Our forecasting track record

What changed, in plain words

• AI buildout: every part of the chain fell together on Tuesday, not just the chips. Memory maker Micron dropped 7.0%, the broad chip fund 4.1%, data-centre power supplier Vertiv 6.8%, grid contractor Quanta 3.6% and machinery maker Caterpillar 4.6%. — The market repriced the whole business of building AI, chips to electricians, in one session.
• Asia overnight: the selloff intensified, and Asia's memory-chip giants took the worst of it. Korea's KOSPI fell 5.80% with SK Hynix down 9.8% and Samsung Electronics 7.8%; Japan's Nikkei 225 fell 3.16% with SoftBank down 10.3%; Shanghai fell 2.4%. — That is where the world's AI memory actually gets made.
• Government borrowing: the US 30-year rate hit a 19-year high on Tuesday, then fell back overnight. It touched 5.336%, the highest since 2007, before easing below 5.27% by early this morning; Japan's 10-year reached its highest since September 1996, then rallied back. — The spike that scared shares had already partly unwound before this morning.
• Weak bond auction: even at the highest rate since 2001, buyers had to be coaxed into last week's US 30-year debt sale. The 25-billion-dollar auction cleared at 5.216%, with bids of only 2.39 times the amount offered and dealers left holding 11.5% — both below their recent averages. — When governments pay more and still struggle to find buyers, the extra cost tends to stick.
• Trade: the 50% US tariffs on Canadian dairy, drink and vehicles were called off late Tuesday, under two hours before they were due to start. They covered about 20 billion dollars of imports; Canada says they are postponed to end of day Friday 21 August. — A price rise on some Canadian goods was called off — but only until Friday.
• Oil and the Strait of Hormuz: Brent crude closed at 91.02 dollars after the US–Iran interim agreement lapsed, and shipping through the chokepoint nearly stopped — five commodity vessels crossed on Saturday and none on Sunday, against 31 the previous weekend. — Very little oil is moving through the world's most important shipping channel, though no actual shortage has appeared yet.
• Copper, and what it was not: copper's fall was not an AI-demand signal — a squeeze in the London metal market eased after big deliveries of physical metal, taking New York copper to about 6.55 dollars a pound from above 6.80 on Monday. — Two things fell on the same day for unrelated reasons; we are deliberately not counting this one as AI news.
• Gold: gold slipped 1.2% to 4,366 dollars as government borrowing rates surged, and dipped to about 4,325 in Asian hours, with gold-mining shares falling roughly three times as much as the metal. — Higher returns on safe government debt make holding a metal that pays no interest less attractive.
• Today's calendar: the Federal Reserve publishes the minutes of its July meeting at 2pm New York time, with flash business surveys on Friday and the Jackson Hole central-banking symposium late next week. — Dated context, not a prediction: these are the scheduled moments that could move rates expectations.

Headline: the long end repriced the AI buildout, and the radar's non-chip legs are what proved it

The clean read on 2026-08-18/19 is not "chips fell." It is that the entire AI buildout chain repriced together, in one session, on a cost-of-capital shock rather than a demand datum. Compute demand (SMH −4.09%, MU −7.02%, NVDA −2.34%, ASML −4.26%) fell alongside data-centre power and cooling (VRT −6.80%), transmission and interconnection contracting (PWR −3.62%) and heavy equipment (CAT −4.63%). Those last three are on the radar precisely because the traded book expresses AI as a chip trade and cannot see the rest of the chain. They all fired. That is the sensor working, not a coverage gap.

The transmission channel is named and mechanical. The 30-year Treasury yield touched 5.336% on Tuesday — the highest since 2007, against that year's 5.44% peak — which does two things to a buildout financed forward: it raises the discount rate applied to AI profits that arrive years from now, and it raises the actual financing cost of the fabs, substations and data centres under construction. Bloomberg's account of the selloff names AI-related corporate issuance as one of its own drivers, alongside a national debt approaching 40 trillion dollars, a July deficit of 432 billion, and waning demand from traditional buyers of long-dated paper. So the coupling runs both ways: the buildout is helping to push the long end up, and the long end is now repricing the buildout. That is the most important structural change of the week and it is a genuine reflexive loop, not a metaphor.

The honest counterweights matter as much. First, the spike reversed: by early Wednesday the 30-year had given back roughly nine basis points to below 5.27% and the 10-year about six to below 4.69%, while JGBs rallied hard on a strong Japanese five-year auction. Reporting "yields are spiking" this morning would be one session stale. Second, the equity damage is a give-back, not a breakdown — over the six sessions from 2026-08-10 the same names are mostly still higher (MU +9.26%, PWR +5.34%, ASML +4.01%, SMH +0.06%). Third, and most decisively, the corroborating gauges refuse to confirm: high-yield credit spreads sit at 2.70 percentage points and are narrowing, the 2s10s slope is positive at +0.52, HYG and LQD were unchanged, and the VIX rose only to 15.84 with a steeply upward-sloping term structure. A cost-of-capital shock that was breaking the financing leg would show up in credit first. It has not.

Breadth says the same thing from the other side. The cap-weighted S&P fell 0.69% while the equal-weight version fell 0.45%, and the Dow fell only 0.22% as healthcare and energy gained. This is a leadership derating with rotation underneath it, which is the benign version of a concentration unwind — and it is exactly the disconfirmer we have now put on the record as P-0095 rather than left as commentary.

