One quick and easy way to understand the state of EdTech in the US right now is to consider the most popular platforms that students use to learn everyday: YouTube and Roblox. The largest EdTech companies in the US, in other words, are not EdTech companies at all.
They are, however, user-generated, free, and appealing to students—all of which make them genuine threats to the proliferation of learning platforms that are seldom any of these things. They are also the reason I’m fairly bearish on almost all new EdTech products and platforms, both in their ability to educate effectively and as investments that will ever produce venture-like returns.
As a quick disclosure, I run an education and EdTech company myself, invest in EdTech, and use EdTech platforms to continue my own education. And in many ways, it is a promising era for EdTech, not only because of the record number of investments in EdTech since the pandemic. Video technology is finally good enough to scale large, live classes. That in turn opens a massive market for educators to become the next big class of creators with giant followings—like the Hong Kong and South Korean tutors who have earned millions of dollars a year over the past decade as educator-celebrities. Meanwhile, continued advancements in machine learning can empower educators with more data to be responsive in real-time to students’ challenges, confusions, and successes, so they can customize accordingly. These three points all make an especially bullish case for live classes, which can leverage tech to create an interactive and gamified experience that feeds data to top educators who can now reach hundreds of students at once.
Nevertheless, in an era where content is free, any successful venture needs to enhance the educational experience itself for teachers, parents, and students—and the near-impossibility of doing so is why I’d like to indulge in the case against EdTech. In particular, any American EdTech company is likely to come across three challenges: first, the difficulty of scaling high-quality education that is traditionally one-on-one; second, the difficulty of retaining students after one-off offerings; and third, the difficulty of appealing to students when parents and schools are the paying clients.
Challenge #1: The Best Education Doesn’t Scale.
Benjamin Bloom’s 2-sigma problem has dogged education for decades with its central proposition that one-on-one tutoring bolsters students’ by two sigma: that is, students go from averaging in the 50th percentile in classrooms to the 98th percentile with direct instruction. That, in turn, has raised a crucial question that’s never been properly answered: how can direct instruction be scaled?
There have been some nuanced criticisms of the overall applicability of the 2-sigma problem, to be sure. Doesn’t the quality of the tutor matter? Doesn’t the subject matter? Does rote learning like basic memorization really require a tutor, or could it be amenable to software-based programs like MOOCs and apps? (Indeed, evidence indicates it can. It’s the critical thinking and conceptual understanding that’s harder to come by without a human coach to customize to the student’s needs.)
Nevertheless, the obvious point—that top tutors can get far better results than standardized programs by coaching, customizing, and caring—is not in dispute. Nor is its point that these experiences can’t be scaled. It is notable that Marc Andreessen featured Bloom’s 2-sigma problem prominently in his seminal essay, “It’s Time to Build,” by asking why we haven’t “built systems to match every young learner with an older tutor to dramatically improve student success?” It is perhaps even more notable that Andreessen’s firm, which has offered some of the best coverage of EdTech, has not invested in any such company.
To understand why it’s nearly impossible to scale direct learning, it’s worth studying what is likely the most successful one-on-one tutoring firm in the world, VIPKid. A platform that connects tens of thousands of American educators to hundreds of thousands of Chinese students to teach English, VIPKid has arguably found the sweet spot that allows it to scale direct instruction. On the one hand, each student works one-on-one with a teacher to practice speaking English outloud—the kind of learning exercise that becomes exponentially better with fewer students and would be vastly compromised in a larger class setting. On the other hand, each teacher is working entirely from a pre-written curriculum that requires little prep or effort: the educators are, in many ways, actors reciting a script that can be scaled to as many people as are available to offer it. As of 2019, had 90k North American teachers on the platform with around 700k students in China—and a valuation of around $6B, backed by Tencent, Sequoia, and Kobe Bryant.
