Actionable insights — The AI Bubble and the Market Risks Ahead

1:34 1. When a stock re-rates on big news, check whether the company's bonds agree

1. On any headline that sends a stock sharply higher (a giant contract, a backlog jump), pull the same issuer's bond prices or credit spreads over the same window.
2. If equity rallies while spreads widen or bond prices fall, treat it as a divergence: credit investors are pricing a counterparty or funding risk the equity story ignores.
3. Look at the unsecured debt specifically. A double-digit yield on a "secure growth" company is distressed pricing, whatever the equity multiple says.
4. Compare the credit market's view with sell-side models: if models assume debt issuance that current yields make uneconomic, the model is wrong or the equity is.
5. Ask who the marginal buyer of that debt would be (here insurers and annuities). A thin buyer base is its own risk.

• Neocloud and hyperscaler-adjacent bond yields and CDS; new issuance coupons versus secondary yields; insurers' disclosed exposure to AI-infrastructure credit.

3:28 2. Size a capex boom top-down: installed base → annual carrying cost → revenue that must pay for it

1. Ignore quarterly guidance. Estimate total capital in the ground today plus committed spend to a fixed horizon (here end-2027).
2. Pick a mid-range useful life and compute annual depreciation (installed base ÷ years).
3. Add the other running costs (power, staff, cooling, water) and the cost of capital of the most levered builders, then a normal profit margin, to get the revenue the asset base needs every year.
4. Estimate the actual end-customer revenue (here the labs that are ~70% of demand), subtract their own margin, and see what is left to pay for compute.
5. If the gap is multiples rather than percentages, precision doesn't matter: "my margin of error for being a lot more right than wrong is gigantic."
6. Stress-test it with the smartest bulls you know; silence in place of a numerical rebuttal is information.

• Updated hyperscaler capex commitments; disclosed AI-lab revenue run-rates and gross margins; any change in depreciation lives used by the builders.

19:01 3. Check whether enterprise demand can fund it: what share of corporate free cash flow must be diverted?

1. Take the free cash flow of the largest companies (the natural buyers) in a clean base year before the boom distorted it.
2. Express the revenue gap from insight 2 as a percentage of that FCF.
3. Judge plausibility. Moving FCF to an expense line cuts EPS unless productivity soars, so adoption starts low (a few percent) and ramps.
4. Watch how the market treats companies that cut spend. If restraint is rewarded, the capex bid is losing its shareholder licence.

• Enterprise AI budget surveys as a share of IT/FCF; stock reactions to capex cuts versus raises at earnings.

10:40 4. Test the pricing leg: internalization, aggregators, high-grading and oligopoly lobbying

1. Track unit pricing (dollars per compute) directly; a revenue model that ignores pricing overstates the payback.
2. Look for customers bringing the capability in-house to protect proprietary data. Every large internalizer removes high-value demand.
3. Identify intermediaries already embedded with thousands of customers who could pool buying and push price down; they are the beneficiaries, and the owners of the capacity lose.
4. Apply the commodity concept of high-grading: the first tranche of capacity serves the best customers at the best prices, so later tranches earn less.
5. Read sudden coordinated calls for regulation from incumbents as a possible bid for a licensing moat and pricing power, a sign that pricing is already the problem.

• GPU-hour and token price indices; software vendors launching compute-brokerage products; licensing thresholds in any AI legislation.

17:27 5. In an overcrowded pie, find the player that gets squeezed out, and everyone levered to it

1. List the main claimants on the revenue pool and each one's funding need over the next two years versus committed capital.
2. Check whether private funding is exhausted: the last round's valuation, how far marks sit above cost, and whether insiders are seeking liquidity.
3. Rank who has powerful allies to close the next raise; the one without them is the marginal casualty once pricing competition starts.
4. Separate the fate of the technology from the fate of the equity: assets can keep running through bankruptcy while investors are wiped out.
5. Screen public vehicles for concentrated exposure to the likely casualty (large stakes, backlogs, financing). Violent daily swings on funding headlines are the tell.

