SK Hynix reported the best quarter in its corporate history overnight: operating profit of ₩60.5 trillion, up 557% year on year, on revenue of ₩79.3 trillion. Both lines missed consensus by roughly 6% — and the capex line, read precisely, was held at the high end of the already-guided ₩40 trillion range while the street had modelled a raise. Seoul’s answer was a second crash day: the KOSPI fell another ~7% after Tuesday’s −10.8% close at 6,024, tripping the eighth market-wide circuit breaker of the year, with Hynix down another ~9% and Samsung ~7%. Two sessions, roughly seventeen percent of Korea’s index value gone — against the best numbers the country’s largest chipmaker has ever printed.

The regime now has a precise definition

Yesterday we wrote that in-line is the new miss. Today’s tape generalises the rule: a record is not a guide. The market has stopped pricing the printed line entirely; only the forward line clears. The ledger from 48 hours, arranged as pairs: Corning beat the quarter and guided in line — it cost 16%. Hynix printed a record and declined to raise its spending plan — sold, twice. Meta arrives tonight carrying ten beats on both lines in its last ten prints — and five of those ten reactions were red, the perfect record the market stopped paying. Against all of that: Teradyne printed a record with a Q3 guide roughly 21% above consensus — paid 13% after hours and up double digits again in the cash session — and Advantest printed a record with a ~35% full-year operating-income raise, one quarter into its year. The split is not semis against the rest, and it is not beats against misses. It is raised futures against unraised ones.

The duopoly is the counter-evidence

The two companies that verify everyone else’s silicon — the testing duopoly — both raised the ceiling of the market they share, inside twelve hours, while Seoul was crushing the memory makers whose output they test. Why that is a structural fact rather than a good quarter — why test is the one layer of the AI build-out that accumulates across every phase instead of rotating with one — is the subject of today’s study, The Verification Tax. The opening section and the full table of contents are free; the complete piece, including how the book classifies and holds the layer, comes with the free C account.

The open is testing the floor with the Fed in the room

The first half-hour of US trading is running the experiment in real time. The semiconductor fund opened near 484 — four points above Tuesday’s 480.50 flush low, the line this book treats as the whole floor. Teradyne holds a double-digit gain while Micron trades ~−1.4% and Nvidia ~−1.1%: discrimination, live, inside one sector, on the same morning. And the two names reporting tonight — Microsoft ~−0.9%, Meta ~−0.8% — are barely down into their own prints, which says the market is withholding judgement until it hears the one sentence it now prices: the spending plan.

Tonight decides the week’s score

The Fed decides at 18:00 UTC with a hold at 3.75% fully priced — the presser half an hour later is the event, and the drift in the price of time into the meeting has been quietly dovish, which is cover for the tape rather than a catalyst. Then the real verdict: Microsoft (est. $4.21 / $87.6bn) and Meta (est. $7.13 / $60.2bn) after the close, alongside ARM, Qualcomm, Lam Research, Vertiv and Amphenol. The score we keep is unchanged from Monday and now has its cleanest possible test: the two biggest capex-raisers of the cycle, reporting into a regime that just demonstrated — on Hynix, at record scale — what it does to companies that print the past without raising the future. A diary observation, not advice: regimes are named by their reaction functions, and tonight the reaction function reports at Mag-7 scale.

C · members block

Into tonight

480.50 is the only line that matters until 18:00. The semi fund is testing the flush low with the decision hours away. Hold it through the presser and the bottom attempt survives its hardest session; close below it and the discrimination story loses its price anchor — the book stops talking about bottoms and goes back to watching the reclaim band from a distance.

Teradyne’s close is the intraday verdict. A double-digit gain held into the bell confirms the tape is sorting — the verification layer paid while the commodity layer reprices. A fade back toward flat says Korea’s block-sale crossed the Pacific after all, and the duopoly evidence gets parked until Tokyo’s answer on Advantest tomorrow.

After dark, listen for one sentence per call. The EPS lines are almost irrelevant tonight; the capex language outranks them. If either Microsoft or Meta raises the spending plan and gets paid for it, the regime that punished Hynix for standing still becomes a regime that sorts — and a sorting market is one this book can work with. If both are sold on beats, the top-callers collect their second external witness at the only scale that matters, and the next accumulation window moves further out, not closer. Thursday’s echo — Samsung’s print and Seoul’s reaction to it — will tell us whether Korea agrees with whatever New York decides tonight.

Verdict: mixed — the tape answered with a rejection sequence, and two of the three branches partially fired. The lines said: SOXX above 530–532 Friday overturns distribution, below 522.24 extends it. The week did both and settled on neither — a six-percent reclaim to Tuesday’s 552.69 close (the old 554 floor tested from below and confirmed as ceiling), then the prints landed, and Friday closed 527.01: back below the band, above the washout. Branch by branch: the handoff did not confirm as leadership — the Handoff Board’s cross-index ratios retraced on the week (Operate/Build −3.2% to 1.118, Use/Build −4.1% to 1.243) as the washed-out build layer bounced. The exit was rejected — HALO finished second of the four, the equal-weight S&P was flat, ten of eleven sectors rose Friday. The internal fired cleanly: Rubin stabilized (+0.4% on the week, best of the four) while verification beat design by 4.6 points inside it — the Verification/Design ratio rose +4.7% through the crack. And the two sessions after the prints re-ran the handoff cascade with force: Rubin −4.3%, Agentic Infrastructure −2.4%, HALO −1.1%, Winners −0.1%. The rotation advances on event days and retraces on flow days — which told us where to look this week: not at the week, at the reactions. The running ledger lives at the track record.

1 · The signal

The switch from AI capex to AI opex and AI beneficiaries does not happen on a date. It happens layer by layer — and its first confirmed tell fired twice this week, on both sides of the same trade, inside 24 hours. When the demand side and the supply side of the build are both spending past their free cash flow, and the market sells both their beats, the capex trade has a short-term peak behind it in the reaction function — whatever the tape does next week.

Tell one, the funder: Alphabet’s quarterly capex ($44.9B) exceeded its operating cash flow ($39.1B) — quarterly free cash flow printed negative $5.9B, the first cash-burning quarter since the IPO — while the 2026 guide rose to $195–205B, 2027 was flagged higher still, and the quarter was financed in part with $21B of net new borrowing. Revenue beat. Cloud grew 82%. The stock closed the week −8.9%, below its 200-day.

Tell two, the builder: Intel printed its fastest revenue growth since 2011 — roughly double the earnings estimate — and raised its own capex plans. The after-hours gain was double-digit; by Friday’s close it was −7.9%. Its print record now reads three beats, zero relief rallies, eight declines.

And the other side of the ledger got paid the same day. On Friday’s broken tape, SAP rose +9.3% and ServiceNow +7.4% — not on vibes, on the first attributable agent economics of the season: ServiceNow’s contracted AI ACV crossed $1 billion in the quarter (net-new AI ACV +40% sequentially, agentic deployments up 9× in nine months, million-dollar-plus deals tripled); SAP’s current cloud backlog reached €22.9B, +27%, with AI in more than 90% of its 50 largest deals. The market sold the companies that spend on AI past their cash flow and paid the companies that bill for it. That is the switch, mid-handshake.

The confirmation ladder. One trigger fired in ink this week, one produced its first print, three remain open — and four more capex guides land within 48 hours this coming week.

2 · What it is

Strip the week to its two windows and the mechanism is visible. The full week was mean reversion: the washed-out build layer bounced hardest (Rubin +0.4%, best of the four indices; the storage sub-index +7.3% after the prior week’s −26%), while last week’s winners gave back (Winners −3.8%, the trust cohort −6 to −7%). The post-print window — Wednesday’s close to Friday’s, the two sessions after Alphabet, ServiceNow, IBM, Tesla, SAP and Intel reported — ran the opposite tape, and that one carries the information: Rubin −4.3% · Agentic Infrastructure −2.4% · HALO −1.1% · Agentic Winners −0.1%. Damage ordered exactly by distance from the capex checkbook. On Friday alone the Winners rose +2.2% — Control-plane +4.2%, Applications +3.6%, Enterprise +3.2%, the Models cohort +5.6% — on the same tape that sold the build-out 3.8%.

Inside the build-out, the market did not sell by sector textbook — it sold by order-dependence. Storage −9.8%, power semiconductors −8.3%, interconnect −7.6%, substrates −6.9%, foundry −6.2% across the post-print window — against lithography −1.9%, photomask −1.1%, gas & chemicals −1.1%, robotics −0.8% and systems & servers −0.1%. What is priced on future orders repriced; what is priced on scarce content, contracted backlog or installed-base workloads held. NVIDIA — the one name in the complex priced on workloads — finished the week green (+2.0%) and fell 0.9% on a Friday that took its neighborhood down 4–10%. The forward map — short term against medium term, the phases from August to early 2027, and the sector leaders inside each index — is this week’s companion piece: The Rotation Is Inevitable. The Date Is Not.

3 · Why it matters

Because the three-stage sequence — the market rewards building AI, then operating it, then using it — is the frame that decides portfolio construction for the back half, and this week moved us measurably along it. The switch is confirmed in the reaction function: capex is now interrogated, not celebrated, on both the demand and supply side of the build. It is not yet confirmed in leadership: the operating layer lost the week (software’s broad basket −5.2%) even while winning the verdict days, and the Handoff Board ratios still sit well above their June 30 base (+11.8% operate-over-build, +24.3% use-over-build) but dipped this week. Both facts belong in the same sentence — a switch that advances on event days and retraces on flow days is early, not broken. Our judgment stands a step further out: the transition window is the second half of 2026 — the Rubin-generation ramp is what moves agentic AI from technically possible to economically rational, and the first fully opex-and-beneficiaries-led market is a 2027 story. Confirmation requires roughly two earnings cycles of the pattern this week started: capex reactions negative, attributable AI revenue compounding, estimates rotating. Probability, not prophecy.

