The "death zone" of venture capital ☠️
How Parsnip can shortcut the difficulty of a venture-scale outcome in edtech
Mountaineers know that there’s a “death zone”1 where there’s too little oxygen to sustain life. Apparently, VCs know that there’s a death zone for their investments, too. Michael Narea from Transcend Network2 first introduced us to this idea.
It’s difficult to make a long list of venture-scale outcomes in edtech. Duolingo is one, but was backed only by generalist funds with no edtech investors. Masterclass might be another, but some would argue that it’s more entertainment than education. Coursera is maybe the most prominent “canonical edtech” example, but it’s not clear if they’ll ever turn a profit.3 What’s going on here?
It turns out that selling learning (or learning-related products) is hard, and most education or learning companies are poor venture investments. Statistically, edtech companies haven’t been fund-returners and thus purely by pattern matching, VCs have an immediate allergic reaction to them.4 Some investors may not even meet with us for this reason, others mention it, and for everyone else we pre-emptively point it out to them—because it’s important to understand why, from first principles, Parsnip’s approach is different.
Through navigating the idea maze in education, we’ve come to understand the landscape better, and want to share our mental model of why learning-related businesses are hard. Let’s first explore B2B businesses, then B2C.
“Traditional” B2B edtech
Edtech has historically referred to software that’s sold to educational institutions. We can divide these products into two categories: learning products and institutional SaaS.
Learning products are tools to help students in schools learn—such as digital curricula, interactive reading, or educational games. But these all face the 3-party problem5, which requires them satisfy:
the buyer, the administrator or school district that’s paying for the product
the teacher, who is overseeing and supervising use of the product
the student, the end user who’s actually using the product
It’s hard enough to build a good product in the first place. It becomes far more difficult when trying to satisfy 3 stakeholders at once.
Furthermore, through the lens of B2B sales, schools and universities are more complex than enterprises—because of the idiosyncrasies across educational institutions, one sale might be vastly different from another, making a repeatable sales motion harder to achieve.
Insitutional SaaS products encompass many varieties of B2B SaaS software that happen to be sold primarily to educational institutions to help with organizational logistics. Examples of these include Blackboard, ClassDojo, and Powerschool.
Many successful businesses in this space have profitable unit economics. Among edtech startups, these are the most likely to result in outsize outcomes, yet are still hamstrung relative to enterprise B2B as selling to schools is less repeatable than selling to other businesses. And with historically fewer unicorns compared to other B2B SaaS businesses, this category is still unappealing to many generalist investors.
Finally, despite the comparatively better financial outcomes here, these software products for institutions aren’t directly helping students learn.
“New age” B2C education/learning products
More recently, education software has started bypassing distribution through institutions and instead targets consumers directly. Duolingo is a pioneer in this space—these products are mainly aimed at individual learning, and are often powered by AI.
Giving consumers direct access to learning is more straightforward and sidesteps the 3-party problem. Selling a product becomes easier when the person paying is the same person that’s learning—or at least the parent of the learner. But there’s a more pernicious limitation here of wide vs. deep business models.6
The first category of “wide” learning products:
target broad, mass markets
are characterized by low CAC and low LTV
have “casual users” that are cheap to acquire, but won’t pay much.
compete with many other products for users’ attention
The canonical example here is Duolingo, with 83M MAU in 2023 but only 6-7% paying users. Coursera, also with millions of users, is still not profitable. Khan Academy is similarly popular, but depends on significant support from donations.
More subtly, the need to appeal to a broad market while simultaneously increasing retention and monetization creates pressure to drive toward entertainment over learning. This focus on growth and revenue can cause product features to converge to the lowest common denominator that appeals to a mass audience. For example, Duolingo has transitioned from being primarily learning focused at launch to taking on more Candy Crush-esque features in order to maximize active users and revenue.
In summary, wide learning products require enormous scale to succeed, while also facing pressure to muddle the learning experience in order to be profitable.
“Deep” learning products can avoid these issues, and:
target niche markets
are characterized by high LTV and often high CAC as well
have “power users” are willing to pay for actual learning, often significantly, but who can be hard to acquire
have fewer substitute products or competitors
By the nature of targeting niche markets, deep companies aren’t household names. These businesses can have highly profitable unit economics, but are hard to scale beyond a specialized market, limiting their size7 and making venture-scale outcomes difficult.
