Last week on the Prof G Markets podcast, Ed Zitron — one of the most persistent critics of the AI industry — was making his case that not only is AI a bubble, it’s a useless bubble. His point is that unlike the dot-com crash, which at least left behind a lot of dark fibre and cheap servers that Amazon and Google would later repurpose, AI’s infrastructure is a dead end. The GPUs training and running these models have almost no alternative use. When the VC money runs out or the post-IPO bubble pops, there’s nothing to salvage.
I don’t agree with Zitron. But his challenge is worth taking seriously, because buried inside the bubble argument is a more significant question: would you keep using AI if you had to pay the real price for it?
My personal answer is yes. Let me show my working.
The Subsidy Is Real
Let’s start with what Zitron gets right. The AI companies are subsidizing their tools, heavily and deliberately. This isn’t contested, it’s their explicit strategy.
The true cost of running inference on a frontier model has two components. First, hardware: a server loaded with NVIDIA H200 GPUs costs somewhere between $400,000 and $500,000, and that’s before you’ve run a single query. Bottom-up analyses of these hardware costs suggest a real inference cost somewhere around $6 per million tokens generated, versus public API prices that can be as low as $0.60 per million. That’s a 90% discount before you factor anything else in.
The second cost is energy. Inference consumes electricity, and electricity costs are rising in most markets. A single complex AI query can consume ten times the energy of a standard search. Unlike hardware, where there’s a predictable cost-reduction curve as chip density improves and utilization gets smarter, energy pricing is uncertain. Whether small modular reactors, hydro expansion, or some other generation breakthrough brings costs back down is an open question. Zitron is right that this isn’t a solved problem.
Where I part ways with him is on trajectory. Anthropic’s gross margins reportedly swung from -94% in 2024 to approximately +40% in 2025. Inference costs have fallen dramatically over the past two years through model efficiency improvements, better engineering, and specialized hardware designs.
Zitron frames inference costs as a runaway train. I think he’s being unhelpfully pessimistic about a genuine, measurable trend in the other direction. The subsidies will normalize. The question is whether value persists when they do.
There’s a second force Zitron underweights: competition. DeepSeek’s R1 model, released in January 2025, matched OpenAI’s best reasoning model at roughly 3–5% of the training cost. Its V4-Pro is now priced at a fraction of a cent per million tokens — 97% cheaper than comparable offerings from OpenAI. Meta’s Llama series is free, open-weight, and increasingly capable. The competitive dynamic here is deflationary by design. Even if Anthropic and OpenAI stopped subsidizing tomorrow, the existence of near-parity open-weight alternatives caps how high frontier model prices can realistically go. You can’t charge $50 a month for something a developer can self-host for $2.
Measuring What You Can’t Measure
Before I make the case for the value of AI, I have to acknowledge how hard it is to measure.
Martin Fowler wrote a piece in 2003 — still worth reading — called CannotMeasureProductivity. His argument is that we can’t measure software productivity because we can’t reliably measure software output. Lines of code is a bad proxy. Story points are a bad proxy. Anything that can be gamed will be gamed, with predictable outcomes. The field has not solved this in the intervening twenty years.
Birgitta Böckeler, Thoughtworks’ Global Lead for AI-Assisted Software Delivery, makes a related point: new tools don’t change the challenge of measuring outcomes. Adding an AI coding assistant doesn’t give you a new productivity metric — it just makes the existing measurement problem more visible, because everyone now wants a number to justify the spend.
The most rigorous study I’ve seen on AI’s effect on developer productivity produced a result that should humble everyone who claims to know. METR ran a randomized controlled trial — the same gold-standard methodology used in clinical drug trials — with sixteen experienced open-source developers. The developers completed real tasks from their own codebases, randomly assigned to either allow or disallow AI tools. Developers using AI took 19% longer to finish tasks than those working without it but the same developers, asked afterwards, estimated they had been 20% faster. A 39-point gap between perception and reality.
