I’m of two minds on AI as EdTech. On one hand, I’m as addicted to and fascinated by the magic text box as anyone I know. I use it for hours a day. I study it. I absolutely believe this is a civilizationally significant technology. On the other hand, I’ve lived through enough EdTech hype cycles to know how they go. The product doesn’t have to be bad for it to fail to be transformative. Well-designed “personalized learning” courseware does produce results in meaningful contexts. MOOCs are genuinely helpful for some learners. But education is a minefield of hidden constraints and contextual differences for product development. I defy you to name five EdTech software product categories that are more than niches.
“AI” can mean either a product or a technology. As a technology, its importance is undeniable. As a product, it’s economically broken in ways that are particularly unsustainable for EdTech. Even the consumer AI you may be using for a fixed monthly fee is subsidized. It costs more than you pay for it. Sometimes wildly more. That works as long as venture capitalists and other investors are willing to tolerate losses for an indefinite period in exchange for winning some race they think they are in. But if inflation bites deeper into the real economy and spending slows down—which is plausible, given how much longer it will take the oil shock of the Iran war to clear the system than is widely appreciated—you may start experiencing more of the usage caps that Claude users have complained about. Or worse; you may be charged for actual usage. We know what that will look like because it’s happening in businesses that pay for usage right now.
This post is a marker of sorts for e-Literate. The AI trends are becoming clear enough that I feel like I have something to say about them again. Before we can think clearly about how AI will affect education and EdTech, we have to clear out the hype and the noise.
Today I start here: Current unit economics for AI are unsustainable for education and will not work for the integration path that SaaS-based EdTech is on. Full stop. If you’re focused on advancing frontier models or multi-agentic systems, you’re looking in the wrong places for the broad impact of AI on education over the next four or five years.
I’m actually feeling pretty good about the pace at which AI will be sustainably adopted and the shape it will probably ultimately take. I’ll have more to say about that in future posts. But for today, I will focus on what can’t happen. Educational institutions cannot buy products that they cannot afford, and EdTech vendors cannot build products that they have to subsidize. Costs will have to change. Many product companies and schools alike that are making assumptions that tomorow’s AI is going to be like today only more powerful are moving in the wrong direction.
AI labs can’t afford fixed cost
OpenAI is not projected to become profitable for another three or four years. Anthropic is in better shape. They are projected to become profitable this year. But here’s the thing: Between 80% and 85% of Anthropic’s revenues come from their API business, not their (infamously usage-limited) desktop subscription usage. (OpenAI gets about 70%-75% of their revenue from APIs.) APIs, in contrast to the subscription, is pay for usage.
And that’s where the problem is.
Check out these opening two paragraphs from an article at The Verge:
Uber blew through its entire 2026 AI coding budget by April. Microsoft revoked its developers’ Claude Code licences six months after enabling them. One company reportedly ran up a $500 million Claude bill in a single month after forgetting to set usage limits. A Priceline employee told TechCrunch that a routine Cursor contract renewal came back four to five times more expensive.
The pattern is the same everywhere. Per-token prices have collapsed, but the push for autonomous AI agents has sent consumption through the roof. Companies that gorged themselves on all-you-can-eat subscriptions in early 2025 are now scrambling to understand where the money went, and whether any of it produced a return.
Did you catch that? One company which, for some mysterious reason, apparently prefers not to be named, blew half a billion dollars on AI usage in a single month. And nobody knows if the money spent was worth it. Notice also the use of the word “agent.” The next time you see the word “agentic” on LinkedIn, which will be the next time you look at LinkedIn, think “cost.”
Now watch what happens in the next section of the article:
“Six months ago, I would have a conversation with a customer and it would be all about ‘What can it do? Is it good enough?’” Alexander Embiricos, OpenAI’s head of enterprise, told TechCrunch. “Now the conversations are about, ‘We’re spending so much. What visibility do you have? What token controls do you have?’”
J.R. Storment, executive director of the FinOps Foundation, described the shift bluntly. “In April and May, I started hearing from companies: ‘Oh my god, we are 3x over our entire 2026 token budget and it’s only April.’ The whole conversation shifted from tokenmaxxing and ‘go fast’ to ‘we need guardrails, how do we control this?’”
Priceline’s senior director of IT finance, Chris Reed, drew a comparison to the telecom billing era. “It’s like the crack-cocaine epidemic. They let you try it to get you hooked, and now you’re kind of beholden to it.” The company has begun placing token limits on certain groups. Reed said he is already seeing discrepancies between vendor-reported usage and Priceline’s internal data.
