The AI Bubble Isn't in Your $100 Subscription: It's in the Trillion-Dollar Bet
Subsidized AI subscriptions don't prove there's a bubble. The numbers in Anthropic's S-1/A point somewhere else: compute, commitments against the future, and your dependency.
By Carlos Casanova · Language: English
In this article
- 1. The myth of the subscriber who bankrupts the lab
- 2. Three catches in that math
- 3. The gym model
- 4. The receipts: subscriptions are one-fifth
- 5. What is actually maxed out: compute
- 6. If there’s a bubble, it’s the trillion-dollar bet
- 7. The toll isn’t the fee: it’s the dependency
- Conclusion: the right question
There’s an argument you hear at every dinner-table conversation about artificial intelligence: “it’s a bubble, just look at how they give tokens away.” The long version says subscriptions are subsidized, the labs burn cash on every user, and the party ends the day they charge what it really costs. It sounds good, it has tweets to back it up, and it makes whoever says it the adult in the room.
I’ve repeated it too. Until I watched someone sit down with the 400-plus pages of Anthropic’s IPO filing and run the full math. The conclusion made me uncomfortable: there may well be a bubble, but the subsidized subscription is the wrong reason to be right. A caveat: the figures below are the ones that analysis cites from the document filed with the SEC; I have not audited the filing myself. What interests me is the reasoning, because it changes the question.
1. The myth of the subscriber who bankrupts the lab
The myth has a basis. In January 2025 Sam Altman wrote that OpenAI was losing money on Pro subscriptions. And every so often a developer shows up with a spreadsheet:
- The 28% one — burned the equivalent of $152 in API-priced tokens using just 28% of the weekly limit on a $100-a-month plan. Extrapolated to the maximum, about $2,250 a month.
- The eight-month one — paid roughly $800 in subscription fees and consumed more than $15,000 at API prices. A single month, the wildest one, was worth $5,600.
With numbers like those, anyone would conclude the flat plan is an unsustainable gift.
2. Three catches in that math
Here’s the thing — that math has three holes, and all three point the same way.
- Prices dropped — the $5,600 month was calculated at old rates. At current ones, the same usage would be closer to $1,450.
- Price is not cost — what the lab would charge you through the API is not what it costs them to serve you. Between the two sits the margin, and the margin is exactly what we are arguing about.
- Almost nobody hits the limit — the person publishing a spreadsheet is, by definition, the extreme case. The honest question is a different one: how many weeks do you max out yours?
3. The gym model
Gyms sell more memberships than they have treadmills, because they know most of us won’t show up. The labs make the same bet: that almost everyone has a life and isn’t squeezing tokens every second of the day.
Put bluntly: on a flat plan you don’t just pay for the tokens you use, you pay for the ones you don’t. The power user who extracts $15,000 from a $100 subscription doesn’t prove the model is broken; that proves there are plenty of others paying the fee to open the chat three times a week.
4. The receipts: subscriptions are one-fifth
Anecdotes aren’t evidence, so let’s go to the filing. According to the S-1/A updated with the third quarter, in the first nine months of 2026:
- Subscriptions (Pro, Max, Team) — about $6.9 billion, 21% of revenue.
- Consumption-based — about $26.3 billion, 79%.
- Total compute bill — about $18.8 billion, and that covers everything: training, research, and serving every user.
Now the sanity check. If every subscriber cost three times what they pay, subscribers alone would account for $20.7 billion: more than the company’s entire compute bill, training included. It doesn’t add up. So most subscribers are nowhere near their limit, and the real business isn’t even there. It’s in the meter that runs per token.
5. What is actually maxed out: compute
What’s at the limit isn’t your plan, it’s the hardware. In the third quarter revenue went from $11.6 billion to $17 billion, and compute spending nearly doubled to $9.9 billion. Part of that jump, around $5 billion, was capacity rented on short-term commitments, considerably more expensive than capacity booked in advance. It’s Uber surge pricing, but for data centers.
Even paying that premium, they brought in close to $1.70 for every dollar spent on compute. And this is where the nuance matters, because it’s where optimism overshoots: that is revenue, not profit. The same quarter closed with a $567 million GAAP operating loss. Money is being burned. Just not where we were told.
6. If there’s a bubble, it’s the trillion-dollar bet
The number that should really keep us up at night is another one: roughly $895 billion committed to compute over the next decade. Almost a trillion.
The loop is simple and fragile at once. Demand needs compute. Compute needs cash. And that committed cash needs demand to keep climbing, year after year, without stumbling. So far the bet is paying off: every Fortune 10 company and over half of the Fortune 500 spend at least six figures a year, and nobody signs that for a toy.
But look at how Anthropic sizes its own market: not as global software spending, but as the sum of all knowledge-worker wages, some $40 trillion. They don’t want the license budget. They want the payroll. And if the bet is about replacing salaries, someone will have to explain who does the buying afterwards. That is the subject of a recent paper, The AI Layoff Trap, and it deserves its own post.
7. The toll isn’t the fee: it’s the dependency
I pay for subscriptions and I also run models at home. This analysis didn’t make me cancel anything; it made me read the invoice differently.
- My subscription is someone else’s margin — it’s a gym membership with a good price-to-use ratio as long as I’m one of those who actually show up. It is not an acquired right: limits get adjusted when it suits the gym.
- The real risk is capacity — if compute is the scarce resource and there’s surge pricing upstream, the day demand tightens you’ll feel it first in limits and latency, not in a headline about bubbles.
- The emergency exit gets built beforehand — having an open model you can serve yourself or buy from another provider isn’t basement nostalgia; it’s not depending on a single tap. I covered it in Cold War 2.0: the value of openweight models is in the margin war.
Conclusion: the right question
Next time someone says “it’s a bubble because they give tokens away,” ask for the receipts. The subscription isn’t the weak link; it’s the best-understood part of the business. The weak link is a trillion-dollar commitment that only works if all of us keep consuming more every quarter.
If you want to keep pulling that thread — what stays scarce when producing costs almost nothing — I work through it with exercises in No more excuses, an interactive course of eight short chapters, about 40 minutes, no sign-up.
And don’t ask whether your $100 plan is sustainable. Ask something else:
If your AI provider raised the price or lowered the limit tomorrow, how much of your work would stop?
Idea and starting point: Avisha (@distilled_ai) — is the AI bubble about to burst?. The figures come from that reading of Anthropic’s S-1/A (filed with the SEC on November 2, 2026) and the sources cited in the video: botfarm.run, ksred.com, Anthropic’s pricing page, and Falk and Tsoukalas, The AI Layoff Trap. The dependency and sovereignty angle is mine.