The Argument in Plain Terms
Ed Zitron has been saying for two years that the AI industry is built on a foundation of investor credulity, not product economics. His latest case against OpenAI and Anthropic is his most specific. He names numbers. He names patterns. And the pattern he's describing is not subtle.
OpenAI has never turned a profit. Not a quarterly profit, not an annual one. The company projects $12 billion in annualized revenue while spending more than that on operations, infrastructure, and compensation. The gap is not closing. It is widening with each new model generation that requires more compute to train and serve.
Zitron's word for this is a con. Not incompetence. Not a timing problem. A con.
That characterization will annoy a lot of people who use and value these products. That reaction does not make it wrong. The distinction between "this product is useful" and "this company's economics are viable" is doing a lot of work in this debate, and Zitron's contribution is forcing them apart.
The Unit Economics Problem
Here is the specific number that anchors his argument: ChatGPT costs more to run per query than OpenAI charges for it. The product is sold below its cost of production. The difference is covered by Microsoft's infrastructure subsidy, baked into a deal that values OpenAI as though it will someday generate returns commensurate with its $80-plus billion valuation.
That day has not arrived. The path to it is not obvious.
Microsoft is not doing this out of altruism. Azure gets a massive marketing asset and a first claim on OpenAI's most commercially valuable outputs. But the arrangement means OpenAI's reported revenue figures are partly a function of a related-party relationship, not pure market demand. Zitron's point: if you strip out the subsidy, the business looks different.
The talent side compounds this. Researchers with $10 million-plus compensation packages are not unusual at the frontier labs. Some of these researchers are working on problems that may take a decade to produce anything commercially shippable. The cash burn is structural, not transitional. And unlike a startup that is burning cash to acquire customers, the AI labs are burning cash to maintain technical parity in a race where the finish line keeps moving.
The market has accepted this, so far. But market acceptance of a burn rate is not the same as market validation of a business model. Zitron's argument is that the distinction is being deliberately obscured.
The AGI Goalpost Problem
OpenAI's fundraising has always been tied to AGI. The company was founded on the premise that AGI is both imminent and worth building carefully. When AGI doesn't arrive on schedule, the definition shifts. "We're closer than ever." "This next model is qualitatively different." "AGI by 2025." Then 2026. Then "we'll know it when we see it."
Zitron reads this as a feature, not a bug. An always-approaching finish line justifies continued investment without requiring the company to demonstrate it has crossed one.
The mechanism: if AGI arrives, the investment was obviously justified. If it doesn't, the goalposts move and the pitch resets. Investors who got in early are already up on paper. The ones writing checks at current valuations are betting on an outcome that keeps not materializing on schedule, underwritten by a definition that keeps shifting.
The secondary effect of this dynamic is that anyone who questions the timeline or the definition gets positioned as someone who doesn't understand what's coming. The AGI framing inoculates the pitch against scrutiny. Skepticism gets recast as a failure of imagination rather than reasonable due diligence.
This is the structure of the argument. It is not about whether large language models are useful. It is about whether the companies building them can ever produce returns that justify their current valuations, and whether the AGI framing serves to postpone that question indefinitely.
Anthropic and the Safety Cosplay Charge
Zitron's critique of Anthropic is sharper and, in some ways, more interesting. He calls it "safety cosplay." The company's entire public identity is built around being the responsible lab, the one that takes existential risk seriously. Dario Amodei has given speeches about the potential for AI to cause catastrophic harm.
And they keep shipping faster models.
Zitron's logic here is simple: if you genuinely believe you are building something that could cause catastrophic harm, the sincere response is to stop building it, or at minimum to slow down. Anthropic does neither. It raises money, hires researchers, and ships Claude updates on an aggressive schedule, the same as every other lab.
The safety positioning, in his reading, functions as a marketing differentiator aimed at enterprise customers who want to tell their boards they chose the responsible option. It is not a sincere constraint on behavior. It is a brand attribute that costs nothing operationally and generates significant commercial value with a specific buyer segment that values it.
This is the most contestable part of his argument. Anthropic would say that if powerful AI is coming regardless, better to have safety-focused labs at the frontier than to cede that ground to less careful actors. Zitron's response: that argument conveniently justifies doing exactly what you would do anyway, at the pace you would do it anyway, with the funding you would seek anyway. It is unfalsifiable and therefore uninformative.
Where Did the Productivity Revolution Go
The macro case against the AI investment cycle rests on a single stubborn fact. The United States and its allies have committed more than $500 billion to AI infrastructure, model development, and adjacent services since 2022. GDP growth attributable to that investment is not showing up in the numbers.
Productivity statistics remain within historical ranges. The sectors that were supposed to be disrupted first, legal, finance, software development, have seen some tool adoption but no measurable contraction in headcount or cost per output. The productivity revolution keeps getting described as imminent.
Same pattern as AGI. Always coming.
Zitron's observation about who is actually making money is useful here. The companies capturing the most margin from AI spend are the cloud providers: AWS, Azure, Google Cloud. They sell the compute that trains and runs the models. They take a cut of every API call. Their capital expenditures are rising, but so are their revenues and margins. The picks-and-shovels business is working exactly as the cliche predicts.
The AI labs are in a different position. They are the ones bearing the cost of frontier research, burning cash on talent and compute, and charging prices that don't cover their costs. The cloud providers are the picks-and-shovels play. The labs are selling the shovels below cost and hoping the gold rush justifies it later. The cloud providers collect a fee whether or not the gold rush materializes.
The 18-Month Clock
Zitron's prediction: a significant reckoning arrives in 18 to 24 months when the infrastructure debt comes due. The data centers being built now were financed on the assumption that AI demand would keep growing fast enough to service the debt. If it doesn't, the write-downs begin. Not from one company. Across the sector.
The historical parallel he invokes is the fiber optic buildout of the late 1990s. Enormous capital went into laying cable that the internet would eventually need. But the timing was wrong by several years, and the companies that over-built went bankrupt even though the infrastructure ultimately got used. Being right about the long run does not protect you from being wrong about the timing.
The counter-case is real. AI tools are genuinely useful. Enterprise adoption is growing. Coding assistants, document summarization, customer service automation: these are real products that real businesses are paying for. The question is not whether AI has value. The question is whether that value scales to the valuations the labs are currently carrying, and whether it scales on a timeline that services the debt being taken on today.
Zitron's answer is no. Not at these prices. Not on this timeline. The companies most exposed are not the cloud providers, who have diversified revenue and can absorb a correction. The companies most exposed are the ones whose entire value proposition is frontier model capability, who have no revenue that doesn't depend on the AI boom continuing, and who are spending ahead of any plausible path to profitability.
That description fits both OpenAI and Anthropic. Precisely.
He might be wrong about the timing.
He is not obviously wrong about the structure.
That is the uncomfortable part.