1. The match

Both goals came early and close together — the 20th and 23rd minutes. He was withdrawn at the interval of a match designed around rotation, which is why the 35-minute figure appears in every account: two goals in that window is an unusually dense return even for an exhibition.

All-Star games are not evidence of much on their own. The reason this one is worth noting is that it sits inside a run rather than standing apart from it.

2. The season so far

The consecutive-scoring run is the number that matters more than the MVP award. Three straight league games with a goal, inside a four-match winning sequence, is the difference between a signing that sells shirts and one that moves a table position. LAFC went from mid-table to second during that stretch.

3. What is at stake on Saturday

The Vancouver fixture on 2 August Korean time is a straight swap for position: both clubs are on 33 points, and the winner takes clear second with a claim on the lead.

There is a second layer to it. Vancouver eliminated LAFC in last season's playoffs, which gives the match a shape beyond the standings. Squad availability had not been published at the time of writing, so how much of that history actually gets settled on the pitch is not yet knowable.

1. What changed

The headline change is the width of the road, not the speed limit. Under JESD270-4, HBM4 doubles the I/O count from 1,024 to 2,048 and independent channels from 16 to 32. Per-stack bandwidth therefore roughly doubles without pushing per-pin speed to a difficult place — a deliberate choice, because signal integrity and power both degrade quickly as pin rates climb.

The second change is less visible and matters more for the industry's structure. The base die that sits underneath the stack and manages data flow moves from a memory process to a foundry logic process. That allows far more capable control circuitry, and it makes memory vendors dependent on foundry partners for a component they previously made themselves. The boundary between memory manufacturing and logic manufacturing is now blurred in a way it was not two generations ago.

2. The numbers

JEDEC sets the standard pin rate at 8Gbps, which yields roughly 2TB/s per stack. SK hynix has stated it achieved operation above 10Gbps; a straight conversion puts that near 2.5TB/s, though no supplier publishes a throughput figure at that speed. Against HBM3E 12-high — 36GB at roughly 1.2TB/s — the per-stack gap is close to a doubling, and suppliers cite power efficiency gains above 40% versus the prior generation.

That gap is the whole argument for the generation. An accelerator's usable throughput is bounded by how fast it can be fed, not by how fast it can multiply — a large language model reads its entire parameter set from memory for each token it produces. Doubling bandwidth per stack moves the ceiling directly. Higher stacks and a logic base die do raise unit cost, and market researchers broadly expect HBM4 to price above HBM3E, but no supplier has published a number.

3. What is still open

Several things commonly reported as settled are not. The 12nm-class foundry process said to be used for SK hynix's base die has not been confirmed by the manufacturer. Pricing is entirely unpublished. The timing for 16-high 64GB parts — allowed by the spec but not yet in volume — remains unannounced, as does the schedule for HBM4E, for which samples have shipped against a 2027 production target.

Heat is the constraint least discussed and most likely to bind. Stacking higher and clocking faster raises total dissipation even as efficiency per bit improves, and cooling design increasingly decides which parts a customer can actually deploy. SK hynix has said it is developing an approach it calls iHBM that lowers thermal resistance by more than 30%; nothing comparable has been detailed publicly by its competitors.

1. What deleveraging is

Deleveraging means reducing debt relative to assets. A company can do it by selling a division or issuing shares. A leveraged fund almost always does it by selling positions, and usually not on a schedule of its own choosing.

The trigger is typically a margin call. A fund borrowing against its portfolio must maintain collateral above a threshold. When prices fall, the collateral value falls, and the lender demands more. If the fund cannot post cash, it sells.

The critical property is that this selling is price-insensitive. A fund meeting a call sells what it can move, not what it judges overvalued. Liquid, widely held large-caps go first — which is why a liquidation drags down exactly the names that look least deserving of it.

2. Why it feeds on itself, then stops

The loop in the middle is why declines from liquidations are steeper than the news justifies. The exit at the bottom is why the rebound is equally disproportionate. Nothing improved. One participant stopped selling.

The 31 July session fits the pattern precisely. Foreign investors bought about 7.2 trillion won net after weeks of selling. Retail investors sold roughly 8.25 trillion won into that bid. Neither side had new information about Korean companies; the supply that had been overwhelming the bid was simply no longer there.

3. What this does not tell you

A liquidation ending is not a valuation signal. It says the technical pressure has lifted, not that prices are correct. Markets that rebound off the end of forced selling frequently give back part of the move once ordinary two-way trading resumes.

It is also difficult to identify in real time. Traders described the July unwind as finished only after the rebound, and whether the book was fully absorbed remains an interpretation rather than a confirmed fact. Estimates of the position's size vary widely enough that no figure is used here. The practical lesson is narrower: when a decline is driven by a seller who has to sell, the relevant question is how much is left, not what the assets are worth.

1. The three tiers

The system runs on felt temperature — the heat index combining air temperature and humidity — rather than air temperature alone. That matters in a Korean summer, where humidity frequently makes 33C feel considerably worse than the same reading in a dry climate.

