This article has been in my queue for about a month now, but every time I sit down to work on it something else has blown up in AI Land. So here I am, finally getting something written.

This started out as an article about OpenAI hiring Peter Steinberger, the developer who vibe-coded OpenClaw (but OpenAI did not acquire OpenClaw itself). My take at the time was that OpenAI was making two plays: first, it was buying into the massive, worldwide hype fest surrounding OpenClaw¹. But the other, arguably more important play is related to token usage. Agents use a metric shit-ton more tokens than your standard civilian-grade chatbot, and by getting users to use more tokens could mean more revenue.

Of course, I also believe that OpenAI has completely lost the plot–I mean, they bought a podcast for chrissakes.² Nevertheless, the token story remains, and has exploded over the past few months.

It's the Tokens, Stupid

An article in the Times summed it up pretty well: a typical 7,500-word student essay might require about 10,000 tokens over several hours of chatbot usage. By comparison, one developer cited in the article estimated that he uses more than a billion tokens per week.³

As it turns out, this is the big story with the sudden explosion of AI agents and especially "Claws." Dozens of companies have joined the gold rush by attempting to build their own OpenClaw analogues, including Nvidia, Alibaba, Anthropic, Tencent, ByteDance and Xiaomi, driving token usage skyward. A combination of ArkClaw and Seedance usage recently pushed Bytedance's Volcano Engine past 100 trillion tokens per day.⁴

But there's a catch: token pricing is nowhere near reality. Most companies offer subscription plans for around $20 per month, with power user options up to $200 per month.

These subscriptions don't even come close to covering the real cost of providing the service, though. One industrious Claude user concluded that the $20 Pro plan actually allows $163 in token usage, while the top-end $200 Max plan offers the equivalent of $2,700 in API cost.⁵

To make things even worse, it seems as though almost no one is paying to use LLMs, relying instead on free options.⁶ It's no wonder these companies are hemorrhaging cash.

The Economics Still Don't Work

The truth is that the economics of AI just don't work. Traditional products and services benefit from scale. If you're making pencils, there is a big up-front cost to build the factory, hire people, buy materials and so on. But over time your costs come down: the factory is paid off and you can get a better price on your materials as you buy them in larger quantities (scale works for your suppliers, too).

AI is the literal definition of "we lose money on every transaction but make it up in volume."

Same goes for software. Once you've written your office suite or CRM or whatever, the incremental costs for maintenance drop over time. Sure, hosting costs may grow, but SaaS products tend to be high-margin affairs.

AI, on the other hand, is the literal definition of "we lose money on every transaction but make it up in volume." Training costs are largely fixed, and more capable models cost more money to train. GPT-4 reportedly cost more than $100 million for training alone.

Running AI models isn't cheap, either. Data centers cost money, high-end chips cost money, electricity costs money, and the more people use your product the more of those you need.

Training is expensive, usage is expensive, and while there are some opportunities to lower costs, such as relying on smaller models, there are limits. The harsh reality is that the economies of scale don't work well for AI purveyors.

Mountains of Debt, No Profit in Sight

OpenAI claims it will be profitable by 2030 and Anthropic thinks it can get there in 2029. Assuming that's possible–a big assumption–that's still 3-4 years away and these companies are burning money at an insane rate.

You might ask, "where is all this money coming from?" Some of it is from what I'd call "standard deals" like banks, venture capital, or corporate bonds.

But there are other sources of "funding," including some creative circular deals among AI players including Oracle, OpenAI, and Nvidia. Last September, Nvidia announced that it would invest $100 billion in OpenAI, which OpenAI would then use to buy Nvidia chips. In other words, Nvidia was essentially funding its own sales.

That deal collapsed in February, but it's hardly the only one. The Stargate deal saw OpenAI promising to buy $300 billion in cloud computing capacity from Oracle over five years. Oracle, in turn, committed to buying $40 billion in Nvidia hardware for the data centers it plans to build for OpenAI.

Seems a little shady, no?

But Wait! That's Not All!

There is another funding source, and that is the private credit market. You're probably familiar with private equity, but private credit is different. Normally, private equity deals involve borrowing from banks, who then resell the debt to investors. In a private credit deal, however, the private equity companies borrow money from each other.⁷

You might think that this is risky and kind of insane, and you'd be right. (You might also be vaguely reminded of the shenanigans leading up to the 2008 crash and you'd also be right.) These private credit funds draw investors, but typically limit withdrawals in order to remain solvent. But the private debt industry is looking shaky, with redemption requests for many funds exceeding 10 percent in recent months. But because redemptions are capped–limits of five percent per quarter are common–investors are starting to get mighty nervous about getting their money out.

