The AI Spotlight Follow
One Critic Says 2027 Is When the AI Boom Breaks. We Checked His Biggest Claims Against This Week's Numbers
AI Money

One Critic Says 2027 Is When the AI Boom Breaks. We Checked His Biggest Claims Against This Week's Numbers

Ed Zitron argues generative AI is a financial con that unravels in 2027: unprofitable labs, circular money and a trillion dollars of data centers. Here is his full case, the strongest rebuttals, and a scorecard of his claims against Nvidia's latest earnings, big-tech capex and OpenAI's actual losses.

Advertisement

Most AI coverage is written by people who like AI. This story is about someone who doesn't, and whose argument is serious enough that investors, analysts and the people building the technology keep having to answer it.

Ed Zitron, a tech-industry PR veteran of about 16 years turned newsletter writer and podcaster, calls generative AI "at its heart a con." In a long interview that has been making the rounds again, he laid out a specific timeline: the money runs out, the biggest labs can't raise more, and 2027 becomes the year the tech sector breaks.

You can disagree with every word and still learn something from checking his claims against reality. So that's what this does. First his case, in his own terms. Then the strongest pushback. Then a scorecard against the newest numbers, because some of his claims look better than others.

A caution before we start: this is an argument, and an opinionated one. Nothing here is investment advice, and a forecast, by anyone, is not a fact.

Advertisement

The thesis in one paragraph

Zitron's case: the AI industry has sold itself in the language of magic (curing cancer, replacing all jobs) while actually being "boring cloud software that's extremely expensive, unprofitable and unreliable." He argues the revenue is tiny against the spending, that most of it comes from two unprofitable companies funded by the same giants selling them computing power, that heavy usage is being subsidized and that when the subsidy ends, the whole structure unwinds.

The six pillars of his argument

1. A trillion dollars of spending, tens of billions of revenue

Capital expenditure, or capex, is money spent on long-lived assets such as data centers and the GPUs inside them, as opposed to monthly costs like electricity. Zitron says the largest companies have spent over a trillion dollars on AI capex and want to spend another trillion next year, "to make tens of billions of dollars."

His example is the Stargate Abilene site in Texas, being built by OpenAI and Oracle: 1.2 gigawatts of power across eight buildings, which he says will each hold 50,000 Nvidia GB200 GPUs. His point is the concentration: enormous cost, power and labor in one spot.

He also says that other than Microsoft, the big builders have started taking on debt to fund it.

2. The revenue comes from two companies that can't pay for themselves

He claims that, for the cloud giants, around 70% of AI revenue comes from just OpenAI and Anthropic. Both lose money and, he says, depend on investments from the same cloud companies:

  • Amazon sent $50 billion to OpenAI this year and $5 billion to Anthropic.
  • Google sent $10 billion to Anthropic.
  • Analysts, he says, expect more than $400 billion of revenue from those two labs over the next three and a half years, a large share of the cloud giants' future growth.

He calls this circular: investors fund the labs, the labs buy computing from the investors, the investors report growth, markets reward them and more money flows in. He also says the big companies don't separately disclose their AI revenue.

3. "Adoption" is partly forced

He calls generative AI "the largest non-consensual push of technology in history": an assistant appears in your search engine, your documents, your word processor and your online shopping whether you asked or not. People are told constantly that they must use it or fall behind, so usage numbers overstate genuine demand. Much of it, he says, is used like a search engine, "and that's not worth a trillion dollars."

4. Customers are being subsidized, and the bill is arriving

Language models are billed in tokens, roughly three-quarters of a word each, like a taxi meter. Zitron argues that flat-rate subscriptions hide the real cost. A heavy user paying $100 might cost the provider several hundred or even $1,000 in compute, he suggests, though he concedes the companies don't disclose real figures.

There is evidence that bills bite when they become visible. Uber's CTO said that rapid adoption of Claude Code exhausted the company's entire 2026 AI budget within a few months, and Uber later said it brought costs down through prompt caching, more efficient default models and visibility into engineers' usage. In Zitron's telling, this is what happens when "the moment people actually had to pay for it," enthusiasm meets a budget.

5. Reliability: a tool you pay for whether it works or not

You pay for tokens whether the answer is right or wrong. Zitron points to hallucinations he's seen: for example, a financial terminal's AI assistant getting a stock price badly wrong. When it's only a stock price in a draft, no harm done. When it's a doctor's transcription tool, a hedge fund model or a code change that leaves a security door open, the stakes rise, and problems compound when users stop checking.

He also argues that benchmark gains are measured on tests built for language models, and that charts showing AI completing longer tasks leave out the fact that many such results mean success about half the time.

