My agents fact-checked his NDC keynote against the transcript, SEC filings. Here's what survived.
Richard Campbell gave a keynote at the NDC (Norwegian Developers Conference) 2026 titled “After the AI Hype — What’s Real, and What’s Next.” It’s a good talk. It’s also the kind of talk that makes you want to sell every AI-related position you own and hide in Treasury bills or gold.
Before doing anything that dramatic, I asked my local agent — Hermes Agent 99, running Codex 5.6 — to pull the full 50-minute transcript, fact-check Campbell’s claims against primary sources (SEC filings, investor relations, the original research papers, Gartner’s own documentation), run three independent review passes, and measure his argument against Leopold Aschenbrenner’s Situational Awareness thesis. This is what came back.
Don’t use this video as a sell-everything signal. Campbell’s central warning is important, but the talk mixes a valid bubble-risk framework with enough major factual errors that it can’t stand alone as investment analysis.
His best point is the one most people get wrong: a genuine technological revolution and an investment bubble can coexist. The internet transformed civilization, and dot-com investors were still destroyed. Fiber got laid, and the companies that laid it sold for cents on the dollar. You can be right about the technology and wrong about the company, the price, the financing, or the timing.
He’s also right that data-center capital expenditure is now macroeconomically significant, that circular financing deserves scrutiny (chip vendors investing in customers who then buy their chips), that many AI products are undifferentiated and unprofitable, and that announced data-center capacity is not the same as financed, constructed, energized, and utilized capacity.
And his conclusion is more modest than his tone: useful technology survives the bust. Focus on solving real problems. The talk itself doesn’t logically support liquidation.
The fact-check turned up errors that sit directly under his investment argument:
He says NVIDIA has roughly $5 trillion in cash. That’s the market capitalization, not cash. NVIDIA’s April 2026 filing showed about $50 billion in cash and marketable debt securities plus $30 billion in marketable equities.
He implies buying NVIDIA stock gives NVIDIA your money. Ordinary stock purchases are trades between investors on the secondary market. The company receives proceeds only when it issues securities.
He presents GPT-4 as a confirmed one-trillion-parameter model. OpenAI never disclosed the parameter count; the trillion-plus figures were industry rumor.
His origin story for the term “artificial intelligence” — scientists trying to raise money from the military — is wrong. The phrase comes from the 1955 Dartmouth workshop proposal, which requested $13,500 from the Rockefeller Foundation.
His histories of ImageNet and Siri are garbled, and his claim that AI companies have “hardly any results” collides with Microsoft’s disclosed $37 billion AI annual revenue run rate, growing 123%.
A speaker who confuses market capitalization with corporate cash, and secondary-market trading with corporate financing, should not be trusted with your liquidation decision.
Aschenbrenner is currently stronger on the direction: continuing capability progress, the physical scale of the buildout, and power, memory, networking, and cooling becoming strategic bottlenecks. Campbell is stronger on the financial mechanics: capex cycles overshoot, circular financing hides weak counterparties, and valuations priced for perfect execution get punished.
Neither has proved his extreme case. Aschenbrenner hasn’t proved late-2020s artificial general intelligence or that trillion-dollar clusters will earn adequate returns. Campbell hasn’t proved that capability progress has peaked or that the infrastructure cycle has already turned.
The three independent review passes converged on one clean test separating a structural boom from a telecom-style overbuild:
Does externally funded, recurring AI gross profit compound faster than depreciation, financing costs, and efficiency-driven price erosion?
Watch that, not the Gartner curve. The Gartner Hype Cycle is a qualitative adoption framework — not a valuation model or a market-timing instrument, and Gartner itself has never claimed otherwise.
Technologically, yes: the biggest spenders are profitable incumbents, the revenue is increasingly real, and the infrastructure serves multiple workloads.
Financially, no: valuation still matters, capex still cycles, suppliers still overbuild, customers still default, and a correct technology thesis can still produce a disastrous stock return.
The agent ran my own portfolio against Campbell’s argument, and his strongest point does apply to me: several of my positions express the same AI-capex thesis through correlated layers — chips, storage, optics, power, and the neocloud financing layer. If hyperscalers pause to digest capacity, those correct together, even if the long-term thesis is intact.
So the real question isn’t “is Campbell right?” It’s: could I hold this basket through a severe multi-year drawdown without being forced to sell? If yes, his talk is a stress test, not a sell signal. If no, the problem is concentration and risk tolerance — not the Gartner curve.
My probability-weighted read: a real AI supercycle, plus at least one serious capital-cycle correction, with brutal rotation between winners and losers along the way. Aschenbrenner is probably more right about the direction. Campbell may be right that too much capital arrives too early — and that some investors get wrecked while the technology wins.
None of this is financial advice. Cecil Ray Burnett III writes as Claudius Fabulous and Leathery Tendons