4 · What to look for

The lines this signal is scored against: the reaction rule on the four prints — beats with raised capex guides sold again = trigger one confirmed across the complex; capex guides paid = the crack was macro, the switch clock pauses. Operate/Build and Use/Build rising on the week = the handoff advancing from reaction function toward leadership. SOXX: a reclaim of 530–532 that holds into Friday reopens the bull branch; a close below 522.24 extends distribution. And NVIDIA green or flat through four capex guides keeps the bridge thesis intact — the bridge failing collapses every branch into one.

5 · Bull case / bear case

BullBearWednesday inverts the rule: Microsoft’s capex guide gets paid because the AI revenue beside it is big enough to fund it — the one company where the spend and the attributable return sit in the same income statement. The complex reclaims 530–532, the switch clock pauses without resetting, and the back half broadens instead of rotating.All four prints run the Alphabet pattern — good quarters, higher guides, sold — and the estimate rotation begins. Capacity beta breaks the 522.24 washout on volume, the opex layer keeps winning only relative battles in a falling tape, and the market spends August pricing the gap between the build’s cost and its return. The switch confirms the hard way: through the denominator.

6 · How it may define the next weeks

Whichever way the four guides resolve, the frame for August is set: every AI position now answers to one of three prices — orders (capacity), content and contracts (constraints and backlog), or workloads (utilization and the operators billing for it). The reaction function has already chosen; leadership has not. If the reaction rule confirms across the complex, August is about rotation down the chain and the estimate revisions that follow. If it inverts, the build gets one more leg and the switch waits for the Rubin ramp itself to force it. Either way the instruments are in place: the Handoff Board scores the migration daily, the phase map lays out the road to early 2027, and next Sunday scores this signal. Probability, not prophecy — the referee wears a ticker.

This week’s edition of Closelook@Hypergrowth, dated July 26, 2026.

First, the housekeeping, stated plainly: this letter moves to the weekend. Closelook@Hypergrowth now publishes Sundays, joining Global (Saturday) and US (Sunday) — the mid-week slot is retired, and this edition covers the stretch since July 15. It was a stretch worth two letters. The flow scan’s least favorable pre-print configuration — Netflix, flagged here in black and white — resolved as a growth print sold to a 52-week low within 48 hours. The single warning row on our ledger — KLA — did not stay single: the roll spread through the equipment sleeve, and by Friday the semiconductor complex had broken every base the last edition celebrated defending. The opex flows held their regime through all of it. And the market added a new sort this letter now has to score: on Friday, SAP and ServiceNow were paid nine and seven percent on the same tape that sold Intel eight percent for the season’s best print. The ledger called the last round of shots. Below, we score it — including where it was wrong — and load the next round.

1 · This Week’s Action

The cross-asset backdrop — the floor changed. The last edition wrote “the friendliest possible macro floor under a rotation.” That floor is gone: September hike odds went from one-in-five to four-in-five, oil traded through $100, and Treasuries fell with equities — a discount-rate shock, whose full cross-asset map belongs to Saturday’s Global letter. This letter needs exactly one conclusion from it: a rotation that was running on a friendly discount rate now runs against a hostile one — and it kept running anyway. That is the fortnight’s most important fact about the flows.

The tech shelf — the bases broke, and we said they’d hold. Full honesty against the last edition, which celebrated the semiconductor bases surviving their hardest test: within the fortnight every one of them went — the old floor was rejected on a retest from below, and the complex ended Friday with its first close-basis crack in the June capitulation structure since spring (the level-by-level map, and what the reference book’s contract does about it, is the US letter’s job). This letter’s job is the regime statement underneath, and it is the cleanest one we have ever printed: SMH −11.9% over one month with three green names out of twenty-two — against cloud software +9.2% and broad software +3.8%. A twenty-point monthly spread between the AI trade’s hardware and its software. The rotation is no longer a flow rumor. It is the tape.

The factor view. The equal-weight S&P closed the week at −0.1% against the Nasdaq’s −2.6%; value +1.8%, growth −2.7%; low-vol green. The stress lives precisely where the index weight lives — a concentration event, second episode running, with the average stock declining to participate.

Our own board — the fresh scan. The Rotation Ledger below is the Closelook Directional Flow scan across all four growth buckets, sorted by the 21-day flow delta, scanned July 23. Read it against the last edition’s version:

2 · The State

The four buckets, briefly. New readers: we cut the growth trade into four functional indices — Rubin is the AI build-out, AEI the AI-opex layer, AW40 the agentic winners (the application cohort 2026 punished hardest), and HALO functional growth beyond AI.

The week the ladder flipped twice. The verdict week sorted the buckets in a new order: Rubin +0.4% — the build-out stabilizing after its washout (+92% YTD) — while AW40 gave back 3.7% (−23% YTD), AEI −2.9% (+38%), HALO −1.9% (−1%). Then Friday re-sorted it again: AW40 +2.2% on the day SAP and ServiceNow were paid for claiming the orchestration seat, while Rubin fell 3.8% with the cracking band. Two flips in five sessions is not indecision — it is the market pricing the AI trade as layers rather than as one theme, which is the founding premise of the four-bucket framework. The scoring just moved onto our home field.

The receipts, in names — what the ledger had already said. Two of the fortnight’s verdicts were this letter’s own calls, so they get scored first. Netflix: the last edition printed, in advance, that it carried “the scan’s least favorable pre-print configuration.” It grew, beat, and was sold to a 52-week low within two sessions — the flow state knew what the estimates didn’t. IBM: missed with a cut, mainframe revenue −42% against distributed +37% — the legacy decay this letter has tracked since the 25%-day, now visible inside a single income statement. The rest of the print week (Intel’s sold double-beat, the SAP and ServiceNow orchestrator bid — the full print-by-print record is the US letter’s) matters here as a sort: the market has started pricing enterprise software by which layer of the agentic stack a name occupies — the framework in this week’s long read, The Agent Is Not the Product. The Orchestrator Is., six claimants, at most three seats. For this letter the consequence is concrete: the AW40 cohort’s fate now depends on layer, not on growth rate — and the ledger’s job is to catch the market’s layer-sorting in the flows before the prints announce it.

The boundary, name by name — Friday’s dependence ladder. Friday’s session sorted the AI-hardware complex by a single variable: how much of each name’s value depends on the next hyperscaler order. Our numbers, Friday close, in that order: ARM −8.1% (royalties per chip deployed — the purest order-dependence in the complex, on its richest premium) · Micron −7.0% (HBM is ordered, not consumed) · Applied Materials −4.7% and KLA −3.8% (equipment order books feel capex changes first) · Vertiv −4.5%, Quanta −4.3%, Eaton −2.7% (the data-center power layer) · AMD −3.3%, Broadcom −2.7% (custom accelerators in between) · and at the far end NVIDIA −0.9%, the one name priced on workload execution rather than on anyone’s order book. Now cross-check the ladder against our own ledger, in public: KLA is the scan’s biggest vindication — flow rolled to accelerating-down two scans before Friday priced the reason. And ARM is the scan’s open contradiction: the ledger has it accelerating-up (+31.7) while Friday hit it hardest of all. One of the two is wrong — either the flows are catching accumulation into weakness, or the delta hasn’t caught the repricing yet. That row is now the most interesting single test on the board, and we will score it here either way.

The structural read — scored honestly. Our working QQQ count (Wave 1 into early 2025, Wave 2 the spring break, ongoing Wave 3) took its first real hit: Friday’s close at 684.23 sits below the 686.37/686.76 double-washout floor — the exact escalation line the US letter flagged — while the S&P closed green the same day. A two-point break of the floor with the broader market refusing to confirm is not a completed answer; it is the question, posed at the count’s expense. The honest statement: a fast reclaim of 686–694 keeps the Wave-3 read intact with a deeper internal correction; acceptance below it forces the count’s alternative — that the correction is a degree larger than we have been treating it. We hold the read, at reduced conviction, and let next week grade it.

3 · The Outlook

The tells from the last edition — all three, scored.

• Opex persistence, week three: confirmed — in flow. The scan shows the cluster unbroken and broader: Palo Alto (+42.7), Okta (+42.0), Datadog (+37.5), Fortinet (+36.5), CrowdStrike (+32.7) — all accelerating-up, week three-plus. The doctrine said three weeks is a regime; the flow regime is in ink. The price test arrived simultaneously: the rate shock taxed the whole layer mid-week (CrowdStrike −9.8% on the week, Palo Alto −8.7%) before Friday’s orchestrator bid. Flows lead, price got stress-tested, the regime held. Watch the same cluster through the season’s software prints.

• The conversion rate: pending, closer. Duolingo did not convert — it remains “reversing-up” — but it climbed to #6 on the entire ledger (Δ21 +43.4 against flow still at −103), and the reversal cluster recruited: Snowflake, Rubrik and Axon joined it this scan. Accumulation is broadening; conversion is still the trigger the tactical sleeve waits for.

• KLA — the contagion question: answered, against us. The last edition asked whether KLA’s roll would stay idiosyncratic. It did not. KLA deepened to −138.5 with flow now negative outright, and the roll spread exactly along the sleeve we named: Photronics, Lumentum, AAOI, Sandisk. This was the “stall is the top” scenario’s first domino, and the sector-degree price action — 554 rejected from below, the 582 base gone, the band cracked — confirms the equipment layer’s rest became a de-rating. We wrote we would refuse to pretend otherwise; we are not. The build-out’s core leaders tell the other half, below.

The leaders’ dissent. In the same week the band cracked, NVIDIA closed green (+0.5%, above both moving averages), AMD green, Micron +4% after its −24% month, Dell +9.3% on the ledger’s accelerating flow, DuPont leading the entire scan on the materials side. A complex whose index breaks while its leadership and periphery refuse to confirm is exactly what the last edition called “value migrating from the crowded core to the supplying periphery” — the barbell inside the build-out. The equipment sleeve is repricing; the architecture and materials layers are not. The bucket is not one trade either.