He further claims AI-assisted coding is making software buggier. His evidence is the flood of machine-written code pushed to open-source hosting platforms, more outages and well-meaning newcomers shipping code they barely understand, and the human tendency to check less after the AI has been right a few times.

6. The "five myths" and the doom question

Asked to respond in one sentence to popular beliefs, his answers:

  • "The AI industry is creating enormous economic growth." No. He says spending is mostly chip makers feeding startups, and a few companies feeding each other, and that outside the two big labs the market is tiny.
  • "AI will replace all human jobs." He says there's no economic data supporting it.
  • "The U.S. must spend trillions to beat China in an AI race." He asks what the race is for, saying large models were built without top-end Nvidia hardware.
  • "These models could fall into the wrong hands." In his view, they already are, in the hands of powerful executives.
  • AI safety warnings: He sees "doom" talk as marketing, a scary far-future story that distracts from today's harms, such as slop, errors, copyright disputes and cost.

The punchline: 2027

His prediction is a chain of dominoes. If OpenAI and Anthropic can't keep raising money, the cloud giants lose the growth they've been counting on, orders for chips drop, tech-heavy stock indexes fall and, because index funds hold these stocks, ordinary retirement savings feel it. He says he personally keeps his savings in cash.

The strongest pushback

His interviewer and other critics offered serious responses.

Adoption is real. ChatGPT reached 100 million users in about 60 days, faster than TikTok (nine months), Instagram (two and a half years) and the web. Enterprise surveys show nearly nine in ten organizations using AI in at least one function. A solo founder who isn't a native English speaker described it as "transformative" for writing copy and creating images.

Money-losing doesn't mean wrong. Amazon Web Services lost money for years. Uber burned tens of billions. Spotify lost money for a long time. Zitron's reply, which is the crux of the debate: scale. He says AWS's total capex over its first dozen years was about $29.7 billion (inflation-adjusted), and Uber's losses were about $33 billion, whereas OpenAI lost $20.9 billion in a single year and raised $217 billion in six months, according to him. The difference, he argues, is "magnitudes."

Disruptive technology starts worse. The cars-versus-horses example: early cars were less reliable and more expensive, then overtook the horse. If AI keeps improving and getting cheaper, the economics may work later. Zitron's response is that GPU costs haven't fallen the way chip costs did historically, and that even Nvidia's newer systems claim efficiency in terms of power or dollars per megawatt, not clear cuts to what OpenAI and Anthropic pay.

There is real value. Even Zitron's interviewer said he doesn't dispute that practical uses exist; the disagreement is whether value is proportionate to investment. Zitron partly agreed, describing the spending as a gold rush: big companies whose core businesses are running out of growth saw one new thing to buy, got rewarded by markets for buying GPUs, and rushed in.

"AI" isn't only chatbots. Zitron himself makes this point against the industry: protein folding, robotics and autonomous systems are different technologies from large language models, and the giant data centers are being built mostly for generative AI.

Scorecard: checking the claims against this week's numbers

Here's where we stop arguing and look up data.

Claim: "OpenAI lost $20.9 billion last year"

Mostly accurate, but the full picture is bigger. Reporting based on leaked financials says OpenAI's 2025 operating loss was about $20.92 billion, while its net loss was about $38.5 billion after a large non-cash charge tied to its conversion to a for-profit structure. So the headline number he used is the operating loss.

Claim: "Amazon sent $50 billion to OpenAI this year"

Supported. Reuters reported that OpenAI's $110 billion funding round in February 2026 included $50 billion from Amazon, $30 billion from SoftBank and $30 billion from Nvidia. That also matches his point that a small number of big, interlinked players are funding the labs.

Claim: "More than a trillion dollars of capex, and the spending keeps rising"

Broadly consistent with the latest quarter. For Q3 2026, Microsoft reported about $45.2 billion of capex, Alphabet about $51.4 billion, Amazon about $75.9 billion and Meta about $34.9 billion: roughly $202 billion in one quarter among four companies, by one tally. Full-year 2026 guidance runs around $170 billion for Microsoft, $195 to 205 billion for Alphabet, about $220 billion for Amazon and $135 to 145 billion for Meta. Alphabet expects a significant increase in 2027. One report noted that Google and Microsoft held their capex forecasts, rather than raising them, in October.

Prediction: "Chip orders will fall, and Nvidia's revenue could drop sharply"

Not yet. The newest data points the other way. Nvidia reported its third quarter of fiscal 2027 on November 17: revenue of about $115.8 billion, up 103% year over year, with data center revenue of about $107.6 billion, up 110%. It guided the next quarter to about $130 billion and raised its next-year growth outlook to roughly 75%. By this measure, demand for AI hardware is still accelerating, not collapsing. (Earlier in the interview Zitron cited about $215.9 billion in GPU sales over the previous fiscal year, so Nvidia's current run rate is far above that.)