• Anthropic's raise closing; OpenAI financing terms (down round, structured debt); SoftBank's margin-loan and stake disclosures; Oracle/neocloud contract renegotiations.

27:20 6. When sovereign long ends break everywhere, short the market still anchored to the lowest yield

1. Chart ten-year yields across major sovereigns. When they rise together, treat it as a global repricing of long-dated government credit, not a local story.
2. Find the outlier still at the lowest yield and ask what happens to its equities (often bond-like defensives) if that yield merely doubles.
3. Use a milestone in a previously suppressed market (here the JGB ten-year reaching 3%) as confirmation that no curve is immune.

• Swiss 10-year yield breaking 1%; SNB policy shifts; Swiss defensive-heavy index relative to global equities.

28:36 7. Test whether a levered economy can service its coupons: unit demand from demographics versus the M2 needed

1. Start from the rule of thumb that money supply (M2) must grow ~2–3% a year to avoid a credit reset, because levered companies need nominal growth to pay coupons.
2. Model the working-age (18–65) population ten years out, stripping out cohorts not in the labour force, to estimate unit-volume demand growth.
3. If units shrink, nominal growth must come from price or currency. With long-end borrowing exhausted, the remaining lever is money printing.
4. Translate that into assets: printing favours gold; the weakest currency bloc is where the pressure breaks first.

• M2 growth by bloc versus the 2–3% line; working-age population releases; central-bank balance-sheet expansion restarting.

31:15 8. Rank sovereigns relatively ("outrun the slowest"), then map election calendars to currency-breakup risk

1. You don't need a sound sovereign, only one that looks better than the worst. Rank blocs by government share of GDP, deficit overshoot and political fragility.
2. Pull polling for upcoming general elections; flag countries where an ostracised protest party now leads everyone else combined.
3. Where government spending is most of GDP, fiscal overspending is the growth, so a new government's currency stance becomes the macro variable.
4. Express it as short the fragile currency / long the hard asset (euro vs gold), sized for a catalyst you wait for rather than chase.

• UK and French election dates and polling; protest-party statements on the euro; OAT–Bund spreads; gold in euros versus in dollars once bond yields stop rising.

35:02 9. Ask whether companies have built capacity for deficit spending, making the deficit impossible to cut

1. Compare sovereign debt/GDP with corporate debt/GDP. A rising government share with a flat corporate share means the state is carrying demand.
2. Companies ignore a one-off crisis stimulus ("an ephemeral bolus") but build capacity for one that persists for years. Check capacity plans in staples and consumer names.
3. Divide the deficit by population to show the per-person transfer households unknowingly depend on.
4. Conclude the policy trap: cutting the deficit leaves excess capacity and collapsing pricing, so the deficit keeps running, and any forced cut is deflationary.

• Consumer-staples capex and capacity-utilisation commentary; deficit-reduction proposals; negative pricing power showing up in volume/price splits next year.

50:02 10. Read a scary inflation print as next year's comp: model the base effect and forecast rate-of-change disinflation

1. When a chart shows prices up sharply year over year, don't stop at the level. Ask what that year-ago number does to next year's comparisons.
2. Roll the base forward: a +7% y/y print becomes the comp nine to twelve months out, and matching it would require a new shock.
3. Build a quarterly nowcast of the inflation rate of change. A peak followed by a projected halving is a disinflation signal the Fed's level-based framework will see late.
4. Use cycle analogues (2007: ten-year yield peaks ~5.29%, then falls straight down) to anticipate the policy flip from hikes to panic cuts.
5. Note the positioning trap: when the market needs the hike that it feared a month ago, the setup is fragile either way.

• Monthly import/export price indices as the 2026 base rolls in; CPI rate of change versus the nowcast; the first Fed pivot language after any hike.