While most of the big players aren't funded through private credit, some of the companies they rely on are, such as data center company CoreWeave. Private credit company Blue Owl recently failed to secure $4 billion to fund a CoreWeave data center in Pennsylvania, which puts Blue Owl in a very tight spot. Either Blue Owl has to come up with the money on its own or the project goes away. (I'm sure the local residents would be pleased with the latter result.)

There are rumblings of an impending private credit crisis, and if that happens then things could get very bad for the AI industry.

Companies Look to the Enterprise

Some AI vendors are beginning to figure out that they can't rely on $20 subscriptions that no one pays for, and the $200 plans are even worse for profitability. Companies are beginning to modify their strategies along a couple of major lines: restricting model access and increasing token pricing.

Anthropic Drops the Hammer

Perhaps one of the bigger stories of the past couple weeks was the announcement that Anthropic would begin forcing users of third-party harnesses such as OpenClaw to use its pay-as you-go plan instead of the standard subscription.

While this pissed off a bunch of Lobster bros, it makes perfect business sense. The platform was not built for a 100x (or more) increase in usage, and it's just not viable to allow virtually unlimited usage for $200 a month. In fact, the announcement email said that very thing: "We’ve been working to manage demand across the board, but these tools put an outsized strain on our systems."

Alibaba Restructures

It's instructive to look at Alibaba, which has completely restructured its AI business over the past month or so. First, it launched Alibaba Token Hub (ATH), followed shortly after by Wukong, which it hails as an "AI-Native Agentic Platform for Enterprise." Stripping out the buzzwords boils down to "OpenClaw but integrated into our platform and without massive security holes."

Alibaba also named Li Feifei as its new Cloud CTO. Li isn't a data scientist or AI researcher; he is an infrastructure specialist with a track record of building reliable, scalable data platforms. This seems to indicate that Alibaba is shifting its focus toward offering a reliable, useful platform for businesses.

Finally, the new top-tier Qwen3.6-Plus model is only available via API for paying customers. It looks as though the subsidies may be coming to an end.

Meanwhile, AI companies are also beginning to increase prices. Zhipu, Tencent, and Alibaba have begun raising prices, so expect this trend to continue.

AI Development is About to Get A Lot More Expensive

We're about to see what of AI-assisted software development really costs and whether the improvements in productivity (if any) are worth the trade off. I think this will begin to affect your average vibe coder as well. Expect to see more companies following the lead of Anthropic and Alibaba by strongly incentivizing the use of their own tooling over third-party options, while jacking up prices.

It's hard to justify $600 to make a few dropdowns work in a prototype.

This doesn't just apply to first-party providers, either. Figma recently began charging to use some its AI tools notably Figma Make. While it's unclear how their credit system maps to tokens the prices are exorbitant.

I have a colleague who worked on some fairly basic interactions and prototype flows in one of her designs. She used a little more than 28,000 credits, which would have cost between $600 and $750 depending on the plan. For one small set of interactions. Fortunately, this was during Figma's "try it for free" period, but it's hard to justify $600 to make a few dropdowns work in a prototype.

Though the industry is shifting toward the enterprise, I think we'll see the $20 basic subs stick around, though prices may increase a bit. However, we may also see increasing restrictions on model usage, as companies push free or low-tier users to smaller, cheaper models. This, in turn, could lead to user dissatisfaction, resulting in fewer subscribers. And to be honest, shedding the low-end users might not be the worst business plan.

Naturally, ads will also stick around because Silicon Valley seems incapable of finding any other way to make money, but there is no way in hell that running ads in AI chats are going to come anywhere near paying the bills.

As I was editing this post, OpenAI reported that it had made $100 million in the first two months of its ad pilot program, so evidently someone is paying for ChatGPT ads. But I'll also point out two things: first, digital advertising has always been and will always be a whirling bullshit machine full of fraud and lies; and second, OpenAI tends to be very creative with how it comes up with financial numbers. Note also that OpenAI killed off their shopping initiative, presumably because they discovered that building an e-commerce platform is a bit more difficult than slapping a "buy" button on a chatbot.

Mitigation Strategies

I've looked at OpenClaw's context management up close — it's bad.

So restricting model access and raising prices are two emerging strategies as AI companies begin to search for a path to profitability. But there are other steps that can help as well.

