The AI Investment Bubble: Paper Glory, Cash Gravity
Nov 17, 2025
Strip the confetti off it. The AI investment bubble isn’t a mystery; it’s an accounting costume. Headline “trillions” glam up paper scaffolding—warrants, deferred capacity, vendor credits, SPVs—while real cash flow limps behind the parade. When money is cheap, story outruns math. When money tightens, gravity reminds you what’s funded and what’s fluffed. I’m not here to scream apocalypse or chant salvation. I’m here to separate paper glory from cash gravity so you can position with a spine, not a feed.
First move: call optionality “investment.” Warrants with penny strikes pad press releases but don’t fund anything today; they are contingent leverage in a tuxedo. Second move: call reservations “revenue.” Multi‑year compute commitments are capacity IOUs, not cash in the bank; the demand risk still sits there like a live wire. Third move: loop the cash. An SPV raises funds, buys GPUs, and leases them back to the opco; revenue circulates inside the echo chamber while risk hides in the vehicle you don’t model. Fourth move: barter with sparkle. Vendor credits, offsets, and “strategic partnerships” that settle invoices in favorable accounting, not in dollars that survive the quarter. Ever notice how every deal arrives at round numbers with a comma and a zero that feels good on TV? That’s not precision. That’s choreography.
Unit economics: where illusions break
Tokens don’t pay the power bill; customers do. Inference gross margin starts at price per thousand tokens and then gets hammered by GPU amortization, electricity, cooling, networking, orchestration overhead, and the very human cost of failure and retries. Average utilization—not the launch‑day peak—is the killer. Below roughly half capacity on average, ROI falls through the floor. Every glossy dashboard loves peaks; creditors care about means. Then there’s the grid. Substations and transformers arrive on their schedule, not yours. Interconnect queues run long, water rights get political, liquid‑cooling supply chains don’t bend to a keynote. No megawatts, no model. If that sentence hurts, good. It should.
Watch GPU rental curves for the current generation parts, spot versus term, and compare rates to financing costs—that spread is the tell on sustainable supply and pricing power. Track used GPU resale prices and time‑to‑sale; discounting on the used market is demand fatigue disguised as “market dynamics.” Read cloud capex to sales and depreciation schedules; explosive capex paired with lagging depreciation and slippage on new facilities tells you who is stretching optics. Monitor power permits, interconnect queues, and liquid‑cooling lead times; none of these care about valuation. Pair reported requests per second with revenue and the sales and marketing spend required to hold it; if ARPU falls while S&M rises, the product is pulling the company backward. Finally, follow receivables growth against revenue; when receivables sprint, someone is booking hope.
Three scenarios and the if–then for each
Base case is the grind. Demand keeps rising but pricing compresses; margins thin, deployments stagger on power constraints. In that tape, own boring cash flows tied to reality—power, cooling, networking, grid equipment—and be neutral or underweight the tourists whose AI story is a costume over weak core economics. If you want a commodity tailwind tied to build‑out, keep it small and run hard stops; copper and aluminum do not owe you a smooth line.
Stress case is the margin crunch. Utilization misses plan, revenue recognition wobbles, cancellations hit, credit tightens. Then avoid or short the capacity sellers whose “revenue” is kiosk tickets for a show that hasn’t opened. Buy audited cash flows on 20–30 percent air pockets and make sure the margin line is stabilizing before you get cute. Sell headlines, buy cash.
Blow‑off is the vertical melt with an air pocket at the end. Think “AGI next quarter,” funding at any price, then one permit delay or grid miss and a 30–50 percent drop that feels personal. Then trim 20–30 percent into weekly parabolic candles, add hedges, and reload leaders only after 10–15 percent price and time cool‑offs with improving unit margins. Don’t call the top. Respirate your risk.
A diligence decoder you can run in five minutes
Cash or paper? If you can’t point to dollars arriving now, assume theater. Revenue or capacity IOU? Real usage beats minimum commitments. Owned or leased? On‑balance sheet capex is pain you can measure; SPV loops are pain you discover late. Peak or average? Ask for average utilization, penalties for under‑use, and contract escape hatches. Power path? Show me the substation, the interconnect date, the cooling vendor, and the megawatts under contract. Pricing pressure? Who survives a 30 percent price cut and who dies at 10? If they can’t answer, you just met a valuation that depends on applause.
Enter in tranches over six to twelve months; your first buy is a hypothesis, not a victory. Don’t add to losers unless it was scripted in your plan; hope is not a hedge. Two‑loss day stop. Cap daily portfolio risk at 1.5R. Pair structures where it makes sense: short the narrative‑only capacity sellers against long positions in utility‑like cash flows—grid gear, cooling, power—so that you own the invoice while you question the pitch. On days engineered to hijack your nervous system—CPI, Fed, megaconferences—run orders‑only and trade the sheet you wrote yesterday. The last filter is the most important: the recognition tax. If you need an audience to buy it, pass.
What’s different this time (and what isn’t)
We’ve seen this choreography. Dot‑com vendor financing. Channel stuffing during the Wintel era. Non‑cash “economics” at unicorn factories that printed culture while bleeding dollars. Same playbook, new costumes. What’s new is the power constraint and the scale of the hardware bet. You can spoof software metrics for a while; you can’t spoof electrons. Infrastructure is truth serum. Also new: politics baked into supply chains. Export controls, sanctions, and litigation threads now wrap around the stack. Theater meets permitting, and permitting doesn’t clap.
Take the most popular products and run the plain math. Price per token sounds big until you subtract amortization on GPUs at current quotes, electricity at contracted rates, cooling overhead, orchestration layers, reliability retries, and the cost of humans shepherding broken outputs back into acceptable results. Now lower price by competitive pressure and destroy peak utilization by showing me Tuesdays at 2 a.m. Then ask: would I lend to this at today’s rates? If the answer is no, why is your equity conviction higher than a lender’s patience?
Operator’s field card (paste it by your screen)
Write three truths and obey them. One, power is king; without locked megawatts you are pitching a rumor. Two, average utilization pays the debt; peak charts pay the keynote. Three, free cash flow is oxygen; the rest is perfume. Make every buy answer those in writing. For entries, split into thirds with clear invalidations. For trims, let the tape ask for it—weekly parabolic candles, euphoric superlatives, and vol term inversion into “good news.” For reloads, wait for time and price to do some work; don’t buy the first red; buy the first reclaim with improving margins.
This is the part people skip. Rate your state one to five before any order. If you’re above three—fear, FOMO, shame—pause for ninety seconds and breathe four-four-four-four. Halve the size or pass. Keep a scar ledger. Tag every exit: good loss with rules obeyed, good win with rules obeyed, bad win with rules broken, bad loss with rules broken. Aim for seventy percent good trades weekly. Add one rule that would have blocked your worst mistake. Scars you measure become instructions; scars you brag about become folklore, and folklore is how accounts die.
The AI investment bubble won’t pop because a clever thread arrives. It will deflate through margins when average utilization disappoints and power shows up late. It will expose cash when headline “trillions” meet lenders who ask for dollars, not decks. Your edge is not to predict the day. Your edge is to trade the receipts: rental curves, used hardware spreads, permits, interconnects, utilization means, and gross margin lines that either harden or hollow. Trim when the choir sings. Buy when the curve tightens and the numbers answer to gravity. Don’t trade the narrative; trade the invoice. And if the next headline tries to flatter you into a hero, turn the screen, open your sheet, and make the call your future self can live with on a bad day.