The bucket system now has its own gauge — and it was built for exactly this week. The AI Handoff Board is the four-bucket thesis turned into an instrument: nine ratios measuring whether value is migrating along the chain this letter is organized around — Rubin builds → AEI operates → AW40 uses — each normalized to 1.00 at June 30. The readings, Friday close: operate-versus-build at 1.12 (+12% in four weeks — the opex bucket re-rating against the build bucket), use-versus-operate at 1.11, and the full-chain scoreboard use-versus-build at 1.24. This week all three cross-ratios dipped — which is not a contradiction: Rubin stabilized while the operators gave back, the same double flip the bucket table above shows. The macro event that lit the fuse — the capex-sustainability double-print and the four-layer architecture it forced apart — is the US letter’s story, told at the sector-and-index degree; what belongs here is the system statement (the migration the deltas flagged in early July is now visible in level terms across the whole chain) and, below, the names. Confirmation, in bucket language: two earnings cycles in which AEI-bucket names earn relative estimate upgrades while Rubin-bucket equipment and memory estimates flatten. Wednesday’s referee block is cycle one. (And the opex layer is not an American monopoly — Walldorf proved that this week; the non-US expressions of the layer are next Saturday’s Global material.)

The regime gauge. Money Temperature at 50, up from 44 — dead-neutral. The pairing the doctrine likes (composite steady, rotation internal) survived the fortnight; the composite turning down while rotation runs remains the exit signal, and it has not printed.

The week ahead — the referee block. For the four-bucket framework, next week is not an earnings calendar — it is the referee. The Wednesday/Thursday mega-prints (the budgets that fund the build-out bucket’s entire demand thesis) either re-confirm the capex path at full size, or hand the rotation its biggest receipt yet; the print-record angles and levels live in the US letter. Cadence on Monday matters here for one bucket-specific reason: it prints one session after Kimi K3’s promised open-weights release, and a verified design shock reprices seat-based software economics across two of our four buckets at once. Whatever the verdicts, the ledger scans them before the narratives do — that is the point of it.

Share Closelooknet

Read more

Closelooknet is a reader-supported publication — to receive new posts and support the work, become a free or paid subscriber.

This week’s edition of Closelook@US Stock Markets, dated July 26, 2026.

Last week ended with a question — was Friday’s round trip a spring or the last reclaim? — and this week answered it twice. First the bulls’ way: the semiconductor index rallied six percent off the washout, reclaimed its June zone, and on Tuesday closed within two dollars of the broken 554 line. Then the bears’ way: a rate repricing that took September hike odds from one-in-five to four-in-five, an oil print through $100, and a Friday that closed the chip complex below its capitulation band — on the day Intel reported its fastest revenue growth in fifteen years and was sold eight percent for it. In between, six of the biggest prints of the season landed, and the market applied one consistent rule to all of them: it did not score the earnings. It scored the layer of the stack the company lives in. That rule — who gets sold on a beat, who gets paid on one — is the whole letter.

1 · This Week’s Action

The tape, day by day. Monday drifted; the week was always going to start Wednesday night. Tuesday gave the bulls their evidence: SOXX closed at 552.69, a full reclaim of the June zone that stopped almost exactly at the old 554 floor — tested from below, not regained. Wednesday was verdict night: Alphabet held the beat, raised its capex path again and showed Cloud accelerating; ServiceNow beat on essentially every line; IBM missed with a cut, its mainframe segment down 42% against distributed compute up 37% — the product cycle turning inside one income statement; Tesla missed. Thursday the macro arrived: spot crude traded through $100, September-hike odds spiked toward 80%, the tape fell 2% — Tesla −13.7%, Alphabet −6.5% through its 200-day on a raised guide, gold failing its haven test — while SAP printed positively and rose 6% off its 52-week low and Intel rallied 10% into its own report. Friday sorted the survivors: Intel’s double-beat was sold 7.9%, SOXX fell 4.4% to 527.01 — below the 530–532 band, QQQ broke 694 on its third test — and on the same tape SAP rose another 9.3%, ServiceNow 7.4%, and the S&P closed green with ten of eleven sectors up.

The sector read. Energy led for the third straight week (XLE +3.6%, +33% YTD) with utilities (+2.1%), real estate (+1.4%), materials (+1.2%) and industrials (+1.1%) behind it — the defensive-and-real shelf, again. The bottom two sectors are really two stocks: consumer discretionary −5.9% is Tesla’s −19% week wearing a sector costume, and communications −4.1% is Alphabet’s. Tech itself closed just −0.7% — because inside it, one complex cracked and another rallied, and they nearly cancelled.

The factor read — the average stock skipped this week. The equal-weight S&P closed at −0.1% against the cap-weighted −1.1% and the Nasdaq’s −2.6%; value gained 1.8% while growth lost 2.7%; low-vol and min-vol were green. For the second stress episode running, this is a concentration event, not a market event — the damage lives where the index weight lives.

Inside tech. The weekly numbers hide a violent round trip: SOXX’s −1.3% week contains a six-percent reclaim and a six-percent rejection. The month-scale ledger is the honest one — SMH −11.9% over a month against cloud software +9.2% and the broad software basket +3.8%, a twenty-point spread inside “tech.” And July’s semiconductor breadth is three green names out of twenty-two: NVIDIA, AMD, ASML — everything else red, Intel −30% for the month with the season’s best print. The complex is being repriced from its speculative edge inward: on Friday the chip-designer sleeve (XSD) fell hardest, the broad complex least.

2 · The State

The rule the reaction function is running. Line up the week’s six verdicts and the pattern is not subtle. Beats sold: Alphabet — held guide, raised capex, Cloud accelerating — closed the week −8.9%, below its 200-day. Intel — earnings roughly double the estimate, revenue +25%, its best growth since 2011 — jumped double digits after hours and gave back the entire move by Friday’s open, closing −7.9%; the reversal tracked the one number in the release that wasn’t about the quarter: Intel raising its own capex plans. Its record now reads three beats, zero relief rallies, eight declines. Beats paid: SAP +9.3% on Friday alone, ServiceNow +7.4% — the two names that used their calls to claim the same seat: orchestrator of the agentic enterprise. Misses executed: Tesla −18.9% on the week, unrescued; IBM sold on the cut. The market is not grading earnings anymore. It is grading position in the emerging stack — and this week it told us which transition it is pricing. We published the orchestration half this week — the Pulse (Agentic AI: The Orchestration Knock Out) and the long read (The Agent Is Not the Product. The Orchestrator Is.). The other half has a name too.

The handoff, named: capex → opex. Every infrastructure cycle runs the same sequence: first the market rewards building the capacity, then operating it, then using it. The switch between the first two does not happen on a date — it happens when the constraint moves from building AI to operating and monetizing it, and the market’s tell is precise: capex announcements stop producing positive reactions. That tell printed twice this week, on both sides of the same trade, within 24 hours. Wednesday night the demand side: Alphabet’s quarterly capex ($44.9B) exceeded its operating cash flow — quarterly free cash flow printed negative for the first time on record — while the full-year guide rose toward a ~$200B midpoint and disclosed purchase obligations swelled dramatically (a number to handle with care: those commitments span far more than chip orders, so they support the build’s visibility without being a semiconductor backlog). A marginal-return warning, not a financing crisis — trailing-year cash flow remains solidly positive — but the question changed from how much are you building to what does it return. Thursday night the supply side answered with the same sin: Intel, spending well ahead of its own cash generation, raised capex again — and the market sold the season’s best print for it. When both sides of a trade are spending beyond their cash flow, the sustainability of the orders in between becomes the risk. That is what Friday repriced.

The repricing, read at the sector degree. Sort Friday’s tape by dependence on future hyperscaler orders and the market’s new rule is legible in the sleeves before any single name: the chip-designer cut fell hardest, the broad complex less, the diversified complex least — the same speculative-end-first hierarchy the whole July de-rate has followed. The equipment sleeve repriced before the chip sleeve, as it always does — capex changes reach order books before they reach shipments — and behind it the physical layer of the build (the power, cooling and grid-connection complex) took three-to-five-percent hits as a group: fewer data centers under construction means fewer of everything those buildings contain. Meanwhile the software shelf next door was green. One number quantifies the whole boundary: the spread between the complex’s most order-dependent name and its most usage-priced one was seven points in a single session — the market drawing the line between what is priced on orders and what is priced on workloads. The full name-by-name ladder is Sunday’s Hypergrowth material; what belongs here is the architecture it reveals.

The four layers, named — the thesis this letter now runs on. This week forced the AI trade apart into four layers, each with its own index expression and its own verdict:

1. AI Capex — the build trade (the semiconductor complex, the broad build-out index): repriced. Priced on order flow whose sustainability is now the open question; the band crack is this layer’s chart.

2. The Rubin constraints — scarcity inside the build (test & verification, metrology, consumables, advanced memory and packaging — the layers that get more scarce, not less, when designs and deployments multiply): held. The Handoff Board’s inside-Rubin ratios — verification-versus-design and consumables-versus-tools — were green through the crack week. When the broad capex trade de-rates, the constraint layers are where the build’s value concentrates.

3. AI Opex — the operating layer (software’s consumption side, security, observability, orchestration; our Agentic Infrastructure index, +38% YTD): paid. The orchestrator Friday was this layer’s coming-out session.

4. AI Beneficiaries — the using layer (the Agentic Winners cohort): selective, print by print. No basket treatment yet — this layer must show attributable economics name by name, and mostly hasn’t.

The market stopped trading “AI” this week. It started trading the layers — and every index needed to score that transition is already on the site.

