These days, when the topic of 2021 comes up in venture capital meetings, there’s a certain silence. Before continuing, someone usually clears their throat after the slides are raised and the numbers are displayed. It’s not that the year was a complete catastrophe. Four and a half years later, the math just doesn’t add up the way it should. The vintage that everyone battled to enter has become the one that everyone is discreetly attempting to remove from the books.
The disappointment is so dramatic because of its scope. In 2021, global venture capital reached a peak of over $800 billion, of which about $257 billion went to AI-tagged companies prior to ChatGPT’s arrival, giving the category a legitimate product story. Funds were being raised at unprecedented rates. Term sheets were being signed in a matter of days or even hours. Additionally, the agreed-upon valuations, especially for software wrappers and applied-AI startups, only make sense if you assume that revenue will increase tenfold annually for five years. It hardly ever does. Here, it hasn’t either.
“Down round,” which describes what happens when a startup raises new capital at a lower valuation than its previous one, is currently the most courteous term in use. Limited partners use the less courteous term, “markdowns of 30% to 70% on positions taken at 2021 prices,” in private. A few of these businesses are still in operation today. Some engineers are performing well. However, the entry point was so high that no reasonable execution could yield a return sufficient to cover the initial check. Speaking with general practitioners lately has given me the impression that they are aware of this and are merely awaiting distribution schedules to provide them with protection.
It is a structural issue rather than a psychological one. The so-called Total Cost of Inference, or the cost of operating AI at scale, proved to be far more expensive and stickier than anyone had anticipated. Every time a startup built on top of Anthropic or OpenAI APIs expanded, their profit margins shrank. The cost of GPU computing remained high. Additionally, the layer above them continued to be commoditized by the underlying models. By 2024, a product that seemed defendable in 2021 appeared to be a feature within a Microsoft release. The true cost of “no moat” is now being revealed to investors who funded thin wrappers.
Exits—or rather, the lack of them—also contribute to the disappointment. Since 2022, the IPO window for venture-backed businesses has been largely closed. When it does open, the few names that make it through are typically infrastructure plays rather than the application-layer software that characterized the deal flow in 2021. The 2021 cohort has nowhere to go in the absence of IPOs and strategic acquisitions at reasonable multiples. They continue to increase bridge rounds, reduce staff, and convince themselves that the upcoming funding cycle will be better. Perhaps it is. It’s still not clear.
Observing this unfold from a distance is unsettling because it appears so predictable in retrospect. The same pattern emerged from the dot-com vintages of 1999: painful overcorrection at the bottom, irrational exuberance at the top, and a few survivors who ultimately justified the entire asset class. AI will most likely take the same course. Eventually, some of the wagers from 2021 will become legendary winners. Most won’t. According to CIO surveys, 71% of executives say they will reduce AI budgets if value isn’t proven within two years. For those who still have paper from the 2021 class, that figure should be more concerning than any markdown.
The speed at which the language has evolved is difficult to ignore. Term sheets from 2021 discussed scale and transformation. These days, they discuss payback periods, gross margins, and “responsible burn.” The discipline is back. Infrastructure and vertical AI are now the focus of the capital, which is moving more slowly and selectively. Most of the 2021 founders, who raised $1 billion on $5 million in revenue, are still operating leaner businesses in the hopes that the next cycle will be kinder. A few of them will survive. Many won’t. Even when meetings remain silent, the spreadsheets continue to whisper this truth.