But three points might make us skeptical of the validity of this approach more broadly—particularly in America where companies like Juni are attempting a similar model. First, ESL instruction is a bad model for other subjects, primarily because no other subject demanding one-on-one work can be scaled so easily. In other words, ESL is fairly unique in requiring one-on-one instruction for practicing the rote skill of sounding out words, even while the instructor can be nearly any native speaker reading the pre-written script. That’s not true for other subjects, which might be practiced rotely through a standardized curriculum without need for one-on-one (CodeAcademy for programming, for example) or might benefit from one-on-one instruction from a specialist for deeper conceptual understanding and critical thinking (nearly all of them). ESL, by contrast, is unique in that it requires one-on-one without requiring the specialist. Second, ESL generates a market for ancillary instruction that no other subject does: it is viewed as a globally-required skill that schools can’t sufficiently teach since their instructors aren’t native speakers.
And finally, VIPKid itself is barely a success. Investors have evidently been concerned for years that VIPKid has no path to profitability and has already tapped out its potential market while incurring higher and higher marginal customer acquisition costs. Indeed, it spends massively on customer acquisition, giving 10 free lessons for each customer referral, effectively paying more and more to get clients who are less and less inclined or financially able to pay for its service.
That leaves two alternatives for American EdTech companies looking to scale. The first is just to forget the Bloom 2-Sigma Effect and focus on creating software that can reach as many students as possible and even learn from their mistakes to customize recommendations and instruction. In rudimentary ways, this has been the playbook for the past decade of MOOCs (Massive Online Open Courses) like Udacity and Udemy, lesson module platforms like Teachable and Thinkific, homework-helper services like Chegg and Course Hero, and most impactfully, the wonderful test prep tools of Magoosh. As Justin Reich’s Failure to Disrupt shows, software-based instruction can be quite helpful for more rote learning, particularly in STEM and standardized tests, but often fails for classes requiring critical thinking, particularly in the humanities. (Presumably, more critical thinking would be good for STEM as well.)
The second, then, is to create different tiers of service, ranging from one-on-one at the priciest level to small group instruction to large online classes. As EdTech shifts towards live learning models, we’re likely to see more and more of these tiered offerings, though they’ll always be an acknowledgement of the compromise they make alongside the rest of us: we can have high-priced quality or cheaper scale, but without a state subsidizing tutoring, it’s extremely difficult to have both.
Challenge #2: Lack of Retention Once Students Finish Courses
Let’s test our hypothesis about the difficulty of scaling direct instruction by detouring back to China, where major education platforms boast millions of students as users. As we suspected, one-on-one tutoring has far less scale than classes: VIPKid turns out to be a minor player in volume, serving a mere 700,000 students, while New Oriental, an old-fashioned brick-and-mortar tutoring franchise for classes, had 10.6 million students last year. Meanwhile, app-based platforms that operate for cheaper subscriptions of course can attract far greater numbers. Zuoyebang claims to have 50 million students per day (and around 170 million in a year), and Yuanfudao is said to have an improbable 400 million students. Even if the same students are being counted multiple times across Yuanfudao’s different services, that’s roughly the entire population of the US that uses a single Chinese tutoring service.
By comparison, Kaplan, the behemoth global tutoring service with revenue well over a billion dollars a year, based out of the US, counts 59,000 students.
We arrive at a second point: China’s size alone can’t account for that staggering difference in scale. China has about five times as many students as the US (roughly 250 million to 50 million), but New Oriental has around 200x as many students as Kaplan, probably its closest American counterpart. Even a one-on-one Chinese tutoring agency is over 10x the size of America’s biggest tutoring platform.
So there’s another factor at play here, I think: retention.
China’s firms have historically served students through a decade-plus pipeline of standardized exams and English-language study from elementary school to college. Note the multiple variables in that sentence that don’t apply to the American education system: multiple entrance exams, standardized curricula, and demand for foreign-language service. Each element enables tutoring services to scale as long as parents can afford them, ultimately achieving a kind of network effect unique to education and luxury goods: once half the population is paying for these, the other half will want to pay for them as the norm. But the fact Chinese families even can pay is an ultimate differentiator. China’s fast-growing middle class is in a position to bank on a flywheel effect of higher class-standing leading to greater financial resources, leading to investment in education in turn, and finally leading back higher class-standing—the kind of flywheel effect that makes sense to few American families in a time of inequality and a dissipating middle class.