That is a striking result. It is also not the whole story.
METR’s follow-up experiment hit a methodological wall: developers increasingly refused to participate in the no-AI condition, even at $50 an hour. METR ended up acknowledging they couldn’t get a clean signal because too many developers would not work without their AI tools for any amount of money.
You can read that two ways. One is that AI tools create a kind of psychological dependency without objective productivity gains, which is Zitron’s interpretation. Or, that revealed preference is evidence: if experienced developers consistently choose AI even if they eschew a $50/hr incentive, those tools are providing something real. The picture is complicated by the fact that AI is a moving target — models and harnesses improved drastically between June 2025 (the original METR study) and their attempted follow-up, so it’s not accurate to say that METR’s study shows today’s AI tools give developers a false sense of productivity.
Two Jobs, Two Kinds of Value
I should be specific about my own experience, because “AI makes me more productive” without specifics is exactly the kind of claim Zitron (reasonably) dismisses as vibes.
As a technology advisor, my job is to rapidly synthesize broad context and apply it to a specific client’s situation. The raw breadth problem is genuinely hard: any single expert’s career covers maybe three or four industries deeply, a handful of technology eras, and whatever they happened to work on. With AI, I can research a sector I’ve never advised in within an hour, at a depth that would have taken days before.
More importantly, I can maintain a richer model of each client. I make notes of each client meeting, each 1-1 discussion, I save strategy documents and other often very detailed company-specific information, and I track each work item that I produce during an engagement. The accumulated context I carry throughout an advisory relationship is qualitatively different with AI than without it.
As a developer, the value is different. I wouldn’t claim AI makes me faster at tasks I already know how to do in codebases I already understand. The true value of AI is in what I’m willing to attempt. It takes a few seconds to ask a model “what are the standard algorithms to tackle this problem?” and get a rigorous answer. If I’m using a coding agent, I can tell it roughly what to do, maybe answer some clarifying questions, then come back 5-10 minutes later and either fine-tune, accept, or abandon its work product. These two usage styles empower me to simply “do more stuff” in a code base. The value of tackling more ambitious scope is hard to quantify, but it’s real.
This is where Zitron’s framing breaks down. A lot of what AI provides is the confidence to start things you’d otherwise skip — which shows up as capability expansion, not time savings on known work.
What Should It Cost?
The value of a tool is derived demand — it’s worth whatever share of productive output it enables. If you produce R dollars of value per year, and AI lifts your productivity by P% (and you can convert that productivity into revenue), the tool is worth up to P% × R.
Let me run the numbers on my own situation. As a solo practitioner on value-based engagements, my capture rate is close to 100% — every efficiency gain flows directly to my bottom line rather than a client invoice. So if my annual advisory output is worth $300,000, and AI conservatively lifts that by 20%, that’s $60,000 of derived value per year. My current Claude subscription costs about $1,200 annually. Even if inference costs normalise and that subscription reprices at 25× — Zitron’s argument — the resulting $30,000 annual cost is still half the derived value. I keep paying.
The interesting threshold question is at what income level does unsubsidized AI remain rational? The math is:
Annual income × productivity lift % × capture rate ÷ required ROI multiple = maximum defensible tool spend
If you earn $150,000, expect a 15% productivity lift, capture two-thirds of that value economically, and want at least a 2× ROI on the tool, the maximum you’d pay is about $7,500 per year. At that income level, a hypothetical unsubsidized $30,000 annual AI subscription would no longer be worth it. The economics remain compelling for higher-income professionals, higher capture rates, or larger productivity gains, but the math becomes more sensitive as those assumptions fall.
Some useful comparison anchors: a Bloomberg Terminal costs around $24,000 per year. A junior research assistant runs $50,000–$80,000 fully loaded. If AI is genuinely doing a meaningful fraction of that work at a fraction of that cost, the pricing asymmetry holds even after subsidy correction.
