When The Bill Comes Due

I think the point where it became clear to me that the AI bubble was hitting a wall came about two weeks ago, when Anthropic launched its Claude Design product.

As someone who is interested in design and is trying to understand what the hell is happening with AI enough to have a thoughtful perspective on it, it struck my interest. I threw it a GitHub project and told it to extract the visual style. Then I ran a second command, and … suddenly, it was out of credits until Friday at noon. It was 1pm on a Friday.

I prefer to actually understand the things I feel skeptical about, because it helps me catch things that highlight bigger underlying problems. And well, it doesn’t get any more blunt than that.

Expect more stories like this. Recently, GitHub announced that it would change its billing for Copilot to usage-based pricing. And there have been mixed messages around how people could use their paid Claude subscriptions for tools like the overly hyped OpenClaw—at first, it was banned, but now it’s not.

The truth is, these mainstream AI products are getting subsidized, and users are addicted to the subsidy. It’s not sustainable, but the goal is to keep you addicted long enough that maybe it will be. Admittedly, some, like noted AI skeptic Ed Zitron, don’t think it ever will be:

It’s very, very important that nobody writing about AI in the mainstream media actually understands how much these services cost, and that any mainstream articles written about services like ChatGPT or Claude Code are written by people who have little or no idea how much each individual task might cost a user.

Remember: generative AI services are, for the most part, experimental products that do not function like any other modern software or hardware. One cannot just walk up to ChatGPT or Claude and start asking it to do work.

And when you become dependent on these tools, you pay more, and then you fall into this trap of paying huge chunks of your revenue just to keep these things running.

It’s costly, it’s inefficient, and we’re going to run into scale limits eventually. All the money for those Super Bowl ads and new products is going to eventually come due.

But then there’s DeepSeek and its ilk.

AI For Cheapskates: Yes, It Exists

To be clear, I’m not saying any AI model is going to solve the inherent structural problems that AI creates. Much the opposite. However, I think that DeepSeek’s story is an important contrast to the massive players that seem to dominate every news headline these days.

The Chinese company, whose access to the highest-end GPUs is limited by trade restrictions, has been building models on a budget, then charging for those same models at rates that put dominant players to shame. Its newest ones came out just last week.

DeepSeek V4 Flash, which is roughly as powerful as Claude’s Sonnet 4.6 model, costs less to use than the low-end Claude 3 Haiku, a deprecated model which came out more than two years ago and is barely functional in comparison. It isn’t as slick, but it has some advantages, notably a very large context window (a weak point of prior DeepSeek models). Plus, you’re not paying the overhead for all the misadventures that come with Anthropic or OpenAI’s models.

On top of all that, since it’s an open-weight model, you’re not stuck using DeepSeek’s servers. The pay-as-you-go cloud service Novita, for example, charges roughly the same rates as DeepSeek does for its new models. (Good luck self-hosting these though, as you’d need a lot of GPUs to do so.)

When DeepSeek first emerged with its R1 model a year ago—built on the cheap and punching above its weight—its emergence was so shocking that it caused the global stock market to shake. This new model did not do that, which has led outlets such as Reuters and The Economist to suggest it might be a failure.

But if anything, the reason it feels like a failure to some has less to do with its innovations and more with the fact that others have followed in its footsteps. MiniMax, for example, has a “self-improving” model comparable to DeepSeek’s, one that actually learns from its mistakes, apparently. And other players like Z.ai and Alibaba Cloud’s Qwen have drawn excitement from self-hosters who want to avoid paying Anthropic or OpenAI any money at all.

They’re Aiming For The Big Bill. You’re Not.

I think the point I want to get across is that AI is not a two-headed hydra. There are other players, but the biggest ones are incentivized to suck up all the oxygen. So you might be forgiven if you think that only two or three companies are building AI models.

These smaller players, who are more likely to open-source their innovations, have to compete on efficiency and technical prowess, not with unlimited resources. That makes them an effective counterweight to the attention-grabbers.

But the big players are always ready to throw their weight around—for example, with constant feature drops. Recently Anthropic announced integrations with numerous major creative tools, including Affinity and Creative Cloud. (And, controversially, Blender.)

I’ve often joked that, in the wrong setting, AI can be a decorative bird, but I don’t think that’s what’s happening here. I think, instead, Anthropic has identified a use case for its technology that is useful enough, and sucks up enough tokens, that companies will be ready to put down some big bucks for its models. And you thought Creative Cloud was expensive.

I think so many companies are going to be sucked in by the efficiency benefits of these tertiary tools, only to find that they’re spending giant amounts of money for modest efficiency gains.

It won’t do much for all the other ethical concerns about this stuff (giant data centers remain a big risk, for one), but a bit of literacy about what all this stuff does will go a long way. After all, modest efficiency gains will be a lot easier to swallow if the amounts of money aren’t giant.

A lot of companies want to optimize with AI, but they’re looking at efficiency, not cost. Which is a shame, because I think these open-weight Chinese models will ultimately be like open-source. Lots of people will ignore them, because they don’t advertise, don’t have sales teams, and don’t build random buzzy-but-expensive things like Sora. But eventually, a cult will form of people who figure out that spending tons of money on bleeding-edge LLMs, when there are cheaper options (including some that can be hosted on your laptop), is not a good investment.

(Plus, dollars to donuts you’ll eventually see an open-source recreation of this design functionality, just as we saw with Claude Code competitors like OpenCode.)

As I wrote recently, humans are often the problem with AI. But they can also be the solution—by being savvy about how they deploy it, by not letting flashy features distract you from your goals.

After all, the bill is going to eventually come due. And you’re gonna want a smaller bill. Trust me on this one.