All of those factors lead not only to massive scale but massive retention of students from kindergarten to college (though the state is starting to crack down on these services). With the exception of private and charter schools, no American educational firm traditionally retains students for so long. In fact, the situation in the states is almost the opposite: most American firms are dedicated to one-off offerings. Numerous companies exist almost solely for the ACT or SAT, learning apps typically target specific ages, live professional development platforms feature single classes, ISAs last a short amount of time, and both MOOCs and cohort-based-classes offer one course at a time.
There are exceptions. Khan Academy targets all ages, Chegg and CourseHero target high schoolers and college students for a roughly eight-year pipeline, Synthesis offers a subscription for enrichment work, Emeritus’ recent acquisition of iDTech allows it to offer coding and design enrichment from kindergarten through the workplace, and Outschool offers ongoing “classes” (really clubs on topics like “The Star Wars Resistance”) without expiration dates that make up 50% of their volume.
Note that these exceptions fall into two buckets overall: either subscription apps for students to use to help with individual assignments in school or enrichment tutoring services for subjects like coding that typically that they need to cover out of school. The game for the former companies is to achieve scale through a cheap service that can retain students for a few years at a time, either as a non-profit or by appealing to cheating. The game for these latter companies is the same as in China—to find ancillary subjects that schools don’t cover but students still need to learn in order to achieve retention for years. Still, there is a difference from their exam-prep-colleagues in China: without an intensive exam-culture, these enrichment programs are offering nice-to-have vitamins for an indulgent middle class rather than must-have painkillers that every student needs.
That leads to a corollary of bear reason #1: as an American tutoring service, you can achieve scale (like Kaplan) by targeting narrow subjects with massive demand, or you can achieve retention (like Outschool) by offering a broad range of subjects with low demand. You can have scale or retention, but unlike in China, you can’t have both.
To see why retention is so crucial for the growth of a company, let’s play with some sample scenarios with small numbers.
For example, let’s say we have an agency whose marketing draws 10 students in year one; each year it will reinvest its profits and draw 10% more students than the prior year. What’s the difference between an agency that has no retention and one that retains half of the prior year’s students over the course of a 17-year educational pipeline? It looks like this:
The agency that retained half its clients ends up with 84 students by year 17, or 14.25% annualized growth, whereas the agency that retained none ends up with 46 students, or 10% annualized growth.
But what if the agency achieves full retention of all its students from year-to-year?
By year 17, this agency has 405 students, or 26% annualized growth. It is a massive difference, even as it outwardly was only drawing 10% more students each year.
Finally, let’s play with one more variable—organic growth from word-of-mouth. What if we keep the scenario above but also assume that 10% of a previous year’s clients will successfully refer the agency to new clients as well? We get a graph like this:
The agency without any retention ends up with 185 students, the agency with half-retention ends up with 339, and the agency with full retention gets 1574 students. Those are annual growth rates of 20%, 24.63%, and 37.19% respectively.
That’s the effect of retention.
Challenge #3: The client is not the user
The user for all EdTech platforms is the student, but only for adult education is the student also the client who actually pays for the service. For traditional EdTech companies, the client would be the school; for newer EdTech companies, the client tends to be the parent.
What that means is that EdTech companies’ incentives for profit are rarely aligned with incentives for making a product that students actually like to use for learning. Take Emile, for example, a new EdTech company offering polished videos of well-coiffed professors delivering AP lectures in a classroom—in other words, a company that basically replicates a monologic class experience to Masterclass-lite videos without offering any dynamic interplay with students that might engage their attention. Left on its own, Emile might go the way of a Quibi with its short-form videos in search of a market, but Emile has another plan, as its COO put it last week: “The way this becomes a billion-dollar company is the government-funded system.” Whether students like or don’t like Emile will be irrelevant if it can convince schools to play its videos, most likely by appealing to district savings rather than improved learning. Where will these savings come from? The platform itself gives its own suggestion: after all, why can’t it effectively replace teachers by offering a deluxe, beautifully-coiffed vision of the classroom for students?
This is, alas, an investment memo, so we’ll leave aside any sanctimonious moralizing about just how awful misaligned incentives make learning for students. Instead, let’s just note that it is extremely challenging to command a marketplace without network effects that come from users actually wanting to use a product and share it with each other. Indeed, many EdTech companies survive through legacy deals with schools that last for years, but in some sense, these also limit their growth—after all, their competitors also have legacy deals with other institutions lasting for years. We end up with the ACT and SAT having deals with different states, for example, and Coursera and EdX having deals with different Ivy League institutions. Incumbents’ long-standing deals give them fortress-like protections, but these are not winner-take-all marketplaces. A marketplace where almost nobody can lose is also a marketplace where nobody can win.