The house instrument already tracks this — and it agrees. We built the AI Handoff Board precisely for this transition: nine ratios measuring value migration across the three-index chain — Rubin builds, Agentic Infrastructure operates, Agentic Winners use — each normalized to 1.00 at June 30. The readings as of Friday: the master ratio (operate ÷ build) stands at 1.12 — the operating layer has re-rated 12% against the building layer in four weeks; the scoreboard (use ÷ build) stands at 1.24. This week both ratios dipped — Rubin stabilizing while the operators gave back — which is exactly what a broadening looks like as opposed to a clean replacement. The board is the standing referee for everything this section claims, updated daily, in public.

The macro engine underneath. None of this happened in a vacuum: Thursday’s flush repriced the September meeting from roughly 20% hike odds to above 80%, real yields pressed toward 3%, and — the week’s signature — Treasuries fell with stocks (IEF −0.8%, TLT −1.3%). A tape where bonds can’t catch the safety bid is a tape repricing its discount rate, and the most extended complex on the board absorbed the hit. Oil is the input behind the odds — spot through $100 this week — but the full term-structure read (and why the December contract near $80 says temporary) is Saturday’s Global letter; for the US tape the transmission is one line: oil → inflation odds → the discount rate → the concentrates.

The levels, and what broke. Tuesday’s 552.69 close is now the defining rejection — the old 554 floor confirmed as ceiling. Friday’s 527.01 close sits below the 530–532 June capitulation band (the first close-basis crack in that band since the spring advance) but above the 522.24 June washout line. One degree up, the escalation line we flagged last week — QQQ closing below the 686.37/686.76 double-washout floor — fired: Friday’s third test of 694 resolved as a break, and the 684.23 close sits two points beneath the floor. And yet the degree above that refused to confirm: the S&P closed green on Friday, the equal-weight was flat on the week, and ten of eleven sectors rose. Last week’s containment question has a split answer — the Nasdaq degree cracked, the market degree did not. Both facts are in ink; neither cancels the other.

The quiet strength, read correctly. NVIDIA finished the week green (+0.5%, above both moving averages) and was the only major AI-hardware name to sit Friday out — not an accident, per the boundary above: it is the one name in the complex priced on workload execution rather than on the next order. A sector-degree break that its usage-priced leader declines to confirm is not yet a bear market in the AI trade — it is the market splitting the AI trade along the order/usage line. Whether the leader eventually pulls the complex up or the complex pulls the leader down remains the cleanest single tell in the US book.

3 · The Outlook

The four indices — the reaction rule, visible in our own instruments. The sold-the-beat/paid-the-plane rule from §2 printed straight through the Closelook index family. On the week: Rubin +0.4% (+92% YTD) — the build-out stabilizing after its washout, even as its ETF proxies cracked. AW40 −3.7% (−23% YTD) — but +2.2% on Friday alone, the orchestrator day, the index’s Control-Plane cohort doing exactly what SAP and ServiceNow’s tape did. The Agentic Ecosystem −2.9% (+38%), HALO −1.9% (−1%). One tape, four different verdicts — the AI trade is no longer one trade, and the full four-bucket flow system behind these numbers is the Hypergrowth letter’s home turf (also out this Sunday). Here the Agentic Winners index earns one sentence: it is where the orchestration thesis gets scored print by print, in an index, in public.

The regime gauge. Money Temperature at 50, up from 44 — back to dead-neutral. The forced-selling phase cooled; nothing euphoric replaced it. At 50 the market funds rotations, not chases — which is precisely what the sector board shows.

What would confirm the handoff — the diary’s checklist. One violent Friday names a transition; it does not confirm one. We will call the capex→opex handoff confirmed when, and only when, several things print together: capex raises keep getting sold across the hyperscalers (Wednesday and Thursday are live tests); analysts stop raising out-year estimates for equipment and memory while orders are still strong; consumption-linked revenue — inference, observability, security, orchestration — visibly accelerates off the installed base; and the opex names show attributable AI revenue, not “AI is a tailwind.” By that standard, July 23–24 was the opening shot, not the verdict: confirmation needs roughly two earnings cycles of relative estimate migration, which points at the second half of 2026 as the transition window. And the both-hands clause, stated plainly: this is a broadening, not a replacement — inference growth itself demands compute, so the usage-priced end of the hardware complex can stay a winner while the order-priced end de-rates. The vulnerable cohort is narrow and specific: marginal, long-duration capacity whose valuation requires hyperscaler capex to accelerate forever.

The week ahead is the season’s center of gravity. Cadence prints Monday — one session after Kimi K3’s promised weights release, making it the one name carrying both the design-shock verification and the sold-beat regime at once. Then the block that funds the entire AI trade: Microsoft and Meta on Wednesday, Amazon and Apple on Thursday — the four largest capex budgets reporting within 24 hours, into a cracked semiconductor band, an 80% hike probability, and month-end flows. Our Microsoft print-record card carries the season’s sharpest stat into Wednesday: twenty beats in twenty quarters — and a stock that has been sold anyway in the recent ones. If the sold-beat rule survives that print block, it is a regime. If it breaks, Friday’s crack loses its sponsor.

Share Closelooknet

4 · What May Lie Ahead

The levels, updated. SOXX: the map now reads 530–532 as overhead (the band, lost on Friday’s close), then 522.24 (the June washout line — Friday closed 527.01, between the two), then the primary rising channel floor near 520 — the line that decides whether the down-channel from the June highs stays corrective — and below it Friday-of-last-week’s 498.54 print. The structural read is unchanged from our chart work: the decline from June still counts as a flag-shaped correction while the channel floor holds. QQQ: 694 and the 686.37/686.76 double-washout floor are both overhead now — Friday closed 684.23, the first close beneath the floor; a fast reclaim of 686–694 would mark Friday as a false break (SPY’s green close argues that branch is live), while acceptance below it points the map toward the June congestion zone. SPY: green Friday, ~745 (the 50-day) overhead, structure intact. Alphabet: closed 319.74 against the 200-day at ~324 — the raise-and-sold print leaves it below the line institutions defend; a reclaim would be the first evidence the collector-selling is exhausting. Intel: the 3-0-8 record needs no level — it needs a reaction regime change.

The calendar just testified — against itself. July is, by a wide margin, QQQ’s most reliable month: +3.45% average, positive 88% of the time — 15 of the last 17 years. July 2026 printed −7.08% — not merely the first red July since 2024, but the worst July in the entire 17-year sample, more than four points beyond the previous worst. When the strongest seasonal tailwind of the year produces the worst print in its own history, the message is not seasonal noise — it is that something in the tape currently overpowers the calendar’s most dependable pattern. And the calendar now stops helping: August is mediocre on QQQ’s history (+0.65% average, 56% positive) and September is the only negative-average month of the year (−0.31%, positive less than half the time). The next reliably green columns are October (+2.50%) and November (+2.92%). The band cracked at precisely the moment the seasonal floor was scheduled to leave the building — which cuts both ways: the bears lose their “it’s just seasonality” objection for July, and the bulls lose the calendar as a crutch until October.

The three bellwethers. The architect diverged from its complex: NVIDIA +0.5% on the week, green, above both averages — the only major AI-hardware name that ignored Friday. Micron +4.0% — the first green week after a −24% month, the memory unwind pausing at scale. TSMC −1.1%, quiet, still below its 50-day (~426) but far above its July low: the paradox holder is consolidating, not breaking. A band crack with all three bellwethers flat-to-green is a very different animal from June 9, when they led the washout. Watch whether that divergence resolves upward (leaders pull the complex) or downward (the complex pulls the leaders) — it is the cleanest single tell in the US book right now.

Read more

For two letters running, the stress in this book came from Asia: first the Tokyo leverage unwind, then a Chinese model that claims to design chips. This week the direction of travel reversed. The market-implied odds of a September Fed hike went from roughly one-in-five to above eighty percent in a single week, spot crude traded through $100, and the US semiconductor complex closed below the support band it had held since spring — on the same day Intel delivered its fastest revenue growth since 2011 and was sold 8% for it.

And the ex-US map? Taiwan finished the week green, Korea’s decline shrank from double digits to one and a half points, the SK Hynix debut ended the week where it started, and the all-world ex-US benchmark outperformed both the S&P and the Nasdaq. This letter is about what it means when the test moves from the supplier’s side of the AI trade to the buyer’s side.

1 · This Week’s Action

The cross-asset backdrop. Oil again — USO +10.4% on the week, the second consecutive double-digit week, now +97.6% year-to-date. Copper miners joined (COPX +5.6%), silver held its bid (SLV +2.7%), and that real-asset shelf sat on top of a board whose bottom half tells the real story: the mega-cap concentrates led the equity decline again (QQQ −2.6%, QTOP −3.0%, TOPT −1.9%, against SPY −1.1%) — and this time the bond shelf fell with them (IEF −0.8%, TLT −1.3%).

That is the week’s signature. In the Tokyo-unwind week, Treasuries caught the safety bid; this week they were part of the problem. When stocks and bonds fall together, the market is not repricing growth — it is repricing the discount rate. Gold (+1.1%) and the dollar (+0.7%) both firmed, which is what that regime usually looks like.

The global sectors. Energy led for the third consecutive week (IXC +4.2%, now +32% YTD) — the oil chart is no longer a footnote to this board, it is the board’s organizing fact. Behind it, the same defensive-and-real shelf as last week: materials +1.5%, utilities +1.3%, REITs +1.3%, industrials +1.1%, healthcare and financials modestly green.

The bottom is new, though: not tech, but communication services (−4.2%) and consumer discretionary (−3.6%) — the sectors where Alphabet’s post-print drift and Tesla’s unrescued miss live. Global tech itself closed only −0.8%, a number that hides everything: inside it, semiconductors were cracked on Friday, software gave back an early-week rally, and the orchestration layer had its best day of the season. The aggregate is calm; the interior is violent.

The regions. No washout, no single story — which is itself the story after two weeks of forced selling. Norway led on the oil bid (+3.6%), Hong Kong +1.7%, Taiwan +1.4% — its first green week in three — Spain +1.2%, Argentina +1.1%.

