AI: The Biggest Capital Misallocation in History | Market Talk with George Noble

ELI5/TLDR

Three sceptics — an AI researcher, a macro economist, and a markets guy — argue that the AI boom is a giant money bonfire dressed up as progress. Their core claim: large language models are unreliable, every company builds the same one (so nobody has a moat), and the whole industry loses roughly two dollars for every dollar it earns. They think the spending is the largest capital misallocation in history, that it’s being kept alive by accounting tricks and circular financing, and that when it unwinds it could make the dotcom crash look small. They can’t say when it pops, only that the structure guarantees it will.

The Full Story

The conversation has three voices. Gary Marcus is the AI researcher who has been flagging the technology’s limits since 2001. Julian Garran is a macro economist who frames everything through capital allocation. Jack is the markets guy who’s been picking apart the accounting since 2023. George Noble hosts and mostly cheers them on.

The technology has two stubborn flaws

Marcus’s whole case rests on two things that scaling hasn’t fixed and, he argues, can’t fix by scaling alone. The first is efficiency: these models burn enormous energy and chips, need the entire internet to train, and still aren’t satisfied.

The brain runs on 20 watts and we’re building out terawatts after terawatts. Like that just cannot be the right long-term solution.

The second is reliability. Models hallucinate — produce confident, plausible, wrong output — and 25 years of “give us more data, we’ll fix it” hasn’t fixed it. The interesting bit isn’t that they err, it’s where that matters. Marcus reaches for the 1994 Intel Pentium floating-point bug: a chip that made an arithmetic error roughly once in a billion operations triggered a $475 million recall and a national scandal. Today an error rate estimated near 20% is shrugged off. For brainstorming a jingle, fine. For anything mission-critical — Noble’s co-host spent 15 years in nuclear power and won’t tolerate a 1% failure rate — it’s unusable.

This feeds the productivity puzzle. A wave of studies (MIT, McKinsey, Bain) keeps finding the return on investment isn’t there. The mechanism is “work slop”: the model nails the style of a good report, because mimicking style is exactly what it does, but a close look turns up made-up citations and invented numbers. Checking it costs you the time you thought you saved. As Bain put it, “the technology worked, the value didn’t arrive.”

Why nobody can win

Marcus’s sharpest point for an investor is the absence of a moat. The whole field converged on one recipe — the same architecture (large language models), the same bet (scaling), the same fuel (the internet). So everyone runs the identical expensive experiment and gets identical results, including the unsolved hallucinations. One benchmark study found open-weight models trail the closed leaders by only about four months.

You can’t have a sustainable business model if everybody’s basically playing the same game with the same recipe. It only takes four months to catch up.

There’s also a deeper, older limit: novelty. Marcus’s 1998 work showed these networks generalise well inside the space they were trained on but break when asked to extrapolate beyond it. His illustration: a Tesla that drove itself straight into a $3.5 million jet at a trade show, because no training data had imagined jets on the path. A human thinks “big, expensive, avoid.” The model looks for the nearest similar example and, finding none, fails. This is why true level-five self-driving — type in any two points, get there safely — exists nowhere; what we have are constrained routes with remote operators, “a glorified monorail.”

Julian’s golden rule and the subsidy chain

Garran draws a line between narrow AI (closed systems, clear rules, useful for 50 years) and the generalised LLMs everyone is excited about. His rule: you can’t build a product on an LLM that’s commercially viable across the whole supply chain.

Picture a healthy property market — bank lends profitably, builder sells profitably, landlord rents profitably, tenant is happy. Each link makes money. The AI chain is the opposite: data centres lose money subsidising the model labs, the labs lose money (roughly two dollars out for every dollar in) subsidising the app makers like Perplexity and Cursor, who lose money subsidising the end user. Subsidies all the way up. It only holds as long as fresh funding keeps arriving.

That’s why the recording date — June 1st — matters. It’s the day Microsoft switched GitHub Copilot from a flat fee to usage-based pricing. The moment people pay what the compute actually costs, they have to ask whether it earns its keep. Garran’s analogy: a chauffeured Rolls-Royce for $10 a day is wildly useful; charge the real cost and every business built on the cheap version folds. Marcus calls the same shift “the death of token maxing” — the brief mania (Amazon reportedly had a leaderboard) where firms rewarded employees for burning the most tokens, i.e. spending without regard to value. Reports of one customer wasting $500 million in a month killed the mood.

The accounting and the Enron echoes

Jack’s contribution is the financial engineering. He walks through “Valor,” a special-purpose vehicle (Apollo, xAI/SpaceX, Nvidia) that bought $5.6 billion of Nvidia chips — of which $1.9 billion came from Nvidia itself (“the Girl Scout whose dad bought all the cookies”). The remaining debt was packaged by Apollo and sold to Athene, an insurer Apollo owns. The chips sit in xAI’s data centre, the debt sits off xAI’s books, and the ultimate holder is grandma’s annuity. If the depreciation bomb goes off — GPUs are wasting assets, worthless in a couple of years — she eats the loss, not Nvidia or SpaceX.

Garran’s framing tool is the “Wicksell spread”: the gap between the policy rate and the neutral rate (a couple of points above nominal GDP growth). Set rates far below neutral — it hit minus 12% in the pandemic, the deepest ever — and you create a screaming incentive to borrow and buy assets, which then get bid to perfection and impair when the cycle turns. His estimate: misallocated capital across the whole US economy now runs at two-thirds of GDP, some 23 times the level on the eve of the dotcom crash. Not just AI — crypto, private equity, private credit too.

