What the Failure Rate Gets Wrong

About 75% of venture-backed companies never return capital to investors. That figure, from a Harvard Business School study of more than 2,000 companies funded between 2004 and 2010, gets cited all the time when someone wants to talk a serious person out of a serious bet.

But what if it's the wrong number to be citing? The number is accurate. It just doesn't measure what people think it measures.

What the failure rate tracks is whether a specific company, at a specific point in time, converted its thesis into a self-sustaining business with returns to investors. What it doesn't track is what the company produced on the way: the knowledge it generated, the researchers it trained, the tools it built, the approaches it either validated or proved obsolete.

In most businesses that distinction doesn't matter much. The company is the value container, and when it closes, the value closes with it. In frontier companies, those that are built on moonshot bets, that's not how the value flows.

The Genomics Wave

In the late 1990s, a generation of companies — Celera, Incyte, CuraGen, Human Genome Sciences — raced into commercial gene sequencing on the premise that controlling genomic data would be among the most valuable positions in medicine. The business models failed. By 2005, Celera had placed its formerly proprietary genome sequence data into the public domain. The sector looked, by conventional accounting, like a record of misallocated capital.

Except the knowledge didn't disappear. The sequencing infrastructure, the laboratory methods, the analytical software, the researchers trained in those programs: all of it redistributed into the next generation of genomics. The companies now editing inherited disease out of human genomes are in many ways building on a substrate that the "failed" wave funded.

Take Colossal Biosciences. The de-extinction technology company reached a $10.32 billion valuation by September 2025 and had reintroduced the dire wolf in April of that year, using multiplexed CRISPR tools, ancient DNA sequencing, and AI-guided genome analysis that accumulated across decades of prior frontier work.

The same knowledge-transfer logic is visible in Astromech, the AI-biotechnology company Ben Lamm founded after co-creating Colossal, building alongside Harvard geneticist George Church. Astromech is developing a predictive engine for biology — applying AI to forecast evolution, disease risk, and system vulnerabilities before they manifest. The company reached a $2 billion valuation within nine months of launch, a direct extension of the genomic and AI infrastructure Colossal and its predecessors spent years building.

The DARPA and Apollo Model

One organization that has deliberately built an understanding of the value of a moonshot into its operating model is the Defense Advanced Research Projects Agency, or DARPA. It accepts, as a feature rather than a concession, that most of its programs won't achieve their stated objectives. Goals are sometimes set at levels that analysts describe as virtually impossible to achieve. The knowledge generated by those near-misses seeds the next generation of attempts and reduces the cost of what follows. Behind every DARPA success, as the agency's own analysts describe it, is a mound of failure that made it possible.

Apollo worked the same way at a larger scale. By 1963, the program was consuming roughly 60% of U.S. integrated circuit production, not because anyone planned to build the semiconductor industry, but because the program's requirements forced that level of use and development. That pressure generated more than 1,800 traceable spinoffs and helped define the structure of the semiconductor industry itself.

The Right Measure

Analysis by Benedict Evans of roughly 7,000 Horsley Bridge investments made between 1985 and 2014 found that 6% of deals generated 60% of total returns. Within that distribution, the companies with the biggest outcomes were often the ones that attracted the most skepticism at founding. The failure rate is a description of that pattern, rather than a consideration of its cause. While only a small percentage of companies succeed, the ones that do are often building on or reimaging the failed moonshots of the past. Those that succeed often do so precisely because companies that failed before them took the same kind of chance.

Rather than ask how likely a company is to fail, the better question for evaluating a frontier company may be: what capability is being produced per dollar deployed? How much technically novel knowledge? How many researchers trained at the frontier of disciplines that barely existed when the company was founded? What tools does the category now hold that it didn't hold before?

Those outputs are harder to model than a return multiple. They're also the primary mechanism by which frontier investment generates durable value, compounding across companies and across time, long after individual vehicles close.