The red column is shallow: Indonesia −2.8%, Malaysia −2.0%, India −1.9%, Korea −1.4% — and Korea’s number deserves a second look, because it was flat through Thursday and took nearly all of its decline in Friday’s US-driven session. The one to frame: VEU closed the week at −0.35%, against −1.1% for the S&P and −2.6% for the Nasdaq. In a stress week manufactured in the US, the ex-US book outperformed the US core across the board.

The Global Compass

New this week, and recurring from here: the Compass — the same four relative-strength questions, answered the same way, every Saturday.

Who leads and lags, at two horizons; and the ratios that decide the biggest allocation questions in this letter: stay home or go global, stay tech or go broad. The Asia-based home-vs-global view — the most heterogeneous one — joins next week.

Regions: the year’s leaders are the month’s laggards. The YTD column still belongs to the AI supply chain — Korea +68%, Taiwan +54% — with Thailand, Norway and Austria behind.

Flip to the one-month column and the same two names sit at the bottom (Korea −21%, Taiwan −7%) while the month’s leaders are the year’s forgotten: China’s large-caps +9%, Norway +8%, Hong Kong +7%, Poland +6%. That inversion is the July story in one picture: the crowded winners digesting, the money staying in the region but rotating down the leaderboard.

Sectors: one sector leads both clocks. Energy is the only sector on the board leading the year (+32%) and the month (+11%) — the oil chart translated into equity leadership. Tech is the mirror: first on the year (+27%), last on the month (−5%).

The month’s quiet strength — financials +6%, REITs +4%, healthcare +3% — is the defensive-and-yield shelf this letter has flagged for three weeks, now visible at the horizon where trends get set.

Stay home vs go global — the US view. The ratio tells a story the S&P’s own chart hides: US leadership over the world peaked in early 2025 and has printed lower highs since. The spring-2026 recovery leg stalled this month below the old shelf, and this week the ratio ticked down again — the rest of the world outperformed America in a stress week America itself produced. One line, one question, updated weekly: is the decade-long stay-home trade resuming, or was the spring bounce the counter-trend?

Stay home vs go global — the Europe view. For a Europe-based reader the same question has a harsher answer: the home-vs-world ratio has been in a downtrend for the full three years of the chart, and every counter-rally — including this spring’s — has been sold. A Europe-only book has structurally lost to a global one, SAP’s week notwithstanding. That is precisely why this letter’s map is built on where the flows land, not where the reader lives.

Stay tech vs go broad. The concentration trade in one line: a three-year structural uptrend that went vertical in May–June — the AI re-rating at global scale — peaked in late June, and has been rolling over since, now back at its 50-day average. Not broken, but no longer leading: the month in which SMH lost 12% while the world index barely moved is exactly this rollover, seen from above. A decisive break of the average would be the broadening signal the equal-weight bulls have waited two years for; a bounce from it would say the concentration regime survived another test.

Share Closelooknet

2 · The State

The engine: a rate repricing, not a growth scare. The week’s driver was Thursday’s flush — tariff-driven inflation prints and Fed commentary pushed the market-implied odds of a September hike from roughly 20% to above 80%, real yields pressed toward 3%, and the Mag7 complex took broad losses.

This is a different animal from the two previous corrections in this letter. Tokyo was positioning; K3 was narrative; this is the discount rate — the one variable that touches every asset on every board above at the same time. Hence bonds red, gold up, dollar up, defensives green: the classic fingerprint.

The print week, read from the flow map. The season’s first full week gets its print-by-print scoring in Sunday’s US letter; what belongs here is the geography of it. The American tape sold its own silicon champions on good news — Intel’s strongest revenue growth since 2011 was met with an 8% decline — while the week’s single best print-reaction on the entire developed-market tape belonged to a European name: SAP, up 9.3% on Friday alone, off a 52-week low, on its claim to the orchestration layer of the agentic enterprise (the full thesis: The Agent Is Not the Product. The Orchestrator Is.). For a letter about where value in the AI economy accrues geographically, that pairing matters: the market spent the week repricing which layer of the stack earns the profit pool — and the layers are not all American, and not all in the index everyone watches.

The silicon crack — located, not mapped. The US semiconductor complex closed the week below the support band it had defended since spring — the first structural crack there since this book’s full add; the level-by-level map and what our book’s contract says about it are the US letter’s territory. What this letter takes from it is the location: the crack happened in the buying market, not the producing one. Taiwan and Korea’s ETFs — the supply chain itself — finished green and nearly flat respectively while the US complex broke. And note what the crack is not: it is not the K3 design-shock scenario — the EDA names were quiet. This was a discount-rate shock finding the most extended complex on the most rate-sensitive tape.

The Asian legs held — scored honestly. Last week’s letter ended with one test: Seoul’s Monday reopen, with the US market front-running a gap-up. The gap-up printed — Korea recovered through midweek, and the SK Hynix ADR rallied from its tested $149 line back toward $170 before Friday’s US flush took it down 8.8% to close at $154.57 — almost exactly flat on the week ($154.03 last Friday). A debut that absorbs a Kioxia collapse one week and a US rate shock the next, and holds its offer line on every close, is building a record. Taiwan’s green week and TSMC’s steadiness (below its 50-day, well above its July low) complete the picture: this week’s stress did not come from the supply chain — it came from the customer’s cost of capital. The two-country thesis was a bystander, not a casualty.

Four structures, one grid. Before the count, the states — because they differ, and the difference is the point. The build-out index is ranging: two months inside a well-defined box, currently at the box floor. The US semiconductor complex is testing: still inside its rising channel from the autumn, but in the channel’s lower half after losing the midline. VEU is resting: flat on its horizontal shelf, capped by nothing more serious than a three-week trendline. Taiwan is correcting: under its June down-line, above its breakout. None of the four is broken; all four are asking their own question. That is a rotation’s chart-book, not a top’s.

The structural read — the count, updated on the chart. The weekly count from the October 2022 low now reads: wave 1 into the early-2023 peak, wave 2 into the late-2023 low, and the long wave 3 as the advance since — with its terminal zone drawn just overhead in the 88–90 area (the high so far: 85.74). After it, the map shows a wave-4 rest — the kind measured in months, with the mid-70s as its natural landing zone — and a fifth wave projecting toward the high 90s. The live question is only where in that sequence this pullback sits: either wave 3 is finishing (one more push toward 88–90 before the rest), or the rest began at 85.74 slightly early. Both branches carry the same portfolio conclusion, which is why we can hold them at once: the multi-year advance is not complete at trend degree, but a months-long consolidation is on the map — before or after one more high. This week’s contribution was behavioral, not structural: through a US-made stress week, the ex-US structure was the calmer one.

3 · The Outlook

The four indices. Last week the ladder sorted by physical exposure — Rubin took −9.3% while the agentic operators finished green. This week the ladder inverted: Rubin +0.4% (+92% YTD) as the build-out complex stabilized after its washout, while AW40, the agentic operators, gave back −3.7% (−23% YTD), the Agentic Ecosystem −2.9% (+38% YTD) and HALO −1.9% (−1% YTD).

Then Friday inverted it again: AW40 rose +2.2% on the orchestrator day while Rubin fell −3.8% with the SOXX break. Read the two moves together and the picture is not “hardware versus software” — it is a market that no longer treats the AI trade as one trade. The rate shock hit the leveraged and the extended; the orchestration repricing paid the control plane; and the Agentic Winners index — whose Control-Plane cohort is the diary’s expression of the orchestration thesis — is where we keep score.

The regime gauge. Money Temperature composite at 50 — up six points from last week’s 44, back to dead-neutral mid-band. The gauge is saying what the regional board says: the forced-selling phase cooled, and no euphoria replaced it. Rotations get funded at 50; chases do not.

4 · What May Lie Ahead

Levels. VEU: the 50-day at 82.98 remains the overhead test, the 80.73–78.77 confluence band (congestion shelf + the trendline through the April 2025/2026 lows) remains the downside tripwire — both untouched this week, both unchanged in meaning. SOXX: the 530–532 band is now overhead; the next support is the primary channel floor around 520, the mid-July washout low — and the down-channel from the June highs still reads corrective (flag-shaped) unless 520 goes. QQQ broke 694 on its third test (closed 684); that level flips to resistance. SPY sits just below its 50-day (~745) with a green Friday and 10 of 11 sectors up — the index-level tape is not broken, the tech-weighted tape is. SK Hynix: $149 remains the line that defines the debut. Korea’s 50-day sits far overhead at ~189 — July’s scar needs months, not weeks.

The calendar turns against the book — now. One more input for the two-month map, and it points the same way as the wave count: seasonality. On VEU’s 17-year calendar, August and September are the only two consecutive months with a negative average — August −1.1% (positive only half the time; the worst prints: −8.5% in 2011, −7.8% in 2015), September −0.4%. The rest of the year is green on average; the recovery columns are October through December. Read it against the structural read above: the fourth-wave rest the chart maps and the calendar’s only soft patch occupy the same two months. Neither is a forecast alone; both pointing at August–September at once is how this letter defines an unfavorable window — for adding, not for holding. And a footnote from the month just closed: July was the seasonally mediocre month for the ex-US book (−2.5% this year) and the seasonally best month for the Nasdaq — which printed its worst July on record instead; that anomaly is the US letter’s story, but its geography belongs here: the seasonal script failed in America, not abroad.

Monday is a verification date — for the other shock. Kimi K3’s open weights and technical report are promised for July 27. That is Monday. The design-shock narrative either becomes reproducible fact or stays a demonstration, and the K3 trilogy (Part I · Part II · Part III) is the framework we will score it against. For this book the question stays the same: does abundant design arrive at Asia’s physical layer as more tape-outs and more wafer starts.