Any learning product can either be wide or deep, but not both—and faces inherent constraints in each case. What do you think of this framework?
Parsnip’s loophole ➰ around wide vs. deep
If it’s difficult to sell learning by itself, what if we combine it with doing?
Our thesis for both the product and the business model emerge from the observation that there are two fundamental blockers to making cooking easier: “how do I cook?” and “what should I eat?”. It turns out that the first is a learning problem that’s solvable with AI skill tree learning, and the second is a logistics problem that’s solvable with an AI copilot for the kitchen.
Taken together, these allow us to escape the wide vs. deep conundrum.
The first part, AI skill tree learning, resembles a wide, mass market business. There’s potential for a strong distribution advantage here—indeed, the >40k downloads of Parsnip so far were all at 0 CAC. Under this model, we’d keep a sizable chunk of learning free, with purchases or subscriptions for more advanced cooking techniques and ethnic cuisines.
The ability to personalize any recipe on the internet enables us to focus on social features and product-led growth, making possible the viral distribution that other food-related products (e.g. the Instant Pot) have benefited from. Yet, a standalone learning product might struggle to deliver enough value over time for users to consistently pay. However, the personalized knowledge map from an AI tutor unlocks an even more impactful idea.
The second part, the AI copilot for kitchen logistics, is not a learning product at all, but a platform business. Each user’s individual skill tree allows an AI to help them make decisions in a personalized way that wasn’t possible before. Our goal is to make cooking 10 times easier—and the fastest, cheapest, and healthiest way to eat, so that it will eventually replace fast food, takeout, or meal kits. At its full potential, this experience is a game-changing experience for anyone who eats. This “AI chef” by your side can deliver continual, consistent value over time and convenience that users are willing to pay for.
In other words, Parsnip’s not limited by selling learning because our plan isn’t to sell just learning. It’s to build a platform business in the food space that can grow through network effects, unlock massive B2B revenue, own the home kitchen, and even unbundle food media from other platforms. Cold-starting this platform by itself would be burdensome due to the need for both users and personalized user data—but the skill tree already provides both!
From AI skill tree to a platform business
Putting these together creates a uniquely powerful combination that eschews the limitations of selling “just” learning, allowing for:
Better learning where learning is tightly coupled with doing, and where what you learn immediately applies to real-world actions—way more fun than getting tested in school!
A better business through a platform that captures more of the value from the fruits of our learning and what it facilitates in our everyday lives.
Finally, AI-powered learning is a wide, low-CAC funnel that channels users into a sticky, high retention, high-LTV platform, creating enormous growth potential.8 The enormous potential here is to capture the home kitchen vertical with a zero-marginal-cost, AI + software powered solution and a highly defensible moat of users’ personalized cooking skills and meal history.
This is the core of our business model thesis. If it works, then we’ll have liftoff 🚀.
In other news, thanks to Dan’s hard work, Parsnip is now part of the Stanford-affiliated StartX accelerator and community. Stats about StartX include:
92% of member companies growing or exited (so we'll try not to be in the other 8% 😅)
3x more likely to reach $100M+ valuation and 4.5x more likely to reach $1B+
average of $11M raised post-StartX
We know better than to get caught up in vanity metrics, but simply put: StartX is a talented, well-connected community that opens many new doors for Parsnip.
We’re also grateful to be featured again in the App Store for New Year’s resolutions!
Transcend runs a fellowship we’ve heard great things about, and they also write a very informative edtech newsletter.
And also to food-related companies, though that’s for a separate post 😊
Not to be confused with the three-body problem and its chaotic dynamical system, or the popular Three-Body Problem science fiction trilogy.
Cheers to Elijah Mayfield for introducing us to this idea.
There are many fantastic businesses in this category, which can reach valuations upwards of $100M and serve their customers exceptionally well. However, for better or worse, these outcomes are still unappealing for the venture funding model.
Our long-term vision is to build this holistic freemium revenue & growth model. But we can’t boil the ocean and it would be inadvisable to build everything all at once, so we’ll likely start with a B2C freemium model on the AI learning component.
Great analysis of a complex challenge