Where Returns Lie in Venture Capital

Recently, I’ve been thinking a lot about the nature of early-stage venture investing. In a world where multi-stage investment platforms are gobbling up LP (limited partner) dollars (allowing these funds to outgun smaller rivals) and AI deals command a 50-100% premium relative to broader software deals, how should “normal” early-stage funds generate returns? As I’ve pondered this more — I’ve concluded that non-consensus picking remains an under-appreciated source of alpha.

In the following post, I cover the following:

• What are the constituent parts of the VC job (sourcing, picking, winning, supporting)?

• While there’s a ton of effort spent on sourcing, winning, and supporting, there’s comparatively less emphasis on true, non-consensus picking

• Why non-consensus investing is much easier said than done

• Several examples where funds have generated outsized returns given their ability to make the right non-consensus investments, as well as opportunities that I’m thinking about

Let’s dive in.

A primer on the VC “job to be done”

Broadly speaking, the “job to be done” for venture can be broken into four main components:

This flow consists of finding great companies (sourcing), evaluating these companies (picking), convincing founders to take your money vs. that of rival funds (winning), and helping companies post-investment (supporting). Theoretically, each job to be done helps drive investment returns.

While funds currently spend a considerable amount of time sourcing (for AI deals this might mean thoroughly penetrating elite research organizations like Open AI, Anthropic, DeepMind, etc. or mapping out key authors of papers on arXiv), winning (paying high prices, using firm brand, moving faster, etc.), and supporting (helping close the right hires, key customer intros), it does feel that picking is somewhat overlooked, and can often feel more like art than science (especially for a nascent, rapidly changing domain like AI). This to me means there are significant returns to be had in overlooked, non-consensus deals.

The art of non-consensus picking

Making the right “pick” is probably one of the most intellectually fascinating parts of investing, and for VC, there are a number of non-investment related dynamics at play here as well. One common framework that VCs use to reason about investments is via a 2x2 matrix. On one axis, an investment is either right (makes money) or wrong (losses money). The other axis is whether a deal is consensus or non-consensus (a consensus deal simply means that the majority of investors might feel a startup is worth investing in).

Going through our 2x2 matrix, it’s relatively easy to reason through consensus deals. Investors at top firms are generally adept at identifying startups or investment categories that seem like they could generate large outcomes (currently it’s all things AI). As a result, there tends to be a bit of a herd mentality where VCs over-allocate to a category and consequently drive up entry prices. As an extreme example, if the blended cost basis of a portfolio consisting of all “hype” AI deals is $50M at the seed stage (vs. a more normal $15M), then this automatically reduces potential returns by >3x (assuming the same exit size)! Therefore, even when investors are right, returns end up being compressed.

The non-returns dynamic here is when things go wrong. For lead investors, they can point to the fact that their failed investment was the best bet, given the information available at the time (and the competing term sheets). For non-lead investors, they can point to the reputation of their lead investors. This is essentially the VC version of the “no one gets fired for buying IBM”. It’s understandable then why it takes incredible conviction to issue a term sheet to a startup when there are no other competing investors around the table! The other dynamic here is VC fundraising — meaning VCs have to raise a new fund every 2-3 years, but the “gestation period” for non-consensus deals might be longer, so for the sake of markups, investors might pile onto consensus deals.

Now, actually making non-consensus right investments is much easier said than done, especially in today’s highly competitive funding environment. My thoughts are as follows:

• Investors are generally getting more skilled (techniques/frameworks to analyze startups are fairly well-known at this point, decline in information asymmetry, etc.), so it’s difficult to truly have a differentiated investment thesis derived from some unique insight. I think the “Paradox of Skill” applies in VC as well (Michael Mauboussin of Morgan Stanley’s Counterpoint Global argues that as investors get more skilled, more of the outcome is decided by luck). In the world of VC, this means two things for me: for consensus deals, there are typically a handful of great startups within a given category (like AI inference). The best VCs will flock to a small handful of assets, and which one of these companies “breaks out” ends up being quite a bit of luck. For non-consensus investments, things get even more dangerous still, as a given fund would need to take additional risks that rival VCs might not be willing to take (whether it’s technology, market, or founder). In a world where venture returns are already balanced on a knife’s edge (a very small handful of investments drive the majority of the returns), taking additional risk means an even higher variance in return profile

• Relatedly, there’s timing risk as well — meaning even if an early-stage firm makes a contrarian bet on a non-consensus startup, it has limited time to become consensus for downstream investors. As an example, if an early-stage firm backs a startup at the seed, and it’s still non-consensus at the Series A, that firm must take additional risk to lead the A, in hopes that by the time the company is Series B ready, a growth stage investor can step in. VC lore unfortunately is littered with companies that were born before the world was ready (General Magic is one company that was decades ahead of its time)

• Even if a deeply contrarian fund identifies a promising startup that other funds might have overlooked, there is a bit of a Dark Forest dynamic here as well. This means that if a contrarian fund issues a term sheet to a startup, that alone becomes a signal for other funds to pile on. This would happen if the fund in question is well-respected, and the startup then shops the term sheet around. Investing here then goes back to being a game of winning, not picking!

