Actionable insights — Newsletter #126

1. Before reconciling a contested dataset, force everyone to state which question they are answering

1. When credible estimates of the same quantity differ by 5x, assume definitional divergence before assuming dishonesty. "Almost nobody in the argument is measuring the same thing."
2. Enumerate the distinct questions explicitly. Here there were three: crude loaded inside the Gulf that physically passes the chokepoint; total liquids leaving the wider region by sea (which includes berths the oil reached overland and that never touch the chokepoint); and total regional exports including routes that skip the region entirely.
3. Sort every published estimate by which question it answers, then compare only within a bucket. Goldman's 15–16m answers the third question; the tracker's 4.175m answers the first.
4. Prefer estimators who publish their arithmetic. TankerTrackers published the equation itself (blockade-line departures minus Oman minus Fujairah), so the method can be audited even where the inputs cannot.
5. Look for convergence between methods that share no inputs. Two independent approaches — satellite imagery plus AIS, and vessel-by-vessel transit logging — both landed near 4 million barrels a day. "That convergence did more to move my own estimate than anything a bank… published."
6. Interrogate the estimator on its own blind spots (missed cargoes, double-counted ship-to-ship transfers, vessels running dark in both directions) and accept a floor with a stated error direction over "a point estimate dressed up as truth."

• New estimates that quietly switch buckets — a "Hormuz flows" headline built on regional totals is a different claim than it appears.
• Any estimator who stops publishing the method: that is when the number stops being auditable.

2. Test an official source against its own prior statements before you argue with anyone else

1. Collect the same authority's successive public claims with their dates, and convert each into a run rate.
2. Difference them. If two statements a week apart imply an impossible interval rate, the problem is inside the source, not between sources.
3. Check whether an anomaly has independent corroboration or independent contradiction — including from parties with an incentive to agree with the claim.
4. Before concluding the source is wrong, construct the most charitable reconciliation available. Officials may be quoting per-convoy figures rather than per-day figures, which converts a false claim into a mislabelled one.
5. State the limit of your own audit honestly, then set a default rule: "when the arithmetic does not close I default to the people who publish their method."

• Cumulative-total announcements: they are the easiest place to catch an inconsistent run rate, because each one implicitly prices the interval since the last.
• Whether the party with the incentive to overstate is also the party with the incentive to keep prices low — that alignment is itself evidence about the number.

3. Build the balance bottom-up from named producers, then cross-check it from the opposite direction

1. Rather than picking a side in a range, rebuild the total from its parts using the most recent published figure for each individual producer.
2. Sum them, and record the vintage and provenance of each input so you know which pieces are soft (a market-commentary account citing tracking data is "indicative," not sourced).
3. Now rebuild the same total a second way from a different starting point — here, blockade-line departures plus the routes that bypass the chokepoint — and see whether the two land in the same neighbourhood.
4. Test the residual for physical plausibility rather than accepting it. To reach the higher published estimates you need 5.5–6.5m barrels a day of products and NGLs on top — but regional export refineries have not restarted and global throughput is running ~5m below the prior year, so "asking the region's product exports to run at full pre-war rates while its refineries are down is not something I can make work."
5. Publish your list of unresolved inputs alongside the answer, so the reader can see which assumption to attack.

• The soft inputs he flagged himself: Ceyhan throughput (750,000 contractual vs ~226,000 observed), recent Yanbu loadings, and the per-truck loading on the Syria corridor.
• A producer's figure being restated — the bottom-up build is only as fresh as its oldest component.

4. Name the specific evidence that would make your own build wrong

1. After reaching a conclusion, deliberately construct the strongest version of the opposite case rather than a caricature of it.
2. Identify the structural bias in your own data rather than in theirs. Preliminary flow prints carry an upward revision bias, because dark tonnage gets confirmed weeks later.
3. Quantify the bias with a concrete instance: two very large tankers that had exited five days earlier reappeared near Sri Lanka and revised a seven-day average up by more than a million barrels in one stroke.
4. State the falsification condition plainly — "if that pattern is systematic rather than occasional, my bottom-up build is too low and the market is right to be calm."
5. Concede where the opposing case has been the better trade, not just the weaker argument: the bear case "has been the more profitable one to hold" over six months.
6. Then hold your conclusion loosely and as an asymmetry rather than a forecast: worse than the price implies, buffers close to exhausted, "so the asymmetry from here favors upside surprises over downside ones."