Print season, week two — a demand-side referendum on the Asian slice. The names reporting Wednesday and Thursday — Microsoft, Meta, Amazon, Apple — are, for this book, not stocks: they are the four budgets from which the hyperscaler dollar flows to Asia’s fabs, memory lines and packaging floors. Their capex guides are the demand side of the entire flow map this letter is built on. TSMC has already answered once this season (capex raised to $60–64 billion); this week the customers themselves confirm or cut. Every Asia-linked position in the book is, indirectly, long those four guidance sentences. Cadence prints Monday, one session after the K3 weights drop — the US letter carries that print; this book only needs its echo at the physical layer.

Oil: watch the back of the curve, not the front. Spot traded through $100 this week; the December contract sits near $80. That 20-dollar gap is the market saying temporary — a supply-and-risk premium, not a new price regime. The crisis signal would be the back of the curve rising to meet spot. Until then, the oil chart is an inflation-odds input (see the hike repricing above) rather than a growth-shock input. December near $80 is the line this book watches.

Read more

On Wednesday night ServiceNow beat on every metric and management framed the company as the orchestrator of AI in the enterprise — and for one after-hours session the market paid the claim, the first bought software beat after three sold ones, before Thursday’s discount-rate flush confiscated it.

Read more

The most revealing number of the four-print night was not Alphabet’s capex guide. It was a pair of numbers inside IBM’s release that point in opposite directions. Z mainframe revenue fell 42%, dragging the company to a revenue miss and a trimmed 2026 growth outlook of 4–5%. In the same quarter, IBM’s distributed infr…

Read more

For two years the AI trade rewarded a story. Announce the datacenter, name the model, raise the capex line, and the stock followed. That reflex is gone. The market has quietly switched from paying for AI ambition to demanding a credible path to monetize the spend — and the tape has already handed out grades, well be…

Read more

It took one week. Kimi K3 landed on the 16th, the semiconductor complex had its worst week in over a year, and last night the first institutional buy-the-pullback notes hit client inboxes. UBS’s trading desk — Michael Romano, who runs hedge-fund equity derivative sales — argues the momentum unwind may be nearing its end: hedge funds …

Read more

Kimi K3 was released on Thursday the 16th in Shanghai; by Friday’s close the two companies that dominate electronic design automation had surrendered fifteen billion dollars of market value, and by this Friday’s close the damage had spread far past its origin: the Rubin Build layer finished the week down 9.3%, the semiconductor …

Read more

On Friday the market ran the entire distance to the June capitulation zone, touched it to the dollar, reclaimed everything by lunch — and gave most of it back by the close. This is what a verdict week looks like when both sides get their day.

Read more

Last week we traced the hyperscaler dollar to Asia and concluded the ex-US rally is a concentrated bet on the two countries that collect it. This week the market tested that bet with everything it had: the largest Tokyo unwind of the year, a Chinese AI model that claims to design…

Read more

Tokyo answered Thursday’s New York session with the week’s heaviest tape: the Nikkei fell more than 4,100 points at the afternoon worst — down 6.2%, under 63,000 — and closed off 4.7% at 63,674. The epicenter was the same one it has been all week: memory. Kioxia dove as much as 16%, and its market value has now roughly …

Read more

TSMC’s Q2, released at 08:00 CET: revenue of $40.2 billion, up 33.7% year over year and at the very top of the company’s own $39.0–40.2 billion guide. EPS of $4.31 against estimates near $3.90.

Read more

Last edition’s thesis was stated plainly enough to be wrong: the absolute flow scores still crowned the semiconductors, but the rate of change said leadership was already passing — from the build-out core to the opex layer and the applications. One week later the tape produced receipts. IBM pre-announced a miss and fell 25% in a day — the steepest drop in decades — for exactly the two reasons this letter tracks: memory costs eating legacy IT budgets, and AI spend cannibalizing the incumbents. On the same tape, the AI-opex layer set records: CrowdStrike up 12% and Okta up 11% to new highs, the whole security line green, Cloudflare closing at a record through its 280 line.

The software index barely moved while an 18-point spread opened inside it. That is not a market getting risk-off about software. That is a market re-sorting who owns software’s future — precisely the hand-off the deltas flagged. The read below updates every tell we named, takes this morning’s ASML print into the frame, and gives the earnings season’s referee metric — the Earnings Dispersion Ratio — the treatment we promised. The schedule of our other publications is at the end.

1 · This Week’s Action

The cross-asset backdrop. June CPI printed a 0.4% monthly decline against a consensus −0.2%, taking the annual rate to 3.5% from 4.2% — and the day after held the move instead of fading it: volatility compressed toward its 52-week low, gold firmed, growth led value again. No haven bid, expectations re-anchoring — the friendliest possible macro floor under a rotation.

The tech shelf. Two days that belong in a textbook: Monday the semi complex fell 4–5% while software held green — SOXX closing forty cents below the 554 line — and Tuesday the complex reversed hard, SOXX +2.6% back to 567.92, SMH reclaiming 600, with the epicenter itself producing the violence:

SK Hynix from two dollars above its $149 offer price to a post-debut high, up 27% in one session, on hard news — 12-layer HBM4 in mass production and shipping to Nvidia. The shelf’s message: the washout was positioning, not thesis.

The factor view. The same re-sort one level up: the growth complex caught its bid back, but the year-to-date column still shows the concentration story — the growth factor carried by an ever-narrower list. Hold that thought; it becomes the EDR section below.

Our own board, updated — and the tells resolved. The Rotation Ledger below is the Closelook Directional Flow scan across all four growth buckets, sorted by the 21-day flow delta. Read it against last week’s version and every tell we named has an answer:

Three resolutions worth naming. The opex persistence held: a second consecutive week with AI-opex the strongest advancing bucket — the security-and-observability cluster (Fortinet, Palo Alto, Datadog, now Okta and CrowdStrike with them) exactly where we said to watch it.

The reversal cluster grew: eighteen of forty agentic winners now flash “reversing-up,” from sixteen — and the bellwether row converted: Duolingo’s 21-day delta is now +44 with flow still deeply negative, the textbook shape of a destroyed name being re-accumulated. And the one warning fired too: KLA — last week’s “single number to watch” — rolled from “stalled” to “accelerating down,” the most negative delta in the scan by a factor of three. One core equipment name is repricing, not resting. The complex bounced without it.

2 · The State

The four buckets, briefly. New readers: we cut the growth trade into four functional indices — Rubin is the AI build-out (~125 names of silicon, memory, packaging, power), AEI the AI-opex layer (~33 names that sell the running of AI — security, observability, data), AW40 the agentic winners (40 application names, the cohort 2026 punished hardest), and HALO functional growth beyond AI.

The year so far: HALO led early, the build-out took the year, the winners were wrecked, the opex layer compounded quietly.

The week the buckets swapped. Two days into this week the scoreboard reads: AEI +1.7% — the only bucket green — after leading Tuesday outright at +2.1%, its best day against the field all summer. Rubin −1.8% after a violent round trip (−3.3% Monday, +1.6% Tuesday). HALO −1.1%, resting. AW40 −0.1% — and read that flat number correctly: the destroyed cohort held its ground through a semi washout that took the leaders down 5%, because money is quietly returning to it. Year-to-date the ladder still reads roughly Rubin +114%, AEI +49%, HALO +4%, AW40 −21% — but the week’s tape is the rotation happening, not pending.

The receipts, in names. Monday’s IBM warning is the cleanest single datapoint the rotation thesis has received. The company named its own killers: customers front-running memory and server price rises — the DRAM squeeze arriving in enterprise budgets — and AI spending cannibalizing legacy IT. Inside the same print, Red Hat grew 11%. The next session completed the sort: ServiceNow −5.8%, Adobe −4.3%, Salesforce −2.1% on the legacy side; CrowdStrike +12% and Okta +11% to records, SentinelOne and Zscaler up 7%, Palo Alto up 7%, Cloudflare through 280 to a record close on the other. The software index — IGV, up 1% — hid an 18-point spread. Every one of those record-setters is an AEI constituent. The index was built for precisely this sort.

The structural read. Our working QQQ count stands: the advance off the late-2022 base as Wave 1 into early 2025, the spring-2025 break as Wave 2, and an ongoing Wave 3 — the stretch where pauses get bought and rotations run internal rather than terminal. Monday-to-Tuesday was that thesis in miniature: a two-day washout bought back inside 24 hours.

Below the index, last week’s two pattern-watches resolved in opposite directions. The SMH diamond — usually a topping shape — is dissolving upward: the complex reclaimed 600 and the base at 582 never broke, with SOXX’s 554 line surviving its test by forty cents and reclaiming fourteen points the next day. The IGV throwback held and paid: the broken downtrend retested from above was the launchpad for the record-setters inside it.

3 · The Outlook

This morning’s print belongs to the frame. ASML — the toolmaker whose order book referees the whole build-out — beat its own guidance on every line (€9.3 billion sales against €8.4–9.0 guided, 54% gross margin against 51–52%) and, for the second consecutive quarter, printed no bookings number at all. In its place: a plan to add 30% to this year’s EUV capacity for 2027, and a Q3 guide stepping to €11–12 billion. Amsterdam opened it around 7% higher and held most of it. The market accepted machines as the order book’s proxy — you do not build 30% more litho capacity against demand you cannot see.

For the rotation read, that is the constructive configuration: the build-out’s fundamentals keep confirming while its crowded positioning rests — a trade digesting, not dying. TSMC reports tomorrow morning; tools and wafers could confirm each other inside one week.

The Earnings Dispersion Ratio — the promised treatment. The US letter introduced it as a cliffhanger; here is the number in full. The EDR divides aggregate index earnings growth by the median constituent’s growth. Consensus has S&P 500 earnings growing 23–24% this year — with normal beat rates, the second quarter alone could print near 29%, growth the index normally only achieves exiting a recession. Goldman’s strategists carry roughly $340 in index earnings for the year. But the aggregate is profit-dollar-weighted: the Mag7-and-adjacent cohort compounds near 30% while the rest of the index grows perhaps 5%; equal-weighted, the market’s earnings grow 9–13%; the median company, 7–10%.