One thing we are deliberately NOT folding into the story: copper. FCX −3.01% and COPX −2.03% look like the critical-materials leg of an AI derating, and they are not. Saxo's account attributes the move to an LME squeeze easing as Trafigura and others delivered physical metal, with New York High Grade at about 6.55 dollars a pound from above 6.80 on Monday. Same-day, same-direction, different cause. Counting it would have manufactured a fifth confirming dimension out of nothing.

Separately, two policy items landed. The first-ever Section 338 tariffs — 50% on roughly 20 billion dollars of Canadian dairy, alcohol and motor-vehicle imports, with no USMCA relief — were deferred less than two hours before their 12:01am ET start, to end of day 21 August. That is de-escalation on a short leash, and it resolves the direction of P-0075/P-0077 for the weekly reviewer. And Hormuz transit has collapsed to essentially zero without an actual supply outage, which is why Brent at 91.02 is elevated but not disorderly.

Readings

Fact-check log

Standing theses — re-scored

• #1 AI-credit fragility — CONFIRM on the mechanism, DISCONFIRM on the symptom. The cost-of-capital channel fired exactly as the thesis describes (long end up → buildout equity down across every leg), and Bloomberg now names AI-related issuance as a driver of the long-end supply glut, which is the reflexive loop the thesis predicted. But the financing leg itself shows no stress: HY spreads 2.70pp and narrowing, HYG/LQD unchanged, no private-credit event. Net: unchanged confidence, better-specified mechanism. — The way we thought AI trouble would arrive is now visible; the trouble itself is not.
  For your money The thing that would actually reach a saver's money — lenders pulling back from risky borrowers — has not happened, and that is the gauge worth following rather than the daily share moves. No reason to change anything on the strength of one session.

  What to watch Matters more if the extra interest risky companies pay starts widening at the same time as AI-linked shares fall. Matters less while borrowing stays cheap and available.
• #5 Sovereign-debt / rates stress — CONFIRM, with the strongest evidence yet. The 2026-08-13 auction is a primary-document confirm of the thesis's own named signal (weak/failed auctions): highest stop-out since 2001, bid-to-cover 2.39, dealers 11.5%, awarded above when-issued. Add a bear-steepening month (30y +13bp, 2y −12bp) and simultaneous multi-decade highs in Japan, Germany and France. Disconfirmer live: the overnight rally and Japan's strong 5-year auction show buyers do appear at these levels. — Governments are paying the most in years to borrow long-term, and last week's auction shows buyers had to be coaxed.
  For your money This is the slowest and most consequential of the risks we track, because long government rates eventually set mortgage rates, loan costs and the value of bond and pension holdings. Slow means there is no reason to rush; it favours spread-out holdings and a cash buffer over any single big decision.

  What to watch Matters more if the next long-bond sales again see below-average demand, or if more governments announce spending without funding it. Matters less if the overnight recovery in prices holds and demand broadens.
• #7 Equity concentration / breadth — CONFIRM on concentration, DISCONFIRM on contagion. Cap-weighted −0.69% vs equal-weight −0.45%, Dow −0.22%, with healthcare and energy up: the index moved because a narrow group moved, which is the concentration risk the thesis names. But the rotation underneath is the benign resolution, and we have put that on the record as P-0095 (p=0.30) rather than asserting it. — A few giant AI companies moved the whole index down while the average company held up better.
  For your money A broad index fund is less spread out than its name suggests when a handful of firms dominate it — useful to know when reading your own balance. It is a structural feature to understand, not a signal to act on.

  What to watch Matters more if the average company starts falling alongside the leaders. Matters less while the rest of the market keeps absorbing leadership weakness.
• #6 Energy & critical commodities — CONFIRM on supply risk, and a deliberate NON-confirm on metals. Hormuz transit at five vessels Saturday and zero Sunday against 31 the previous weekend, with Brent at 91.02 and no physical outage, is the thesis's supply-risk signal without its price signal. Copper's fall was explicitly excluded: an LME squeeze easing, not demand. — The oil chokepoint is close to shut, yet prices are elevated rather than disorderly.
• #9 Labour deterioration — UNCHANGED, with an honest internal tension worth flagging. The deterministic gauge reads calm (Sahm gap 0.00pp; claims four-week average 199,000, −7.3%), but July payrolls fell 23,000 with 103,000 of downward revisions to May and June, and the unemployment rate only fell because participation did. The fast weekly measures and the monthly survey disagree; we are not resolving that by picking the one that suits the story. — Weekly job-loss data looks fine while the monthly job count went backwards — a genuine disagreement, not a verdict.
• #8 AI restructures IT services — mild DISCONFIRM datum. Accenture rose 1.74% on a broadly down day, the opposite of the derating pattern. One session, one name, no bookings data: noted, not scored. — The services company we watch went up while tech went down — interesting, far from conclusive.

Research, not financial advice. Portfolio results are mock capital.
Project snapshot — changes waiting for review: 11 · tracked forecasts open: 58 (checking now: 0) · track record so far: 0.128245 · practice portfolio updated: 2026-08-18.
Past briefs & dashboards: market-sentinel-lyart.vercel.app/reports