Why it unwinds, and what comes after

The trigger they keep circling is funding. Once the ecosystem can’t self-fund, it depends on the debt and equity markets, and those are starting to balk — Oracle’s debt trading poorly, banks “dancing near the door,” the IPO season forcing OpenAI and Anthropic to disclose real numbers. Marcus thinks OpenAI is the weak link, “the WeWork of AI”: no moat, less capital-efficient than Anthropic, a trust problem at the top, trying to IPO into the same window. If it stumbles first, the ripple is both psychological (it was the poster child) and economic (unhonoured commitments).

On the macro endgame, Garran is bleak. The buildout plus the wealth effect adds maybe 3 points to nominal GDP; if it merely stops growing that’s a recession, if it reverses it’s worse. A bailout wouldn’t work — unlike mortgages, GPUs don’t generate a return if nobody wants them. So the response would be reflation: print and spend to cushion unemployment. But Jack argues the 2020-style playbook has stopped working with inflation already near 3%. The “phoenix from the ashes” Garran sketches is a rotation out of long-duration tech into commodities, resources, emerging markets, and gold.

On timing, everyone hedges. Marcus’s image is Wile E. Coyote, already off the cliff, falling only once he looks down — a psychological event, like the tulip mania ending. Noble closes with the stat that lands hardest: roughly 60% of the S&P 500 by market cap now trades correlated to the AI trade. “It’s become like one giant trade.” Even if you own no tech, you own this.

Key Takeaways

Hallucination is a feature, not a bug, of the architecture — these models mimic the style of correct text; accuracy is incidental. “Work slop” is output that looks right on first read but hides invented citations and numbers, costing more time to verify than it saved.
No technical moat. Everyone converged on the same architecture, the same scaling bet, and the same training data (the internet), so results converge too. Open models reportedly trail the closed leaders by only ~4 months.
The subsidy chain: data centres subsidise model labs subsidise app makers subsidise end users. Frontier labs lose roughly $2 for every $1 of revenue. The whole thing requires continuous outside funding.
Token maxing = rewarding employees for consuming the most AI tokens regardless of usefulness — burning money by design. Its collapse (triggered by usage-based pricing replacing flat fees) is the leading candidate for the first domino.
Wicksell spread = policy rate minus the neutral rate (≈ nominal GDP growth + a couple points). Deeply negative (-12% in the pandemic) signals capital being misallocated at scale. Garran pegs total US misallocated capital at ~2/3 of GDP, ~23x the dotcom-eve level.
Special-purpose vehicles keep GPU debt off the buyer’s balance sheet. In the “Valor” deal, Nvidia funded ~$1.9B of a $5.6B chip purchase from itself; the rest became insurance-held debt — vendor financing dressed as revenue.
GPUs are wasting assets — depreciate to near-worthless in ~2 years — which is why a Fed bailout couldn’t work the way mortgage purchases did: no underlying return.
The novelty problem: LLMs interpolate within their training space but can’t reliably extrapolate beyond it. This 30-year-old limit is why no one has true level-five self-driving.
Neurosymbolic AI (LLMs plus old-fashioned rule-based/symbolic tools) is where Marcus thinks the real recent gains came from — e.g. coding agents wrapped in hundreds of thousands of lines of scaffolding — not from raw scaling.
Agents compound errors: chain 60 steps where each output feeds the next, and a single early mistake cascades. Unreliability scales with autonomy.
The concentration risk: ~60% of S&P 500 market cap reportedly trades correlated to the AI trade. The bear case here is a market-wide bet, not a sector one.

Claude’s Take

This is a one-sided panel — three bears agreeing vigorously and a host egging them on — so treat it as the prosecution’s closing argument, not a trial. That said, the prosecution is unusually well-credentialed and the individual bricks are mostly real: the unit economics genuinely are upside-down right now, the moat question is genuinely unresolved, the circular-financing structures (Nvidia funding its own customers, SPVs parking GPU debt in insurance products) are genuinely documented, and the productivity studies genuinely keep disappointing. Marcus’s technical points — hallucination, novelty, no-moat — are his decades-old hobby-horses, and he’s been more right than the hype crowd admits.

Where I’d push back, with the obvious conflict-of-interest flag since I am one of the products being discussed: the panel slides between “the ecosystem isn’t profitable” (true, and they’re careful about it) and “the technology is worthless” (not what the careful version says, and the looser rhetoric repeatedly implies it). Those are different claims. A bubble in the financing of a real technology — which is closer to their actual argument — is exactly the dotcom analogy they keep invoking, and the dotcom bust did not mean the internet was fake. It meant the capital structure was wrong. The strongest version of their thesis is “catastrophic misallocation around a useful-but-overfunded technology,” and they’re sharpest when they stick to that. The Marcus claim that scaling is fundamentally capped is contested by people who aren’t idiots; he states it as settled. And “I can’t tell you when” is doing a lot of load-bearing work — being structurally right about a bubble and wrong about timing for five years is, for an investor, indistinguishable from being wrong.

Scoring it a 6. High marks for density, real mechanisms, and named specifics (Wicksell spread, the Valor structure, the Copilot pricing date, the 60% correlation stat). Marked down because it’s an echo chamber with no steelman of the bull case, some numbers are asserted rather than sourced (“two-thirds of GDP misallocated” is a big claim resting on one analyst’s model), and the rhetorical confidence runs ahead of the genuine uncertainty they occasionally admit to. Worth watching as the most articulate version of the bear thesis — just don’t mistake a well-argued one-sided case for a balanced verdict.