“Tokenmaxxing.” Dear Lord, please save the AI tech bros from themselves, and please let this term mark peak AI tech bro.
How widespread is this “tokenmaxxing” problem? Here’s what Wired has to say on that:
Roughly 300 companies addressed questions or concerns about AI tokens during their earnings calls or in public discussions with financial analysts in April or May, according to a WIRED review of transcripts from the data provider AlphaStreet. That’s a small fraction of the thousands of calls held during the span, but just 93 companies mentioned “token” in April and May a year ago.
Executives at several companies said they are developing or looking to buy systems to help monitor token usage and choose the lowest-priced model for a given prompt. Others said they were still trying to figure out balancing hiring more people and increasing their budgets for tokens to achieve their goals.
Software has rarely come cheap, but the latest generation of AI tools is causing unusual stress in C-suites for a variety of reasons. Prices keep fluctuating. New models that are more powerful—and more expensive—than the last get released every month. And getting entire organizations on board with new ways of working has been a challenge, so AI-fueled productivity gains on one team can lead to bottlenecks for another.
“Epidemic” is the right word. The problem has gotten so out of hand that the widely corporate-supported Linux Foundation just spun up something they call the Tokenomics Foundation. (So much for peak AI tech bro.)
To be fair, this isn’t the universal story. The frame of the Wired piece is a company that feels their investment in AI is worth it. And I’ve seen productivity gains within 1EdTech. There are real success stories. Unfortunately, these are anecdotes standing against a substantial body of evidence. Even some of the largest companies on earth are struggling with “tokenomics,” because the AI race has driven toward function gain, assuming that the benefits would far outweigh the costs. But in the past six months, the financial math has become abundantly clear:
(Seemingly Useful + Unbounded Cost) x Pressure-to-Maxx = Financial Disaster Magnitude
I’m hammering on the absurdity because it’s important to see the hype clearly (and because Schadenfreude). But for education, the real killer is unbounded cost.
Today’s AI economics break EdTech SaaS
I’m old enough to remember when a tiny startup called Instructure claimed that they would beat the LMS market because they were the first cloud-native LMS. I didn’t get it for a while, but they were right. Universities were not good at reliably running what was becoming a mission-critical, can’t-ever-be-down teaching and learning application. The new deal Instructure offered was, “We’ll run your LMS as a service, you’ll pay a fixed price, and you’ll never have to be responsible for keeping it running. We will.”
It turned out to be a good deal that hinged on one assumption: Fixed cost. Schools and colleges function on annual budgets, and they can’t run losses. Variable cost is therefore a deal-killer. SaaS EdTech providers absorb variable costs of network bandwidth, storage, and compute. Even managing that much variability is challenging. Today’s AI is wildly more variable, wildly more expensive, and is both of those things while being subsidized by an industry with uncertain ability to continue to absorb the extra cost. AI delivered via metered breaks fixed cost, which means it breaks institutional budgets. And any EdTech SaaS product is either paying a variable API cost or carrying it themselves by running a model on one of the big cloud providers. Self-hosted Chinese models may make the cost problem a little better (but not enough), while multi-agentic systems—Drink!—will make it much worse.
Let me be clear about what this means: You will not be seeing unconstrained AI in your EdTech platforms any time soon. Or worse, you will see it and then have it taken away, since some platforms have already committed to the financially unsustainable path. There will likely be some AI use that is constrained by the user interface. Instead of the magic text box, you may have buttons or pull-down menus that trigger the AI to perform a predefined course. And you may get dialogue boxes that say something like, “Maximum Size Exceeded: This course is too large to perform the operation you selected.” Consumer-grade stand-alone apps will continue to exist and be used. Some full-on EdTech AI apps will exist for high-value problems that can have more knowable and controllable costs (like defined administrative batch tasks). Next year’s ASU+GSV conference will still have AI as a (substantial) sideshow. This has nothing to do with the ultimate value of AI and everything to do with how new technologies get absorbed into a real economy. A revolution might be coming to education, but it is not here yet.
This is not necessarily a bad thing. The pace at which AI has been penetrating into education has been stressful at best and deeply concerning at worst. In future posts, I will share some thoughts about more viable and healthy paths for AI to move deeper into education, and how a slowing of the pace means we’ll be less likely to build the wrong thing. In the meantime, if you have a tokenmaxxing addiction or know somebody who does, please seek help immediately.
Join the Conversation