Four companies will spend $725 billion on AI infrastructure this year — up 77% from last year's record. Bulls say the demand is real. Bears say it's a bubble. A 100-year-old economic framework says something more useful: both can be true at once — and it tells you exactly which numbers will settle it. Hover or tap any underlined term.
Every AI-bubble debate collapses into two camps shouting past each other. Bulls: "the demand is real — look at the revenue." Bears: "the spending is insane — look at the capex." Both point at true facts. Neither has a framework for how both facts end.
The Austrian school — Ludwig von Mises and Friedrich Hayek, the framework that won Hayek the 1974 Nobel — was built for exactly this argument. Its central concept, malinvestment, is regularly misread as "wasted investment." It means something sharper: investment that looks rational while money is cheap and the boom narrative holds, but whose returns can never justify the capital spent — a mismatch revealed only when financing conditions tighten. The assets can be real. The technology can be world-changing. And the investors can still be wiped out. That's not a contradiction; it's the pattern.
| Railway Mania, 1840s Britain | Telecom fiber, late 1990s US | |
|---|---|---|
| The true story | Railways really were the future of transport. | The internet really was the future of everything. |
| The build | Parliament approved thousands of miles of track; railway shares became a national obsession; investment hit a historic share of GDP. | Carriers laid 80+ million miles of fiber across the US, racing a claim that internet traffic "doubled every 100 days." (Actual rate: roughly doubling per year.) |
| The reveal | Fares and freight couldn't service the capital. Shares collapsed; fortunes were destroyed. | 85–95% of the fiber sat dark years after the crash. Global Crossing went bankrupt with $12.4B in debt; WorldCom became the largest accounting fraud in US history. |
| The afterlife | Britain kept a national rail network — built with other people's lost money. | That dark fiber became the cheap backbone of broadband, cloud, and streaming — a decade later, for buyers at pennies on the dollar. |
That last row is the part both camps miss. The bears were right about the capital and the bulls were right about the technology. The fiber was useful — eventually, to someone else. Malinvestment isn't about whether the thing is real. It's about whether the people paying for it at boom prices ever get their money back, and when.
Austrian theory says a boom turns into malinvestment when specific, checkable conditions hold. Here's each test against the 2026 build-out — with the evidence on both sides:
| Test | What theory looks for | AI 2026 verdict |
|---|---|---|
| 1. Who funds it — savings or credit? | Booms financed by real cash flow self-correct; booms financed by credit expansion overshoot. | SPLIT — and moving the wrong way. The hyperscaler core ($725B from Microsoft/Alphabet/Amazon/Meta) is mostly funded by enormous operating cash flow — genuinely unlike 1999. But the marginal layer is now debt: CoreWeave's $8.5B loan secured by GPUs (the first investment-grade rating ever for chip-collateralized infrastructure) plus another $3.1B publicly syndicated in May. New instruments invented to lever the boom = the classic late-cycle tell. |
| 2. Is capacity running ahead of demonstrated demand? | Watch the price of using the asset vs the spending on the asset. | FLASHING. H100 rental prices fell from ~$8/hr to ~$2.85–3.50 — a ~64% collapse — as 300+ providers piled in, even as capex accelerated 77%. Falling utilization prices + accelerating buildout is the textbook gap. (Counter-signal worth respecting: Nvidia raised H100 rental pricing ~20% in 2026 — the market is genuinely contested, not one-way.) |
| 3. Circular / vendor financing? | When the seller funds the buyer's purchases, "demand" partly measures the seller's own balance sheet. | PRESENT. Nvidia invested $2B in CoreWeave — whose business is buying Nvidia chips. Meta committed ~$21B to CoreWeave capacity. Anthropic and Google together pay SpaceX/xAI ~$26B/yr for compute capacity built partly to serve them. Real contracts, real money — but the same names keep appearing on both sides of the table. |
| 4. Is the narrative doing the underwriting? | Booms run on a growth claim nobody audits (fiber's "doubling every 100 days"). | PARTLY. "Scaling laws + AGI timelines" is the unaudited growth claim of this cycle. Unlike 1999, there IS large, fast-growing real revenue underneath (frontier-lab run-rates in the tens of billions). The question is whether revenue compounds fast enough to service $725B/yr of hardware that dies in 5 years. |
The strongest evidence for test #4 — is the narrative doing the underwriting? — comes not from a bear, but from the man building AGI. Demis Hassabis runs Google DeepMind and won the 2026 Nobel Prize for AlphaFold. He has no round to raise on "AGI is here." And he says the loudest voices have quietly moved the finish line.
His objection is precise. OpenAI defines AGI as a system that can do most "economically valuable work" — i.e., if it replaces enough jobs, we've arrived. Hassabis calls that bar too low, because the human brain is the only existence-proof we have that general intelligence is possible, so that's the bar. His test makes memorization impossible: train a model on a ~1900 knowledge cutoff and ask it to derive special relativity the way Einstein did. Today's AI has read everything ever written — so when it explains relativity, it's retrieving an answer that already exists. Hassabis says no current system can generate one that doesn't — "the best librarian in history," not a mind.
