Why AI Analyzes the Past Well but Venture Capital Bets on the Future
Three quarters of venture capital firms already use artificial intelligence to evaluate investment opportunities. That figure alone sounds like inevitable modernization. But there is a structural tension that this percentage does not capture: language models are extraordinarily good at doing exactly what venture capital cannot afford to do too often, which is looking backward.
Venture capital is, in its most basic mechanics, a bet on discontinuities. Not on markets that expand in a predictable manner, but on moments when a technology or a behavior breaks away from what previous data suggested was probable. Introducing tools trained on historical patterns into that process is useful until it stops being so, and the boundary between both states is narrower than most firms are acknowledging out loud.
The Most Sophisticated Confirmation Bias That Has Ever Existed
Large language models generate responses by identifying patterns in massive text corpora. This makes them extraordinarily capable for analytical tasks with well-defined contours: mapping competitors, identifying regulatory obstacles, summarizing technical literature, flagging risks in a known market. What they cannot do, by construction, is recognize the moment when those contours are about to redraw the entire map.
The history of venture capital is filled with examples where the correct analysis of the present was the very reason the future was missed. When Airbnb raised its first rounds in 2008, the thesis that strangers would pay to sleep in other people's homes was not just counterintuitive; it was directly inconsistent with the data available at the time about consumer behavior. The sentiment analysis of that era pointed in the opposite direction. The same was true in the early stages of the social web: the dominant surveys from the early 2000s showed that the main barrier to internet usage was fear of privacy. Facebook was built, in part, by ignoring that reading.
A well-calibrated system would have flagged both proposals as high risk. And it would have been right, from the perspective of the past. The problem is not that the analysis was incorrect; it is that it was the wrong analysis for that specific decision.
This is where the bias becomes difficult to detect within firms that have adopted AI as a standard part of the due diligence process. It does not manifest as an obvious error. It manifests as a series of very well-documented analyses that systematically favor bets that have precedent and disfavor those that do not. In the short term, that produces more orderly portfolios. In the long term, it produces portfolios that do not generate the returns that justify the asset class.
What the Flow of Capital into AI in 2025 Reveals
The concentration of global venture capital in 2025 illustrates that pattern with precision. Global venture capital funding reached approximately 141 billion dollars in the fourth quarter, an increase of 12% compared to the previous quarter, making 2025 the most active year since 2021. Artificial intelligence represented more than 25% of global venture capital that year, up from 15% in 2024 and 7% in 2023. In the enterprise segment, spending on generative AI grew from 11.5 billion dollars in 2024 to 37 billion in 2025, according to data from Menlo Ventures.
Those numbers describe an industry that, in part, is betting on the future with genuine conviction. But they also describe an industry that, in part, is following the most legible pattern available. AI is today the sector with the most recent historical validation, with the most citations in research papers, with the highest volume of news flow. It is, in practical terms, the market on which an AI tool can most easily produce analysis. The result is a feedback loop that concentrates capital where the signal is clearest, which is exactly the type of concentration that produces the highest median returns but not necessarily the returns in the upper-right tail.
The distribution within the AI segment also deserves attention. Menlo Ventures breaks down 2025 spending between 19 billion in the application layer and 18 billion in infrastructure. Within applications, horizontal tools captured 8.4 billion, departmental solutions 7.3 billion, and specialized verticals 3.5 billion. That level of granularity suggests that the bet is no longer about whether AI matters as a category, but about which layer of the value chain will capture sustainable margins. That is a much more refined question, and it is precisely the type of question where well-executed analysis, with or without AI, can provide differential value.
What historical analysis cannot resolve is identifying which categories that do not appear in any dataset today will capture the next wave. Modular nuclear energy is the clearest example at this moment.
When a History of Failures Conceals a Real Discontinuity
The records on nuclear energy are full of warnings. Three Mile Island, Chernobyl, Fukushima. Decades of failed commercialization attempts. Construction timelines that stretched from years to decades. Structural cost overruns. An analytical system trained on that corpus would produce, in a completely reasonable manner, a high-risk assessment for any startup proposing small modular reactors as an energy solution.
The problem is that small modular reactors are technically and economically distinct from the large-scale nuclear plants that generated that track record. They are designed for serial manufacturing and standardization, not for bespoke construction at each site. And the demand context has changed in a structural way: AI data centers require volumes of continuous and predictable energy that intermittent sources cannot satisfy in an economically efficient manner at scale. Companies such as Microsoft, Google, and Amazon have already begun signing agreements and making investments linked to nuclear generation, which indicates that the demand signal exists and is being formalized in contracts, not merely in statements of intent.
A model trained on the nuclear past would likely see accumulated risk. An analyst who understands what has changed in the economics of energy demand can see a technology reaching the market at the very moment the market finally needs it. The difference between those two readings is not optimism versus pessimism. It is the capacity to identify when an external variable has reconfigured the space of possibilities for a technology that was previously unviable.
That capacity cannot be delegated to a system that does not have access to information about what has not yet occurred.
Imagination Is Not an Analytical Luxury — It Is the Variable the Model Cannot Import
What venture capital has historically purchased is not analysis of the existing market. It has purchased the capacity to imagine markets that do not yet exist and to identify the teams that can create them. That capacity has components that are not delegable to systems that analyze historical patterns: the reading of weak signals, the ability to recognize when an emerging behavior is about to become mainstream, the distinction between a market that failed due to poor execution and one that failed because the timing was wrong.
None of this means that firms should reduce their use of AI in the investment process. Current tools are genuinely valuable for accelerating due diligence on known markets, stress-testing business model assumptions, and structuring competitive analysis. Used well, they make more rigorous the work that analysts were already doing.
The risk is not in using AI. It is in building processes where the absence of historical precedent automatically becomes a signal for rejection. That institutional design expels from the portfolio exactly the opportunities that venture capital should be best positioned to capture.
The capital that flows to where models can already measure well is competing against every firm that uses the same models. The capital that flows to where models still cannot measure well is competing against far less. That asymmetry does not disappear because the tool becomes more sophisticated. It deepens.