• Whether flow revisions keep running one way. Systematic upward revisions falsify the build; occasional ones do not.
• Your own reaction to a revision: revising the estimate is correct, revising the framework because of one print is not.

5. When price contradicts physical scarcity, find the buffer and put an expiry date on it

1. State the paradox as a question with a number in it: how has the world run 7–8m barrels a day short with Brent at $90 rather than $150?
2. Identify the mechanisms filling the gap and check whether each is renewable or finite. Here both were finite: drawn inventories and destroyed demand.
3. Measure each buffer against its own history, not against last month. Global stocks 8.2bn in January (highest since early 2021), an unprecedented 400m-barrel coordinated release, OECD government stocks at their lowest since December 1990, the SPR at 289.7m — the lowest since 1982 and ~40% of authorized capacity.
4. Separate the temporary from the permanent on the demand side: one bank expects 70% of the Chinese gasoline destruction never to return, which is a structural offset rather than a buffer.
5. Add the third mechanism that hides the shortage without solving it: logistics. The extra cost of shuttling, ship-to-ship transfers and longer voyages "is being absorbed by the system and showing up in freight rather than in the flat price."
6. Date the bill. Roughly 550m barrels over one window alone were bridged with inventory rather than production — "bills like that come due in Q4 when the buffers thin."

• Whether inventory draws resume at the earlier pace with the SPR where it is — "the market will find out quickly how much of the current calm was borrowed."
• Freight rates on the benchmark route: they are the visible price of the workaround, and they decouple from flat price only while the workaround holds.

6. Never accept an industry rule of thumb when the technology it was built for is not the one in front of you

1. When a headline number depends on a generic conversion factor, find out what technology that factor assumes. The 500,000 lb per GW uranium rule is built for light-water reactors and carries enrichment losses.
2. Re-derive it from first principles for the actual asset. CANDU units run on natural uranium with no enrichment and no tails assay, so run burnup (~7.5 GWd per tonne) against thermal efficiency (~31%) and you land at 140–155 tonnes per GW-year — about 390,000 lb, roughly 22% lighter.
3. Sanity-check the rule against the industry body's own figure before assuming the market number is right: the WNA's generic 163 t is ~424,000 lb, so even the standard rule was running conservative.
4. Apply the corrected factor to the actual capacity and the actual life, and convert to an annual and cumulative demand number.
5. Then subtract what is already contracted, and the uncovered tail is the investable number. Reactors mandated to 2064 against a supply arrangement running only to 2040 leaves 24 uncovered years.
6. Ask what makes that demand non-optional: it "cannot be deferred, hedged or engineered away" because a running reactor has no substitute, and the owners just committed private capital to keep it running with no ratepayer vote required.

• Other life-extension decisions: they convert into fuel demand immediately, "with no construction schedule to slip and no investment decision to wait on" — the cheapest megawatt-hour in the industry.
• The Korean licence-term doubling and the US Army's NRC-bypassing microreactor awards: both remove regulatory time, which is the binding constraint rather than technology.

7. Rebase a stale feasibility study before you use it — then re-run the cut-off, not just the price

1. Refuse the tempting shortcut of taking an old study and only marking the commodity price up. "I am not here to go all wild gold bull on the company and just bump up the numbers of the old study based on new gold prices."
2. Inflate every cost input on its own driver, one at a time: the FX assumption (a peso modelled at 20 now trading at 17 pushes local costs up ~18% in dollars), statutory changes (a mining duty lifted from 7.5% to 8.5% plus a doubled precious-metals duty), and peer-observed capital inflation from study stage to feasibility (a sister project went from $36m to $58m, so apply the same rebasing).
3. Credit back anything that genuinely improved — a repurchased royalty, for instance — so the rebuild is not just pessimism.
4. Publish the rebuilt number and label it: "those are my numbers, not the company's."
5. Only then run it at spot, all the way through tax and profit sharing to free cash flow and payback, and discount it.
6. Finally, re-derive the cut-off grade, which almost nobody does. Rebase cost and re-run at the current price and the economic cut-off falls, so the pit widens before anyone drills a hole — plus check what was excluded from the shell (open mineralization below the pit, uncounted deeper sulfides).
7. Sanity-test the margin's tolerance: at a thin grade but a huge margin, missing recovery by several points costs a fraction of revenue "which you absorb without blinking" — whereas the same miss at the old price decided whether the project existed.