Divide, and the index’s earnings line is growing about three times faster than its representative member — an EDR around 3.0. The composition underneath: perhaps 50–75 genuine fast-growers, roughly 250 companies in an unremarkable middle, and 150–200 growing not at all or shrinking.

And a slice of the numerator is not operating earnings: Amazon booking valuation gains on its Anthropic stake — Alphabet likely similar — and a build-out in which one side of every dollar is booked as 100% revenue today while the other side depreciates over years. The first real-time reading arrived Tuesday: all five big banks beat — JPMorgan, Wells Fargo, Bank of America, Citi, and Goldman with the largest surprise — and four of the five sold off on the news. Beats are expected now; the bar is the whisper. That is what an EDR near 3.0 feels like from inside earnings season: an index that cannot afford ordinary good news.

Why this letter cares. Concentration is not a short signal — the ratio can stay elevated for years while the fast cohort genuinely compounds. It is a fragility gauge: the higher the EDR, the more the index’s earnings line depends on a handful of prints, and the more each disappointment costs. This week’s IBM day is the mechanism in miniature — when a mega-cap’s earnings break, the repricing is instant and total. We will track the ratio through the season; if aggregate growth converges toward the median (the boom broadening) the growth tape gets safer; if the median converges down toward zero while the aggregate holds, the costume slips.

The macro frame underneath is doing its part — the inflation impulse cresting rather than entrenching is what lets the multiple survive the autumn calendar:

🔒 Be Informed — the rest is for paying subscribers.

Read more

The split inside tech stopped being a divergence and became a decision. Monday’s session sent the semi complex down 4–5% — SMH −4.2%, SOXX −4.8%, the fabless basket −4.8% — while software rose against it: IGV +0.3%, cloud +1.5%, Palantir +2.6%, Datadog +1.0%, Cloudflare green. Two cohorts, opposite directions, and both landed exactly on …

Read more

Japan’s finance minister has urged the country’s giant pension funds — including the roughly $1.8 trillion GPIF — to invest more at home. Markets immediately understood the subtext.

So what

Traders read the proposal not simply as pension policy, but as a form of “stealth intervention”: redirect more Japanese savings into Japanese assets, c…

Read more

Earnings season opens Tuesday against the highest expectations bar of the cycle — and the market just spent a week pricing in the good outcome before seeing a single print.

The signal has two layers. The narrow one: the numbers themselves — the first AI-compute bellwethers, ASML on Wednesday and TSMC on Thursday, print before any US heavyweight opens its books. The broad one, and the more important one: the market’s reaction function. How does the tape treat this year’s leaders when they print above consensus — and when they merely match or fall short, with guidance the focus either way? And the mirror image: how does it treat the laggards, software above all, on their numbers? The prints answer one week. The reaction read answers the month: whether the leaders still have room to run, and whether the laggards have absorbed all their bad news — or not.

This chart is the tape-side referee: SOXX at 581, sandwiched between the 554 floor it reclaimed and the 590 shelf it hasn’t.

2 · What it is

Two weeks ago the semiconductor complex repriced 7–12% in three sessions. This week the market took the other side: SOXX +2.7%, SMH +3.2%, fabless +5.1%, Nvidia +8.3% back above its 200-day — while last week’s hiding places gave the money back. Software slipped (IGV −1.2%), and the prior week’s rotation winners reversed hard across our own universe: HALO‘s defense cohort −9.4% and space −9.9% after leading everything at +18/+17, the Agentic index’s identity tier −6.1% after +14.7%. The dip was bought, the hedges were sold — before the evidence.

And the tape narrowed while it happened: the S&P 500 gained +1.4% but the equal-weight version fell −0.3%, the Dow −0.4%, small caps −0.5%. One theme carried the index into its own examination week.

The bar being examined, in three numbers:

The dispersion is the story: the index-level bar is carried almost entirely by one theme's capex cycle.

3 · Why it matters

A +24% consensus year is post-recession-sized — the kind of number that stands at the start of a profit cycle, not five years into one. When the forward bar crosses roughly 20%, the market has priced flawless execution: small misses and soft guidance get punished out of proportion, and next year’s comparisons turn unforgiving before a single number is reported.

Read more

Last week we drew two lines and said the market would have to choose: SOXX at 554 and NVIDIA at 191. This week it chose — twice. Tuesday broke the first line and looked like the start of a correction. By Thursday the break had reversed, extended, and been paid for with the largest foreign US listing on record. And underneath the round trip sits the real story: where the buying went. Not to the architect — down the stack, to the layers that had not yet repriced. Which happens to be exactly where the AI dollar itself is headed. That is this week’s letter.

1 · This Week’s Action Free

The tape, day by day. Monday was setup. Tuesday was the software day — SOXX broke 554 and our build-out index had its worst session of the series (−5.6%). Wednesday reversed it: SOXX reclaimed the line at 562, NVIDIA rose 3.7% back above 200, Broadcom added 4.8% on the ~$30B Apple deal. Thursday extended: SOXX +3.5% to 581.7 — the second close above 554 — Micron +4.5% to a $1.12 trillion market cap, software green on a semi day for the first all-green session of the four indices since Monday. Friday resolved the week in character: the memory complex took profits into the SK Hynix debut, while the architect had its comeback — NVIDIA +4.0% to 210.96, back above its 50-day, and SOXX closed at 581.3, flat on the day and twenty-seven points above the line it broke on Tuesday. The week’s net: SOXX −0.3%. The round trip was violent; the destination was where it started. That is what a shakeout looks like on a weekly print.

The sector read. Energy led the week (+3.8%) as oil caught a real bid — the “oil isn’t buying the war” read of recent weeks weakened this week. Tech (+1.7%) and communications (+1.4%) carried the rest; every other sector closed red, with materials (−2.0%) and industrials (−1.5%) at the bottom. Narrow, commodity-accented risk-on.

The macro frame. The FOMC minutes put rate hikes back on the distribution as a contingency — the hawkish tail is alive, the 10-year sits near 4.5%, and the long end sold off again into the weekend. Gold eased but held the 4,000 belief line; bitcoin continued its repair (+4% on the week). The rates question is not resolved; it simply was not this week’s headline. Tuesday it becomes the headline.

2 · The State Free

The count, updated. The five-year view puts the week in its structural place. The advance from the late-2022 low counts as a completed five — that was Wave 1, topping in early 2025. The spring-2025 flush was Wave 2. Everything since is Wave 3, ongoing: its own first sub-wave completed into the winter, its second bottomed at the spring-2026 low, and price now rides the upper half of the red channel in the third-of-third — the strongest part of the sequence, if the count is right. Tuesday’s break never seriously threatened the channel midline, let alone the lower rail; the weekly stochastic has reset to mid-range rather than overbought. The bearish alternative — a completed five at the high — needed the break to stick, and it lasted one session. The channel rails are the count’s own invalidation levels: the midline is where the bulls get their first real question, the lower rail is where the count dies.

The semi complex — the group that decided the week. The four semiconductor wrappers tell the shakeout story in one frame: Tuesday’s break, Wednesday’s reclaim, Thursday’s extension — and a Friday that held the ground. The fabless sleeve (SMHX, +3.2% on the week) led the complex out while the broad wrappers ended flat: break → reclaim → extend is the signature of forced selling absorbed by size, not the start of distribution.

Inside the buying — down the stack. The detail that separates this reversal from a simple bounce: on the strongest day of the week the architect sat out. NVIDIA closed red on Thursday while the layers beneath it re-rated — Storage +6.9%, Deposition & Etch +6.1%, Advanced Packaging +5.7% inside our build-out index, and Micron closed the day worth $1.12 trillion, roughly ten times its 52-week low, still near 6.6x forward earnings. Capital did not chase the top of the stack; it bought the layers that had not yet repriced. That broadens the trade rather than narrowing it.

Where the dollar itself is going. The down-the-stack buying is not a one-week quirk — it mirrors the migration of the AI dollar itself. Last week’s Global edition traced the geography: of every $100 the hyperscalers spend, roughly $45–55 stays in the US — but what stays is the design and IP margin, the cloud economics and the concrete, while the physical system bill runs through the supply chain.

The forward path matters more than the snapshot. Generation by generation — Blackwell today, the Vera Rubin platform entering the market this half, Rubin Ultra scaling toward 576-GPU domains, Feynman behind it — our estimate of the rack dollar shifts the same way the week’s tape did: compute’s share falls from roughly 43% toward ~31%, total memory rises from ~29% toward ~39%, and networking stops being an accessory and becomes part of the processor. The buyers who re-rated storage, etch and packaging on Thursday were, knowingly or not, front-running that table.

3 · The Outlook Free

The four indices — the week in one line each. Rubin Build-Out: −5.6% Tuesday, +3.9% Thursday, −0.1% on the week — a round trip that proved the holder base. The Agentic Ecosystem Index led into Thursday and gave back 3.3% on Friday into the debut, closing the week +1.4%; AW40 added +1.7%; HALO lost −2.6%. Year-to-date the ladder reads Rubin +119%, AEI +47%, HALO +5%, AW40 −21% (equal-weight variants). The barbell swapped sides twice and ended the week near balance — a tug-of-war, not a regime change. Friday’s split inside the pair — the architect +4% while the agentic complex gave back — is the same rotation §2 describes, running in real time.

The road ahead — our working map, and the number that will decide it. Here is the path we are working with, stated plainly enough to be wrong. The tape keeps rising into earnings season — mid-season, perhaps to its end, at the outside until the July-quarter reports land in August. Then the calendar turns hostile: the seasonally weakest stretch of the year meets midterm-election uncertainty, and by then the peak in earnings expectations may already be behind us. Consensus has S&P 500 earnings growing roughly 23–24% this year, and with normal beats the second quarter alone could print near 29% — growth rates the index normally only achieves coming out of a recession. If that is the expectations peak, the autumn setup is a correction — and after the election, the year-end rally.