To see what little traction network effects have in education, it’s helpful to consider what is likely the most successful educational network between teachers and families, ClassDojo. ClassDojo’s success undoubtedly is due to the ways that it seems to align interests between teachers, parents, and students—helping all parties connect to each other through a single platform to share work and notes. And yet, even as it builds incentives for families to connect to teachers and teachers to join to connect to students, its slight-of-hand reveals the ways that these parties are not aligned at all: in essence, it provides classroom surveillance for parents who want to watch teachers while allowing teachers to gamify works by sending high fives and other points/badges/rewards to students for their work. Each gets a different way of enacting control over the other, with the student caught in between. For that matter, it is notable that ClassDojo has been completely free for a decade and still only succeeded with 1 in 6 elementary school families in the US, all while being heavily criticized for rewarding surveillance and punitive instruction.
Again, our exceptions mostly turn out to be the usual ones: the apps like Chegg and Course Hero that offer affordable subscriptions for students to find solutions to work. These are, of course, band-aids, not learning platforms, and they can succeed as such with far lower standards to meet.
But there is another exception as well: professional development for adults that ensures the student and client are the same. There is just one issue here: it becomes difficult to retain this student for continued classes (you beat challenge #3, only to face challenge #2).
The Bull Case for EdTech?
So what would a successful EdTech company in the US look like?
One answer would be a platform for teachers to make their instruction more appealing and fun for students. This scalable platform would enhance one-on-one or small-group learning (solving challenge #1) while creating a better student experience that would lead students to return (solving challenges #2 and #3). It would also take advantage of recent shifts towards live learning, user-generated content, and empowerment of the educator-as-creator with data.
It would look, in other words, a lot like Kahoot!, the altogether brilliant learning-game platform that partners with Disney and others for teachers to devise live quizzes that give students points and teachers data on how their students are doing. Kahoot!’s recent acquisition of Clever, a learning management system used by 65% of classrooms, will likely make it the de facto digital toolkit for teachers moving forward—though it too, like most EdTech platforms, is particularly good for rote skills with right-and-wrong answers rather than any kind of nuanced critical thinking.
So what would a successful EdTech company look like that actually helped students think? There is no good answer now, though a number of recent companies are laying the foundation for the kind of live learning platform that could give students the community and accountability to learn subjects they’re actually interested in. Emeritus is likely the furthest, with partnerships with top universities for classes and curricula it can customize in small groups and one-on-one tutoring for students of all ages. Maven’s cohort-based classes might get there by creating educator stars with massive followings, creating different tiers of service, and extending its classes to all ages for a long-term pipeline. And if Curious Cardinals can command the supply of top university students as educators, it can effectively let them customize curricula for students’ needs that can be reused over time while it builds long-term relationships with schools to teach students areas they never could have learned otherwise.
And finally, there is Fiveable. Though it just offers AP classes for the moment, Fiveable has distinguished itself by 1) drawing over 100,000 students in a short amount of time by actually appealing to how they like to learn in lo-fi, crunch-time, largely improvised study sessions, and 2) developing tools for study web that show a deeper comprehension that learning is not about content mastery but having the structure and interaction to engage one’s mind. Fiveable’s dedication to creating a study community suggests it could create a whole new market for p2p learning between students that could extend well beyond test prep into interests they might never have explored together. Imagine this decentralized, Web 3.0 student-to-student platform for students to join together and learn about whatever they like, no matter how niche, and then to teach others in turn—while getting credit whenever possible and developing projects to enhance their college applications as well. That’s not what Fiveable is today. But that’s what it’s on its way to becoming, and there are few companies to be more excited about than ones like Fiveable that actually care about how students learn—and are in a position to ride massive network effects in return.
With special thanks to Yoko Li.
This was great, David.
This was a very in depth look at the EdTech business world. I would also suggest a look at what Flipgrid (acquired by Microsoft) is doing.