Luo Fuli, the head of Xiaomi's MiMo had this to say on social media:

"Two days ago, Anthropic cut off third-party harnesses from using Claude subscriptions — not surprising. Three days ago, MiMo launched its Token Plan — a design I spent real time on, and what I believe is a serious attempt at getting compute allocation and agent harness development right. Putting these two things together, some thoughts:

1. Claude Code's subscription is a beautifully designed system for balanced compute allocation. My guess — it doesn't make money, possibly bleeds it, unless their API margins are 10-20x, which I doubt. I can't rigorously calculate the losses from third-party harnesses plugging in, but I've looked at OpenClaw's context management up close — it's bad. Within a single user query, it fires off rounds of low-value tool calls as separate API requests, each carrying a long context window (often >100K tokens) — wasteful even with cache hits, and in extreme cases driving up cache miss rates for other queries. The actual request count per query ends up several times higher than Claude Code's own framework. Translated to API pricing, the real cost is probably tens of times the subscription price. That's not a gap — that's a crater.
2. Third-party harnesses like OpenClaw/OpenCode can still call Claude via API — they just can't ride on subscriptions anymore. Short term, these agent users will feel the pain, costs jumping easily tens of times. But that pressure is exactly what pushes these harnesses to improve context management, maximize prompt cache hit rates to reuse processed context, cut wasteful token burn. Pain eventually converts to engineering discipline."

Companies must do a better job with their tooling. As previously noted, OpenClaw is an utter train wreck: it's insecure and evidently impressively inefficient. (This also makes me a bit sus about the quality of AI-generated code.)

This may be another area where vendors can gain an advantage. By building super-efficient tooling they can help reduce costs for their customers while decreasing the burden on their infrastructure.

However, this makes me worry a bit about platform lock-in. Tech executives love few things more than locking in their customers, but that's obviously bad for the customers and usually leads to enshittified products. It will be important to ensure that third-party tools continue to have access to vendor models without extra penalties.

But either way, better-quality tool sets that efficiently manage resources are going to be critical to anything resembling profitability. Half-assed, vibe-coded harnesses won't cut it.

Is the Quality There?

The move to enterprise agents is neat and all, but aside from economics I wonder how widespread LLM usage going to work out in the real world. I've used several models for certain tasks including researching for some of my side projects and I find that it can be pretty useful, at least in my limited use cases. (Note that I never use AI in my writing workflow, including research, editing, etc. See my policy page for details.)⁸

Even if the economics do somehow begin to work out in the future and companies build reliable, efficient platforms and tooling, there is still one underlying issue, and that's model quality.

Models still have an unfortunate tendency to just make shit up. I've had it happen to me and AI agent horror stories abound, including the guy who deleted all of his wife's photos or the business owner who trashed his company's database⁹. And both Meta and Amazon have had some nasty experiences with internal agents wreaking havoc.

AI cheerleaders will say, "but, but, the models are always improving, just wait!" But we all know that infinite improvement is not possible and neither is eradicating LLM "hallucinations."¹⁰

The AI honeymoon is coming to an end, and big changes are afoot. Tokens are king and companies are trying to figure out how to make real money with AI. It's not going to be easy, and eventually the true costs of AI usage will become apparent.

This, combined with external forces such as the concerning private debt industry seem to point to some shock waves ahead. I expect to seem some consolidation over the next year or so, and I strongly suspect that at least one big player will make its exit.

As AI begins to come into its own, we'll see where the real value lies and what the true cost benefit is.

1. https://hellochinatech.com/p/openclaw-china-ai-stack
   ↩︎
2. https://atomicbox.studio/openai-just-bought-a-propaganda-division ↩︎
3. https://www.nytimes.com/2026/03/20/technology/tokenmaxxing-ai-agents.html
   ↩︎
4. https://www.jiemian.com/article/14203212.html
   ↩︎
5. https://she-llac.com/claude-limits
   ↩︎
6. https://institute.bankofamerica.com/economic-insights/consumer-ai-usage.html
   ↩︎
7. https://www.economist.com/briefing/2026/04/01/a-guide-to-the-private-credit-crisis
   ↩︎
8. https://atomicbox.studio/policies
   ↩︎
9. https://futurism.com/ai-vibe-code-deletes-company-database
   ↩︎
10. https://theconversation.com/why-openais-solution-to-ai-hallucinations-would-kill-chatgpt-tomorrow-265107 ↩︎