But the 24% is not what it appears to be. The aggregate is profit-dollar-weighted: a small cohort of AI, energy and capital-markets names growing 25–40% pulls the headline while the median S&P 500 company grows perhaps 7–10% — a respectable expansion, not a boom. The index’s earnings line is growing roughly three times faster than the representative constituent’s. And part of the aggregate is not operating earnings at all: when Amazon books gains on its Anthropic stake — Alphabet likely the same — valuation marks become profits, and inside the AI build-out itself one side of every dollar records 100% as revenue today while the other side spreads its cost across years of depreciation. Earnings booms built on that arithmetic can run for a long time — and they can also unwind through the income statement rather than the multiple.

We will track this through the season with one number: the Earnings Dispersion Ratio — aggregate index earnings growth divided by median constituent growth. Around 3.0 today, by our working estimate. What that ratio does from Tuesday’s bank prints onward tells us which regime this actually is: an earnings boom — or a concentration wearing one as a costume. The paid half of this letter maps the levels, the bellwethers and the book; the EDR gets its own treatment in the coming editions.

The road ahead — our working map, and the number that will decide it. Here is the path we are working with, stated plainly enough to be wrong. The tape keeps rising into earnings season — mid-season, perhaps to its end, at the outside until the July-quarter reports land in August. Then the calendar turns hostile: the seasonally weakest stretch of the year meets midterm-election uncertainty, and by then the peak in earnings expectations may already be behind us. Consensus has S&P 500 earnings growing roughly 23–24% this year, and with normal beats the second quarter alone could print near 29% — growth rates the index normally only achieves coming out of a recession. If that is the expectations peak, the autumn setup is a correction — and after the election, the year-end rally.

But the 24% is not what it appears to be. The aggregate is profit-dollar-weighted: a small cohort of AI, energy and capital-markets names growing 25–40% pulls the headline while the median S&P 500 company grows perhaps 7–10% — a respectable expansion, not a boom. The index’s earnings line is growing roughly three times faster than the representative constituent’s. And part of the aggregate is not operating earnings at all: when Amazon books gains on its Anthropic stake — Alphabet likely the same — valuation marks become profits, and inside the AI build-out itself one side of every dollar records 100% as revenue today while the other side spreads its cost across years of depreciation. Earnings booms built on that arithmetic can run for a long time — and they can also unwind through the income statement rather than the multiple.

We will track this through the season with one number: the Earnings Dispersion Ratio — aggregate index earnings growth divided by median constituent growth. Around 3.0 today, by our working estimate. What that ratio does from Tuesday’s bank prints onward tells us which regime this actually is: an earnings boom — or a concentration wearing one as a costume. The paid half of this letter maps the levels, the bellwethers and the book; the EDR gets its own treatment in the coming editions.

Read more

This week’s edition of Closelook@Global Stock Markets is dated July 11, 2026.

Last week, we showed you that the ex-US rally is six regions wearing a twenty-six-region costume — a concentrated AI-chip bet led by Taiwan and Korea. This week, we follow the money to show why it looks that way: we trace what happens to every $100 the hyperscalers spend, how much of it lands in Asia, and which two buckets within the AI rack account for three-quarters of the bill. The market spent the week drawing the same map — and concluded by taking Korea’s champion to New York.

1 · This Week’s Action

The cross-asset backdrop. The board reads risk-on with a commodity accent. Oil led the week (USO +4.0%, now +57% year-to-date) — the “oil isn’t buying the war” read of recent weeks got at least a partial rebuttal. Bitcoin continued its repair (IBIT +4.2%), and the mega-cap concentrates outran the broad market (TOPT +2.1%, QTOP +1.7%, against SPY +0.8%). On the other side, gold eased 0.8% but held comfortably above the 4,000 belief line, and the long end sold off again (TLT −1.1%) after the hawkish minutes. The rates question is not resolved — it is just not this week’s headline.

The global sectors. Energy led (+3.8%) as oil caught its bid, followed by telecom (+2.3%), tech (+1.3%), and financials (+1.1%) — while every classic defensive closed red: healthcare −2.0%, staples −1.2%, utilities −0.8%. That is a risk-on rotation inside a barely-moving global tape (VT +0.2% on the week).

The regions. The board answers last week’s question — and inverts it. We asked whether participation would broaden beyond the two chip countries. It did, but not the way a trend-follower would have guessed: the periphery led — Brazil +4.0%, Southeast Asia +3.7%, China +3.4%, New Zealand +3.0%, Argentina +2.6% — while Korea rested (EWY −3.4% after its +89% year-to-date run) and Taiwan sat flat. The year-to-date count moved from six regions beating VEU to seven; Poland joined the club. Breadth is broadening at the margin — one week, one region at a time.

2 · The State

The event of the week — the listing. The demand question got answered three times in one week. SK Hynix — the HBM leader and the cheapest large AI name in the world — sold $26.5 billion of ADRs in the largest US listing by a foreign company on record. The book was covered multiple times over. The deal priced at $149, a 2.9% premium to Seoul’s close, where offerings of this size normally price at a discount. And when the ADRs opened on Nasdaq on Friday, the first trades printed near $170 — with an intraday high of $177 — before closing at $168.31, thirteen percent above the offer. The ticker converts from SKHYV to SKHY on Monday.

The tension worth naming. The same week the primary market paid up, the secondary market sold: Micron gave back part of its trillion-dollar week, and the memory complex traded red into the debut. That is not a contradiction — it reads like capital rotating from the expensive expression of the thesis (Micron at ~6.6x forward) into the cheaper new one (SK Hynix at ~4.8x at the offer, ~5.5x after the pop). The Korea discount is not closing by argument; it is closing by arbitrage.

Follow the dollar. The listing is the visible end of a pipeline worth tracing in full. The four large US hyperscalers are heading toward roughly $650–700 billion of combined 2026 capex — Alphabet guides to $180–190 billion, Meta to $125–145 billion, and Microsoft says it is expanding AI capacity by more than 80% while roughly doubling its data-center footprint over two years. Our working estimate for where that money lands: of every $100, $45–55 stays in the US or wherever the data center is built, $10–20 goes to Europe, Mexico and other regions — and $30–40 ultimately accrues to Asian production and suppliers. For the incremental AI-compute dollar the Asian share is higher, around $40–50 of every $100, because the marginal AI dollar contains more GPUs, HBM, advanced packaging, server assembly and cooling — and less land and conventional construction. At a $675 billion midpoint, that implies an Asian revenue pool of roughly $200–270 billion a year, with a central estimate of around $235 billion. This is an economic-value estimate, not a reported accounting number — hyperscalers do not disclose the geography of their capex, and an invoice paid to Nvidia in California ultimately funds wafers in Taiwan and HBM stacks in Korea.

Taiwan is the largest single destination — 13–18% of total capex, $88–122 billion across foundry, advanced packaging, server ODMs, substrates, cooling, and power. Korea’s share is narrower but more concentrated — 7–11%, $47–74 billion — because it flows almost entirely into the single most profitable bucket: memory. Korean producers accounted for nearly 80% of global HBM revenue in the first quarter (SK Hynix 58%, Samsung 21%, per Counterpoint). Japan sits one or two layers upstream — equipment, wafers, photoresists, substrates — collecting its share when TSMC, SK Hynix, and Samsung spend to expand. China’s manufacturing footprint is larger than its profit capture. Malaysia and the rest of Southeast Asia are the China-plus-one winners.

Inside the AI rack. Cut the same $100 differently — not hyperscaler capex, but the AI system itself — and the concentration sharpens further. Our 2026 central case: memory takes $38 of every AI-hardware $100 (HBM alone is $18–25 of complete rack cost), accelerator and host-CPU compute takes $40 (with the design and IP margin staying largely in the US, and TSMC capturing the fabrication and packaging slice), and everything else — networking, power, cooling, boards, assembly — competes for the remaining $22.

Memory and compute now divide roughly three-quarters of the AI-system dollar between them. That is why the economic leverage sits with HBM suppliers, accelerator designers, and advanced-packaging capacity rather than with conventional server assemblers — and why the memory story keeps forcing itself into this letter. Within the Asian pool itself, the two largest slices are advanced logic fabrication with packaging, and memory:

The concentration check, week two. The Vanguard four-slice comparison keeps score on the same story from the top down: VEU (large-cap-tilted, chip-heavy at the top) finished Friday at +14.2% year-to-date against VSS (small-caps, no mega-cap chip names) at +8.5% — the gap widened to 5.7 percentage points from 5 a week ago. The rally is still being carried by exactly the names the flow framework says are collecting the capex dollar.

The two legs against each other. Taiwan compounded steadily all year; Korea went vertical from May. This week the two swapped character — Taiwan flat, Korea digesting. That is what a hand-off inside a theme looks like, not a top.

The structural read. The three-year log view keeps the story honest: a staircase from the 2023 low — +69% over the window — with every correction holding above the prior shelf, including April’s fast flush. VEU trades above both its 50-day (82.6) and 200-day (76.5) averages with RSI at 52: stretched in price, mid-range in momentum. Our chart terminal’s machine wave read (beta) counts the advance as an impulse now in a fifth-wave extension, with invalidation far below at 52.54 — nothing in the structure argues against the trend. The risk at this altitude is breadth and rate path, not pattern.

3 · The Outlook

The four indices. The week told the whipsaw story honestly. Rubin fell 5.6% on Tuesday, recovered 3.9% on Thursday — the first all-green day since Monday — and finished the week at −0.1%: a round trip. The agentic side led into Thursday and then gave back 3.3% on Friday, into the debut, ending the week at +1.4% (AEI). AW40 added +1.7%, HALO lost −2.6%. Year-to-date the ladder reads: Rubin +119%, AEI +47%, HALO +5%, AW40 −21% (equal-weight variants). The barbell swapped sides twice and ended the week close to balance — this is a tug-of-war, not a regime change. Friday’s give-back on the agentic side is the same rotation §2 describes: capital moving toward the cheaper expressions of the same theme.

Read more