Ed Zitron on AI bubbles, government subsidies, and more

Is AI a bubble? Is the government going to bail out Silicon Valley? What’s a reasonable expectation for AI ROI?

IT pros know many people with many different answers to these questions—and more.

One of the more prominent voices warning that AI is a bubble ready to collapse is Ed Zitron, CEO of EZ Primary Research. Zitron also writes the newsletter “Where’s Your Ed At” and hosts the Better Offline podcast.

We sat down with Zitron to talk about the business and future of the technology—as well as what we mean when we talk about AI.

(This is Part 1 of our interview. Part 2 is here.)

This interview has been edited and condensed.

Are we still in the AI bubble? How would you describe that at this moment?

We are very much in the desperation phase of the AI bubble, because even with incremental improvements to models, no one can measure the actual ROI of AI. The actual capabilities—forget benchmarks, which are rigged for the companies to literally train the models to perform on—the actual things that these models can do are basically the same as they were a year ago. They might be better at coding, but they are still mindless, thoughtless machines churning out code with no intention, and they have categorically failed to expand new industries. They just haven’t.

We’ve seen startups like Harvey pop up to try and burp out law-style software. It isn’t really working for anyone other than those related to coding platforms, and those coding platforms also all lose money, so that’s one problem. The other is that OpenAI and Anthropic, a few months ago, moved all of their enterprise customers to something called token-based billing, so when a regular user of, say, the $20-, $100-, $200-a-month Claude subscription does start on the subscription, they’re not actually paying the real costs when they burn tokens. SemiAnalysis just revealed that you can burn up to $8,000 worth of tokens on those subscriptions [in] the $200 range. People have just been using these things willy-nilly, because they could burn way more tokens than the $200 would otherwise buy them.

Everyone’s got used to this all-you-can-eat model, but enterprises back in April, or maybe even March, got moved to token-based billing, meaning they’re paying the actual cost of their AI services. This has led to a massive anxiety throughout the economy of anyone using AI, because suddenly businesses went from paying a couple hundred dollars a head to paying $1,500, $2,000, or $3,000 per person using these services.

At this point, suddenly, one flipped from, “Wow, AI’s so good,” to, “Can we measure the ROI of this? What is the ROI of AI? Is there ROI?” And the answer is: not really. Jeetu Patel, who is the president and CPO of Cisco, said that the costs are way higher than the actual return on investment. That’s probably because no one can actually measure the ROI at all.

Is this a sign of a tightening of this market, or is this an indication that these companies are maybe doing this in hopes that something like a federal subsidization of the industry or some sort of deal will come down, whether it’s private or public sector, probably public, to help them defray this cost for users?

One thing that I live my life by is never try and bring a conspiracy out of things, never try and assume that there’s a big plan. I don’t think anyone has a plan. I think what the actual plan for OpenAI and Anthropic was is very simple. They thought that at the beginning of 2026, they would massively increase their revenues by charging people the actual cost of AI. This happened in the short term.

What they didn’t calculate for was that their AI services cost an absolute crap-ton, which meant that their customers, who had pretty much just been told to do as much AI as possible and told their employees to act in that manner, they then immediately, within less than three months, ended up freaking out and saying, “We’re going to cut our spending.”

I don’t think anyone is sitting there thinking, “Oh, we’re going to get a government subsidy for our AI services,” because I don’t think the government is going to do that. I don’t think the government is going to subsidize very much of anything. Perhaps the only thing we could look forward to in the future, and I mean that sarcastically, is an Oracle bailout, potentially.

But I don’t think there is a government subsidy coming for enterprise companies paying Anthropic and OpenAI. I don’t think that’s happening. Both of those companies are already discussing price cuts, which is a sign that what happened was people don’t like the price of it. People think it costs too much, but due to [the] unit economics of these companies, their businesses already don’t make sense and already don’t work. A price cut would only increase their costs and very likely lower their revenues, too.

This stuff has been integrated into the infrastructure of the federal government and is continuing to be integrated into it. The companies could say, “Hey, this is a national security interest, because we need to continue to be functional, so that we can provide this infrastructural technology to you.”

I don’t think it’s that intertwined with the federal government. Large language models don’t do that much. Project Maven is tied to the US government. The Promise software was tied to the US government. Large language models, Anthropic provides those services to the US government, Google Gemini does, OpenAI does, Grok does, but as far as them being integral, they actually have to do stuff, they actually have to do a thingy. There has to be an outcome.

I don’t think that these things are essential in any way that couldn’t be replaced with an open source model, but even then, I don’t even think they’re essential in that manner, either. This whole thing with Anthropic right now is just politicking. Anthropic pissed off the US government and now this government is pissing around with Anthropic. I just don’t see where the actual national security thing would be.

It’s things like facial recognition, ICE processing acceleration, that kind of application.

That isn’t large language bubbles, that’s a different kind of tech.

Well, yeah, but I’m talking about AI as a whole, I’m talking about these technologies.

Not to push back too hard—

No, no, no. Push back as hard as you want, please.

Well, you’re playing the marketing game for them.

Sure. And how’s that?

Because that’s why they called it AI, so that you would conflate everything together. Facial recognition technology used in Minnesota was not large language models. Even if they were using an LLM to look over those, that is not a specialist thing, they could use any kind of model they wanted, that’s not special to Anthropic or OpenAI.

Even then, I’m not convinced that this stuff is remotely essential to anyone. I think it’s only essential as a symbolic thing to help raise capital and to help pump stocks.

Let me try and tease this out a little bit, because I think this is really interesting. I understand that this is marketing, but it’s also the way that people in the industry refer to it. But what you’re saying is that you are specifically talking about large language models when you say AI.

Yes, and when everyone talks about Nvidia, when everyone talks about SpaceX, when everyone talks about OpenAI, Anthropic, when everyone talks about GPUs and data centers, that is what they are talking about, too. They want you to believe that useful, interesting stuff might come out of it, that there might be autonomous cars or cancer research, or all manner of other stuff. And I must be clear, that is not what those data centers are for. Those data centers support AI, as in large language models.

I think it’s just a question of terminology, but don’t you think that they have, at least for now, won that side of the messaging game?

How have they won it?

Well, because that’s how it’s referred to in the media, in the way that people talk about it within the industry.

Well, that’s where the job of the journalist comes in. But it also doesn’t change anything I’ve said. As far as the government subsidizing people’s AI use, I don’t think that that’s something that would happen, or indeed would be anything approaching a realistic path. I don’t think that conflating those things makes sense.

Just to be clear, every data center could disappear tomorrow, every single one—well, all of the AI data centers, I mean, because data centers have been around for much longer than AI—and all that facial recognition stuff would still be there. All of it would still exist. All of it could be developed and grown from there, it’s not dependent on that at all.

Head here for Part 2.