• The gap between the rebuilt value and what the market paid on the news — here roughly 10% of the unrisked number, which tells you what the market is refusing to underwrite (execution, not geology).
• The sequence a board actually needs before sanctioning: resource update, upgrade out of inferred, and a pre-feasibility with real reserves. Permits are necessary, not sufficient.

8. Price a financing off what it does not disclose, and benchmark it against the borrower's own prior paper

1. Start with the headline coupon, then locate it against the risk-free curve so the spread is explicit rather than the rate: 12.25% against SOFR ~3.66% is roughly 860 basis points, "rich against mainstream private credit for a borrower with cash flow."
2. Now list every undisclosed term, because each one is unpriced cost: fees, "minimum-return protections" (which usually means a floor on the lender's return, making early repayment expensive), and a warrant package with no fixed count and no disclosed strike.
3. Price the terms you can compare. Warrants running seven years against the 24-month warrants attached to the last placement means "considerably more option value handed across, whatever the eventual count turns out to be."
4. Convert that into a working assumption and say so: "an all-in cost in the mid-to-high teens… and nowhere near the 12.25% headline."
5. Benchmark against the borrower's own history rather than the market: the company issued 12% secured notes a year ago, so it has "swapped short paper for a 5-year milestone-drawn facility at essentially the same coupon" — a decent outcome for a pre-revenue first-of-a-kind builder.
6. Diligence the counterparty. An arranger with a single public transaction and no verifiable committed capital means "the diligence that closes this deal may well belong to whoever ends up holding the paper."
7. Read the conditions precedent for what they imply about equity. A requirement to complete "an investment commitment in such amount as may be agreed" means the company puts money in too: "the non-dilutive framing does not survive contact with the document."
8. Finally, test whether the money works at contracted volume rather than nameplate. Nameplate covers the interest several times over; contracted volume alone leaves it thin.

• The definitive documents (targeted inside 45 days, with an explicit clause letting the parties move the date): the warrant count, whether security is ring-fenced to the projects or an all-assets pledge, and whether prepayment permits a cheaper refinancing after an uplisting.
• Uncontracted tonnes finding homes — the framework "only scales if those tonnes find homes."

9. Value land on the deliverable-megawatt spread, not the acreage

1. Find the variable that actually reprices the asset. For data-centre land it is not location or size — it is "deliverable megawatts on a timeline a buyer can underwrite."
2. Quantify the spread between the two states with observed transactions: unpowered land at $2,500–$10,000 an acre, power-ready Texas sites at ~$800,000, a Northern Virginia median of $2.8m against ~$28,000 for large unpowered assemblages. "A 100x spread on physically similar ground."
3. Identify the specific contract that moves the asset across the line, and price its cost and its risk — here a surety stepping from $22m to $54m by end-2027 securing 250–300 MW.
4. Value the stake, not the asset: take the comparable-transaction range, apply the ownership percentage, and set it against the market capitalisation. If the attributable land alone approximates the whole company, everything else is being valued near zero.
5. Check the jurisdiction as an input, not a footnote. With opposition blocking or delaying 75 US data-centre projects in a single quarter (~$130bn) and opposition groups more than doubling to 833 across 49 states, "scarcity is migrating toward the jurisdictions willing to say yes."
6. Look for a validating comparable transaction on adjacent ground — someone else paying for a neighbouring site with the same attributes is third-party confirmation of the thesis.
7. Then state the condition under which the whole valuation collapses: "that only computes if you assume the land transaction never happens."

• The county's pending data-centre ordinance — standards-based rules preserve the jurisdiction advantage; restrictive ones erase it.
• Whether the power commitment steps up before a sale is signed: that is the point the asset turns into a liability.