There’s a structural difference between solo practitioners and enterprise teams worth being explicit about. At a firm billing time-and-materials, AI productivity gains typically get passed to the client as reduced hours. Value-based practitioners keep them. This is why the ROI case for AI is stronger for independents than it looks in aggregate industry data, and why consulting firms are scrambling to defend their bill rates in the face of AI tooling.
Another way to look at the value of AI is through the lens of competitiveness, rather than dollar value.
The derived-demand framework treats AI as additive, a tool that lifts existing output. But it’s possible that AI is already becoming table stakes for high-end professional work — not an advantage you gain, but a handicap you take on by not using it.
My clients compare me, implicitly or explicitly, against other advisors. An advisor who can synthesize a sector they’ve never worked in within an hour, bring fresh research into every conversation, and maintain a richer ongoing model of client context is offering something qualitatively different from one who relies solely on accumulated experience. If the advisors my clients might hire instead are AI-enhanced, then AI isn’t giving me an edge — it’s keeping me at the table.
History is instructive. Email wasn’t a competitive advantage for long; it quickly became the floor for professional credibility. Internet research, then smartphones — early adopters gained an edge, then non-adoption became a liability. I suspect we’re somewhere in that transition for AI in professional services: past the point where it’s a clear differentiator, approaching the point where its absence is noticeable to clients.
If that’s right, Zitron’s question — “would you pay the real price?” — has a second, less comfortable answer: for some of us, AI may be a requirement, rather than a choice.
What I Don’t Know
I want to be honest about where this argument is weakest.
I cannot run a controlled experiment on myself. The METR study should make me pause about my own self-assessment. There’s a gap between “I feel wildly more productive” and “here is a rigorous measurement showing I actually am more productive.” There continue to be studies attempting to measure AI productivity gains in various ways, but I don’t think this discussion is settled, yet.
There’s also a selection bias in almost every positive AI productivity story, including this one. People who find AI valuable use it heavily, and people who use it heavily are the ones who write articles about it. The sceptics are underrepresented in the discourse.
The price discovery question is genuinely unresolved, though we’re about to get a lot more data. Anthropic filed confidentially for an IPO on June 1, just a few days ago, targeting a listing as early as October. OpenAI is on a similar trajectory toward public markets. When their S-1s become public, they’ll have to disclose the actual economics. Real cost per query, training vs inference costs, margin, and how the gap between their revenue and spend is actually moving. For the first time, we’ll be able to understand the true cost of intelligence as delivered by these two frontier labs.
That said, going public doesn’t mean the subsidies end. Public markets have historically accepted years of operating losses from companies with strong revenue growth — Amazon ran at near-zero margins for over a decade. If Anthropic’s revenue continues its current trajectory (roughly $9 billion annualized at the end of 2025, reportedly crossing $47 billion by May 2026), investors may willingly fund continued below-cost pricing as a growth strategy. The subsidies could persist well into the public-company era.
Paying the Real Price
Zitron’s provocation — that users wouldn’t pay true cost — is ultimately an empirical prediction, not an economic argument. We’ll find out if he’s right as inference costs continue their downward trend, as IPO filings force real transparency into the economics, and as open-weight alternatives keep the competitive floor in place.
But here’s how I’d encourage you to test it for yourself. Think about the AI tools you currently use. If the price tripled tomorrow, which would you keep? If it went up ten times? The tools that survive that filter are the ones where the intelligence is genuinely worth paying for.
I think the answer is different for different people in different roles. For those of us whose output is scalable — advisors, developers, researchers, writers — and who are on value-based pricing, the math is surprisingly favorable even at unsubsidized rates. For someone doing repetitive, well-scoped work in a time-and-materials role, it’s a harder case.
The AI companies are betting that the first group is large enough and hooked enough to sustain their business when the subsidies run out. I won’t like price increases, but AI is such an ingrained part of my workflow that I’d be hard pressed to give it up. I’m not the only one.