The duration column is the part most often dropped in reporting, and it is the design decision that matters. Advisory and warning both require a sustained pattern, on the reasoning that consecutive hot days compound physiological load. The critical tier abandons that requirement: a single forecast day at 38C is enough, because at that level the risk is acute rather than cumulative.

2. What to do

For heat exhaustion — heavy sweating, weakness, nausea, headache, but consciousness intact — move to shade or air conditioning, loosen clothing, and give cool water or an electrolyte drink in small amounts.

For suspected heat stroke, the response is different and time-critical. The distinguishing signs are a body temperature above 40C, skin that is hot and often dry rather than sweating, and altered consciousness — confusion, slurred speech, collapse.

• Call 119 immediately. Do not wait to see whether the person improves.
• Move them into shade and remove excess clothing.
• Cool aggressively with ice packs at the neck, armpits and groin, where large vessels run close to the surface.
• Do not give fluids to anyone unconscious or unable to swallow reliably. Aspiration is a real risk.

The distinction matters because heat stroke has a mortality curve measured in tens of minutes. Cooling before the ambulance arrives is not optional first aid; it is the intervention.

3. Where the season stands

Cumulative heat-illness cases reached 1,638 with 12 deaths as of 31 July. Officials described the death toll as running at roughly 75% of the equivalent point last year, against higher peak temperatures — which suggests the alert changes and the cooling-centre network absorbed part of the load, though that is an inference rather than a measured effect.

These totals are provisional. Korean heat-illness deaths are reclassified after cause-of-death review, so the final figure will not be settled until autumn. The temperature record that triggered the first critical alerts is covered in a companion piece published the same day.

1. What was announced

The measure came out of the emergency economic committee rather than a single ministry, which signals both the political weight attached to it and the number of agencies that have to coordinate. Three components matter.

The first is the survey itself. About 190,000 buildings are to be inspected between September 2026 and the end of 2027, sequenced so that the highest-risk structures go first.

The second is money. 1.45 trillion won over five years is earmarked to help operators — mostly small and mid-sized — pay for the retrofits the survey will identify. Without it the standards would function as an unfunded mandate on exactly the operators least able to absorb one.

The third is the standards themselves: tighter requirements for exterior cladding, detection systems and sprinklers.

2. The numbers

The two fires are the reason the package exists in this form. The March fire at an auto-parts plant in Daejeon killed 14 people. The July logistics centre fire in Incheon followed while the response to the first was still being drafted. Korean industrial fire policy has historically moved after casualties rather than before, and this sequence follows that pattern.

3. What is still open

The effective dates are the gap. Announcing tightened cladding, detector and sprinkler standards without saying when they bind leaves the most consequential part of the package undefined. Operators cannot plan retrofits against a rule that has no commencement date, and the 1.45 trillion won cannot be allocated against a schedule that does not exist.

The underlying source document could not be opened, so the figures here rest on four press reports cross-checked against one another rather than on the ministry release. The phased breakdown of survey volume — how many buildings in each half-year — appears in only one outlet and is therefore not reproduced. Anyone relying on these numbers for planning should wait for the official text.

1. What was voted on

The measure was S.J.Res.181, introduced by Senator Kirsten Gillibrand under the War Powers Resolution. Procedurally, the vote was not on the resolution itself but on a motion to discharge it from the Foreign Relations Committee, where it had been referred and where it remains.

This distinction matters for reading the result. A discharge motion is a test of whether the chamber wants to consider a question at all. Failing one does not kill the underlying resolution; it leaves it where it was.

2. The arithmetic

Three Republican defections is a meaningful number. Rand Paul's position on executive war powers is long-standing and predates this administration; Susan Collins and Lisa Murkowski represent a narrower institutionalist strand. Against that, John Fetterman's vote with the majority continued a pattern he has held on Middle East questions since the conflict began.

Ninety-nine votes were cast, so one senator was absent. The official roll-call record was not reachable, so that member is not identified here.

3. What is still open

Democrats have forced comparable votes almost weekly since hostilities began in February. The strategy is less about winning any single vote than about building a record and keeping the authorisation question in front of the chamber. On that measure the 49-50 result is closer than most.

The ceiling on the strategy is structural. Even a resolution that passed both chambers would face a presidential veto, and overriding one requires two-thirds in each — a threshold nowhere near reach at these margins. What a one-vote loss does establish is that the coalition is no longer purely partisan, which changes the arithmetic if the conflict escalates or if public opinion moves.

1. Where each supplier stands

SK hynix said on its second-quarter call that HBM4 mass-production shipments began in the quarter, that second-half volumes will expand, and that long-term supply agreements are in place with roughly ten customers. The stock hit its daily limit on the news.

Samsung got there first in calendar terms, shipping commercial HBM4 in February and stating an intent to more than triple HBM revenue year over year.