10. Find the policy incentive hidden in the government's own balance sheet

1. Look for assets carried at a statutory or historical price rather than market. The US carries 261,498,926 oz of gold at $42.2222, fixed by Congress in 1973 — ~$11bn of book value against ~$1.18tn at market.
2. Trace the accounting mechanism by which the gap could be realised without a sale: raise the statutory price, issue certificates for the difference, the central bank credits the treasury account with the gain.
3. Size the sensitivity so it becomes a rate of change rather than a one-off: every $1,000 on the gold price adds ~$260bn of that capacity; every $4,000 adds about $1tn.
4. Connect it to a live funding constraint. Buybacks were doubled to at least $4bn per operation and dealers assumed bill issuance would pay for it — "the TGA is a checking account though and something has to refill it. New issuance does the job and defeats the purpose. Revaluation does the job and does not."
5. Conclude about incentive, not intent: "that is a policy incentive pointing in precisely one direction." And note the timing implication — a Treasury funding from revaluation gains "wants the gold price higher before pulling the trigger rather than after."
6. Check for precedent before dismissing it as fantasy: Roosevelt marked gold from $20.67 to $35 in 1934, booking a $2.8bn paper profit, $2bn of which capitalised the Exchange Stabilization Fund.
7. State the limits so the idea stays falsifiable: no new wealth is created, $1.1tn covers under 3% of $40tn of debt, the cash arrives in the banking system as reserves so base money expands, and the orthodox legal reading requires Congress to move first.
8. Watch for officials holding two incompatible positions — a denial that revaluation is planned alongside a stated intent to "monetize the asset side of the balance sheet." "Those two positions do not coexist indefinitely."

• Any legislative vehicle naming gold-certificate revaluation as a funding source, and the minority reading of 31 USC 5117 that would let Treasury act without Congress.
• The TGA balance itself, and whether buyback funding is met from it or from fresh bills — the funding source is the tell.

11. Identify the marginal buyer, then ask what it does under stress

1. For any market, name who is actually clearing the marginal supply — not who is described as the natural holder. Here it was levered relative-value funds, not pensions or central banks.
2. Size the position and its history so the change is visible: swap-spread positioning ~$305bn against under $50bn in 2022; Cayman basis funds taking a net $1.2tn of Treasuries over 2022–24, ~37% of net coupon issuance, "near enough to what every other foreign buyer on the entire planet managed put together."
3. Classify the buyer by behaviour, not by label. Funded overnight in repo at 20–100x, it "behaves as a market-maker rather than an investor, and a market-maker heads straight for the exit the second volatility makes the trade too dangerous to keep on."
4. Derive the second-order consequence for correlations everyone assumes are stable. If the marginal holder of the safe asset is the same leveraged crowd, an equity selloff shrinks the whole book — so "the flight-to-safety bid turns up for a day and then flips into yields grinding higher while stocks are still falling."
5. Check who the foreign marginal buyer is and what they actually earn after hedging: an FX-hedged 10-year at ~−1.21% means a hedged buyer is volunteering to lose money, and since mid-May a 10–15bp rise in yield produced a 12bp decline in the hedged yield. Higher yields are pushing that buyer out, not pulling them in.
6. Reduce the whole thing to the binary it implies, and then ask which branch is survivable: either the yield goes far above the level everyone calls too high, or the currency goes much lower. With $1.4tn of net supply to place over two quarters, "only one of those is survivable."

• Repo conditions and any further unwind of the basis trade (~$200bn already gone) — that is the mechanism, not the sentiment.
• Whether the expanded buyback operations, now beginning, change anything at all: "we finally get a proper read on whether $4 billion a pop changes anything."
• The dollar alongside the yield. The two scenarios produce the same yield and opposite currencies, so the currency is the only thing that tells you which one you are in.

12. Refuse a measurement that cannot separate the two hypotheses you are testing

1. When a debate is settled by a derived statistic, check whether that statistic is a residual — something computed as whatever is left over after modelling everything else.
2. Ask whether the residual can distinguish the competing explanations. Term premium "cannot distinguish between a market pricing a higher neutral rate and a market pricing fiscal risk."
3. Check whether the underlying price is being actively supported by a participant with a policy motive. "Measuring credibility using the price of a bond somebody is actively supporting is like taking your temperature with your hand in warm water."
4. Look instead for a direct falsification test that does not depend on the contested statistic: a hawkish chair delivering a hawkish speech should pull long yields down if credibility is the operative variable. They went up.
5. Apply the same scepticism to a constraint everyone is waiting to be relieved. Regulators finalised leverage-ratio relief specifically so dealers would stop being penalised for intermediating Treasuries — and the 30-year then hit a 19-year high on a buyers' strike. "Dealers have the balance sheet capacity and the bid still did not turn up, so the capital rule was never the binding constraint."
6. When two analysts reach the same forecast by different mechanisms, treat the agreement as weak confirmation and keep the mechanisms separate: "Different cause, same yield, which is an odd sort of agreement."

• Anyone waiting on a further regulatory tweak to fix the long end: "waiting for a package that already arrived in the mail yesterday."
• The wedge itself — the gap between the policy-rate path and the long yield is the measurable version of the thesis.