Claim: "Enterprises are freaking out about token costs"

Supported by at least one prominent case. Uber's budget story is real and widely reported. But Uber also said it managed to cut costs per token, which supports both camps: costs are real, and they can fall.

Claim: "Only two AI customers matter, and they're unprofitable"

Unverified in full. The 70% figure and the $22 billion "everyone else" figure are Zitron's estimates, and the largest cloud companies do not break out AI revenue in a way that settles them.

Prediction: "2027 is when it all breaks"

Unknowable. This is a forecast. It requires a chain of events: lab funding stalls, cloud giants cut spending, markets fall. So far, the most recent evidence (a record Nvidia quarter and rising capex) hasn't shown the first link snapping. But bubbles can look healthy until the day they don't, which is exactly why they're hard to call.

Claim: "Benchmarks are rigged for models"

Partly fair, partly overstated. Independent measurement groups have pushed for harder, more realistic evaluations, and labs' own tables often flatter the home team. But there are also real-world signs of progress, and outside testers have reported steady gains. The honest position is that benchmarks are imperfect, not worthless.

What a sober person takes from this

  1. The financing structure is the real risk, not the technology. Even people who love AI should be asking where the money comes from. A technology can be genuinely useful and still be financed in a way that goes wrong.
  2. Circularity is worth watching. When investors, suppliers and customers are the same few companies, good news and bad news both travel fast.
  3. Price visibility changes behavior. The moment bills arrive in full, as at Uber, organizations get choosy. Expect more cost scrutiny, which favors efficient models.
  4. Usage numbers aren't the same as value. Ask who is paying, at what price, for what.
  5. Both sides can be right about different timescales. Technologies can change the world and still produce a painful investor shakeout first. That was true of railways and the dot-com era.

Practical takeaways

  • If you build on AI, design for portability and cost control: avoid single-vendor lock-in, track token spend per task and have a fallback model.
  • If you invest, diversify and don't treat any single forecaster, bull or bear, as gospel. Know how much of your portfolio is exposed to a handful of AI-linked stocks.
  • If you work in a company adopting AI, ask for outcome metrics, not usage metrics.
  • If you read AI news, separate three claims that often get tangled: is the technology useful, is it profitable, and is it priced correctly?

Counterpoints and risks

  • Zitron's record as a forecaster is itself uncertain. Calling a bubble early and being right eventually isn't the same as being right on timing, and markets can stay irrational longer than critics expect.
  • Some of his figures are estimates or interpretations. Treat unaudited numbers with caution, including the subsidy multiples.
  • Costs may fall. If efficiency gains and cheaper hardware arrive, today's loss-making usage could become profitable.
  • Revenue is growing. Whatever you think of the base, the top labs' revenue has grown quickly.
  • Reasonable people disagree. The strongest rebuttals above come from people who accept some of his data and reject his conclusion.

FAQ

Who is Ed Zitron? A technology commentator and former PR executive who writes and podcasts about what he calls the "rot economy" and argues the generative AI boom is financially unsustainable.

What is his 2027 prediction? That OpenAI and Anthropic will struggle to keep raising money, triggering cuts in big-tech AI spending, falling chip orders and a wider tech downturn.

Is OpenAI actually profitable? No, based on leaked 2025 financials: an operating loss of about $20.9 billion and a net loss of about $38.5 billion.

Did Nvidia's latest results support the crash thesis? No. Nvidia's quarter reported November 17 showed revenue up 103% year over year and a higher forecast.

Is the AI industry a bubble? Unsettled. The evidence of big spending and unprofitable leaders is real. The evidence of real usage and growing revenue is real too. Whether valuations match the future is the actual argument.

The bottom line

The crash thesis rests on a chain with several links, and so far the loudest links have held: the labs are still raising huge sums, capex is still rising and Nvidia just posted another blowout quarter. But the underlying question hasn't gone away. It isn't "does AI work?" It's "does it pay back the trillions being spent on it, and from whose pockets?" Zitron says the answer is no, and 2027 is when we find out. Plenty of smart people say he's wrong. The responsible way to follow the debate is to keep checking who's paying, what they're paying for and how much of it is genuinely new demand.

Sources

The Most Lucrative One-Person AI Business May Not Involve Building Anything. It Is Selling the Thing AI Can't: TasteUp next · 13 min readThe Most Lucrative One-Person AI Business May Not Involve Building Anything. It Is Selling the Thing AI Can't: TasteSoftware is getting easy to build and attention is not. This walkthrough lays out the one-person creative team model: the three roles, the 10-part interview that captures a brand, a research-backed content map and a no-cold-email way to land the first client, plus a sober look at what the big income claims leave out.Keep reading →