Hayek's Nobel lecture was titled The Pretense of Knowledge — an argument that economies are too complex for anyone to forecast with the precision they claim. Applied here: nobody can call the top of this cycle, including us. What Hayek says you can do is read prices — the market's own information system — instead of narratives. Credit spreads, rental rates, and collateral terms will tell the truth about AI demand long before any keynote or research note does. So we don't forecast. We watch:
| Signal | Why it matters | Bearish read |
|---|---|---|
| GPU rental rates | The live price of using the asset everyone's buying. | Renewed slide through ~$2/hr while capex stays high. |
| Debt share of AI capex | Cash-funded booms fade; credit-funded booms snap. | Hyperscalers shifting from cash to bonds for routine capex; more GPU-collateralized loans. |
| Depreciation schedules in 10-Ks | Quietly extending a GPU's accounting life flatters earnings without changing reality. | Useful-life extensions on accelerators = the fiber-era capacity-swap energy. |
| Neocloud credit spreads | The leveraged edge of the boom — where stress shows first. | Widening spreads / failed syndications on GPU-backed paper. |
| App-layer revenue vs infra spend | The gap that ultimately has to close. | Infra spend compounding faster than end-user AI revenue for another full year. |
| Semis-vs-Bitcoin divergence (retail-liquidity gauge) | Bitcoin historically front-runs retail excess liquidity. Semiconductors ripping (+125%/yr) while BTC bleeds says the rally is institutional capex, not broad liquidity — a boom with no retail bid under it is narrower, and more fragile, than it looks. | The divergence widening further while capex accelerates — an institutional-only boom has fewer hands to pass the bag to. |
Run the bust scenario and the boom scenario side by side, and one asymmetry falls out:
That's the Austrian conclusion in one line, and it's the thesis this site already runs: own the bottleneck, not the boom. Mises would have phrased it differently — the structure of production matters more than the headline — but he'd recognize the trade.
Here's the part both camps miss, and it's the most actionable: a malinvestment bust isn't only a risk to dodge — it's the setup for the next decade's best buyers. Remember the afterlife row of the table above: the dark fiber that bankrupted its builders became the cheap backbone of the cloud, bought a decade later for pennies by patient capital. The losers built it; the winners bought it. So if the AI bust comes, the prize goes to whoever buys the overbuilt capacity at distress prices — but a sharp, AI-specific twist decides which assets to want.
Buy the shell, not the silicon. Fiber lasted 20+ years, so post-bust buyers got a real asset cheap. GPUs rot in 3–6 years — by the time they're cheap, they're obsolete. So the durable wreckage worth owning isn't the chips; it's the data-center buildings, the power contracts, the grid interconnections, the cooling, and the substations — the slow-depreciating shell that outlives any chip generation. A bankrupt neocloud's GPUs are scrap; its power hookup and spot in the interconnect queue are gold.
Who's positioned: patient, cash-rich buyers — the same hyperscalers funding the boom from cash flow could acquire distressed capacity for cents, alongside well-capitalized infrastructure/REIT and private-credit players that can wait. The tell to watch: strategic buyers (not just lenders) showing up on distressed data-center and power-contract deals — that's the afterlife beginning.
Dragonfly Lens maps the full AI build-out chain — and watches the credit plumbing underneath it — with every claim sourced and fact kept separate from hype.
Join the Lens →Sources: $725B hyperscaler 2026 capex, +77% YoY — Tom's Hardware, CNBC; CoreWeave $8.5B GPU-collateralized loan (first investment-grade HPC-secured financing) + $3.1B syndicated facility — CoreWeave 8-K (Mar 2026), CoreWeave 8-K (May 2026); Nvidia $2B CoreWeave investment, Meta ~$21B capacity commitment — same 8-K series; H100 rental collapse (~$8/hr to ~$2.85–3.50, ~64%) — Introl, Silicon Data index, Latent Space; Nvidia 2026 H100 rental price increase ~20% — Crypto Briefing; 1990s fiber overbuild (80M+ miles, 85–95% dark, Global Crossing $12.4B bankruptcy, WorldCom fraud) — Fortune, The Bubble Bubble, Technostatecraft; Hayek, "The Pretense of Knowledge" (Nobel lecture, 1974) — NobelPrize.org; Hassabis "Einstein test" (~1900 cutoff → derive relativity; "today's systems couldn't do that"; AGI ~5–10 yrs) — The News, Zaruko (the Einstein test), Epsilla (path to AGI); Musk's ASI-vs-AGI rebuttal — OfficeChai; Austrian business-cycle theory primer — Mises Institute; SpaceX/xAI compute deals (Anthropic $1.25B/mo, Google $920M/mo) — TechCrunch.
Educational research, not personalized investment advice. Dragonfly Lens is not a registered investment advisor. Figures as of June 2026 from public reporting and SEC filings — verify against primary sources before acting. Companies are named to explain the framework, not as buy or sell recommendations. Past performance does not guarantee future results.