• Funds also can’t afford NOT to invest in a consensus right category like AI, given there will absolutely be monster companies that are born in the current super-cycle (and LPs likely want exposure as well). The question is how investors pick the right high potential (and non-consensus) startup in a consensus category. Because there’s still significant fog of war in AI, there are still likely fund multiplying investments outside of an Open AI/Anthropic yet to be made, which also means the legendary status of future firms is still up for grabs!

As we can see above, non-consensus picking is not easy! But, if a given fund makes enough non-consensus right investments, it will also improve its brand positioning, which in turn will help with sourcing/winning! Annoyingly, this process takes years for mark ups, and potentially over a decade for real distributions to LPs.

Framework for non-consensus thinking

As a bit of an amateur VC historian, I quite enjoy stories of funds making non-consensus bets that generate massive returns. Union Square Ventures’ Thesis 1.0, which led to investments in Twitter, Zynga, Indeed, etc. and DST’s $200M bet on Facebook that generated $4B come to mind. A more recent example of non-consensus investing is that of Silver Lake and Sixth Street’s $1B bet on Airbnb during the height of Covid-19, which came in the form of a loan with warrant coverage that valued Airbnb at $28.56 per share (~$18B valuation). By the time Airbnb opened for trading, it was valued at $146 per share, good for >5x return in less than a year (without factoring in dilution). The funny thing is that all of these non-consensus bets look very obvious in hindsight!

So how should we think about non-consensus picking today?

Here are several angles that I’ve been thinking about:

• It’s generally a good idea to invest in a large market that’s also growing. From that perspective, India is one potential target. For something like a single HQ fund on Sand Hill Road, however, India becomes a difficult target to invest in (unless that fund has local boots on the ground, like what Peak XV was to Sequoia Capital before the split). Given the run-up of India’s Nifty 50, it could also be argued that the country has already become consensus

• Categories that have become “underinvested”, like consumer. But consumer startups have their own issues! My friend Aaron and I have discussed the opportunities & perils of consumer investing at length. While successful consumer outcomes generate the largest outcomes, there’s comparatively less downside protection (vs. something B2B) when a consumer company doesn’t reach escape velocity. As I think about this from a portfolio construction perspective, a consumer-only fund might end up being a basket of zeros! But, if a fund invests in both consumer AND enterprise, then that fund has even less chance of “hitting” the right once-in-a-decade consumer company. As an aside, this is one reason why LPs diversify across fund strategies and vintages (time diversification), so that if one fund truly is a bit of a “stinker”, other funds can help smooth out returns.

• The resurgence of “world of atoms” companies, with consumer robotics being one such example. However, it does seem that there hasn’t been great consumer hardware startups that have come out of the Bay Area in recent years, as founders’ focus has remained on software (as famed YouTuber MKBHD can attest to here and here). It’s arguable that at this point Shenzhen has a stronger consumer hardware culture, and the recent big consumer robotics wins were Chinese companies (DJI, Roborock, etc.). Bambu Lab is a more recent example that has reached hundreds of millions of dollars in sales after only several years. As the next generation of consumer robotic products hit the market, the founder profiles of these companies will be very different from those found in the hacker houses in SF, and these will be founders that Bay Area VCs can’t easily source.

• Taking geopolitical risk. As the major superpowers (US and China) continue to decouple, a bold investor might take the contrarian bet that relations between the two countries will one day renormalize, and invest in Chinese startups (though this is mostly off-limits to US investors for sensitive categories like AI, semiconductors, and quantum). The historical analogy of this was that of Hermitage Capital after the fall of the Berlin Wall. DST’s recent investment in Xiaohongshu is a modern day example of an overseas fund taking the non-consensus bet on China. Given China seems to be making some unforced errors, it’s difficult to imagine these investments being able to exit in the near future. ByteDance has offered to buy back shares from investors as a workaround, however.

As we reach the end of “pure” software investing, some VCs are already beginning to experiment with new strategies, as evidenced by Forerunner’s investment into auto maintenance franchise Stress Free Auto and Slow Venture’s thesis on the Growth Buyout Model. But, as we can see from above, these alternative angles of attack are fraught with danger, and it requires great courage and the ability to endure prolonged pain to make these kinds of investments. Here, a fund that has historically generated great returns will be given more rope by LPs to take greater risks.

Non-consensus thinking in an era of uncertainty

2024 has certainly shaped up to be a very interesting year for investing, given the wobbliness in the US economy, lack of liquidity in the private markets, the high-stakes/frothy AI deal-making environment, and the broader geopolitical uncertainty (more applicable for investors with a global lens). This raises real questions for capital allocators: should investors double down on AI, stay cautious on the sidelines, or bet on something else altogether? All of these things are happening while the VC ecosystem is going through its own period of maturation.

These factors introduce additional degrees of uncertainty as investors look to deploy out of their current set of funds. I think this multi-variate investment environment is precisely the moment when thoughtful, contrarian thinkers get to make their best investments, and I’m excited to see which of the current-day non-consensus right bets become case studies for future investors.

Huge thanks to Will Lee, Aaron Wong, Gonzalo Mocorrea, Yash Tulsani, Monica Xie, Lisa Zhou, Danielle Jing, Liam Armstrong, and Andrew Tan for the data & feedback on this article. If you want to chat about all things ML/AI, investing, and policy, I’m around on LinkedIn and Twitter!