectors where the outputs of generic large language models are insufficiently precise or auditable for practical implementation.
The term “capability-reliability gap” is being used in New York venture talks, and it very much sums up the issue. In demonstration situations, the models are capable of amazing feats. It is a different and much more difficult issue to get them to execute those things consistently, at scale, within regulated businesses, with outcomes that can be examined and justified. In 2024 and 2025, enterprise software purchasers learned this the hard way, and the adoption curve that the most bullish forecasts had anticipated just didn’t show up on time. Businesses who placed large bets on pure infrastructure are now inquiring as to when the money will really come in.
On paper, the emerging hedging technique appears contradictory. Some New York funds are concurrently investing in vertical AI-native applications and foundational model companies, which may eventually reduce the need for such basic models at the enterprise level. A company may support both a general-purpose AI infrastructure player and a legal software company that has developed a proprietary model particularly trained on contract language and case law. There may be competition between those two wagers. A many of them already are. Two years ago, the LPs supporting these funds were not questioning the reasoning behind portfolio building. Now, they are.
Lower-tier venture companies now use the term “capital efficiency” to explain why their current investment slate differs from that of 2023. Only a small number of investors can truly maintain the infrastructure spending necessary to train and operate large foundational models at a competitive scale. The majority of this capital is concentrated with Andreessen Horowitz, Sequoia, and a few other firms with fund sizes large enough to absorb the burn rate. Applying AI to particular, high-value vertical activities where the client acquisition route is shorter and the unit economics make sense without having to win the foundational model race has become the cleaner bet for New York enterprises dealing with smaller pools.
Observing this from the outside, it seems like the story surrounding generative AI is disintegrating more quickly than the technology itself. Although real-world deployment friction that wasn’t apparent from the outside during the peak enthusiasm era has complicated the initial story—foundation models as the platform on which everything gets rebuilt—it hasn’t been shown to be incorrect. The companies that are changing their stances aren’t necessarily negative about AI in general. They are paying close attention to the version of the AI future they are funding. Compared to the pure momentum bets of 2022 and 2023, that is a more sophisticated stance, and it most likely represents a maturing market rather than one that is retreating.
It is still really unknown if the post-generative AI bets—the neuro-symbolic systems, the specialized enterprise models, and the AI-adjacent infrastructure that makes deployment actually work—will yield the returns that businesses are currently projecting. These categories may also be affected by the timeline issue that affected generative AI. However, the hedging is still in place for the time being.