Micron has framed its position in rate-of-change terms: its 12-high ramp is running at roughly twice the pace of its HBM3E ramp, with related revenue already past $1bn.

Three different claims, three different metrics — which is itself the point. No supplier is choosing to compete on the same axis.

2. The timeline

The gap between first shipment and volume shipment is where the money is. Being first to ship a qualified part matters for design wins; being able to supply that part at yield decides revenue. Accelerator makers commit a year or more ahead, so the supplier that can guarantee volume takes the socket even if it was second to qualify.

3. What is still open

The HBM4E production quarter has not been disclosed. The company target is 2027, and market talk of an accelerated schedule is not confirmed by anything the company has said.

Contract detail is thinner than the headlines suggest. Official material refers to about ten customers; some reports say eleven with five-year terms. Volume and pricing are not public in any of them. Micron has published nothing on HBM4E timing at all. The technical differences that make this generation matter — the doubled interface and the logic base die — are covered in a companion piece published the same day.

1. What moved

Three forces arrived on the same day, and only one of them was news.

The first was the end of a forced seller. A leveraged fund had been liquidating Korean positions through July, and reports that a larger firm had taken over its book removed the overhang. Nothing about Korean company fundamentals changed; the marginal seller simply stopped.

The second was earnings. SK hynix said on its second-quarter call that HBM4 mass-production shipments had begun and that long-term supply agreements were signed with roughly ten customers. The stock hit its 30% daily limit, its first under the price band in force since June 2015.

The third was the currency. A firmer won raises the dollar value of Korean holdings for foreign investors, which matters when those investors are deciding whether to re-enter.

2. The numbers

The flow direction is the part worth holding onto. Foreigners bought roughly 7.2 trillion won net after weeks of selling; retail investors sold about 8.25 trillion won into that bid. The exact figures differ by outlet — 7.2196, 7.2412 and 7.2499 trillion won all appear — because settlement timing differs, so the round number is the honest one until the exchange publishes its final tally.

3. What is still open

Whether the liquidation is actually complete is an interpretation, not a fact. Brokerages moved quickly to describe the session as a floor-forming event, but no one outside the funds involved knows how much stock remains to be sold. The size of the liquidating book is itself disputed: reports put it above $20bn or near $45bn at peak, which is a wide enough range that no figure is used here.

A one-day move of this size also says something about liquidity rather than value. When a forced seller leaves a market that has already thinned out, prices can travel a long way on modest volume. The mechanics of that process, and why the rebound tends to follow the selling rather than accompany it, are covered in a companion piece published the same day.

1. What happened

The sequence was compressed. On 29 July Yangsan reached 40.3C, the first time any Korean station had touched 40C during July. On 30 July it did it again, which had never happened on consecutive days. On 31 July it reached 41.4C.

The cause was a stacking of two systems. The North Pacific high, which normally dominates the Korean summer, was overlaid by the Tibetan high pushing east at higher altitude. Two warm air masses sitting on top of each other suppress vertical mixing, so heat accumulates near the surface through the night as well as the day. Tropical-night counts — nights where the minimum stays above 25C — reached a record for the season.

2. The numbers

Health effects tracked the temperature closely. Cumulative heat-illness cases reached 1,638 with 12 deaths. Officials described the death toll as running at roughly 75% of the comparable point last year — lower, but against a higher temperature, which suggests the alert system and cooling-centre network absorbed part of the load.

Korea revised its heat alert structure on 1 June this year, adding a third and highest tier. The criteria and the clinical response are covered in a companion reference piece published the same day.

3. What is still open

The death figures are provisional. Korean heat-illness deaths are reclassified after cause-of-death review, and the final number for the season will not be known until autumn. The 75% comparison is therefore a running estimate rather than a settled result.

Whether this is a new plateau or a single extreme is also unresolved. One station reaching 41.4C does not by itself establish a trend, and the overlapping-high pattern that produced it is not guaranteed to repeat. What is established is that the July ceiling moved: a threshold that had never been crossed in that month was crossed three days running.

1. What HBM is

Conventional computer memory sits in modules some distance from the processor, connected by a relatively narrow bus. That arrangement is flexible and cheap, and its bandwidth ceiling is low.

High Bandwidth Memory takes the opposite approach. Between four and sixteen DRAM dies are stacked vertically and connected by through-silicon vias — copper channels drilled straight through each die rather than routed around its edge. The stack sits immediately beside the processor on a shared interposer.

Two things follow. The connection is extremely wide, because thousands of vertical channels can run through a stack where only a few hundred traces would fit on a board. And it is very short, which cuts the energy cost per bit moved. Bandwidth rises by an order of magnitude; packaging complexity and cost rise with it.

2. Why it became the bottleneck

An accelerator's usable throughput is set by how fast it can be fed, not by how fast it can multiply. A large language model reads its entire parameter set from memory for each token it produces. Arithmetic units sit idle waiting for data.

This is why generational HBM specifications track AI capability more closely than transistor counts do.

The HBM4 step is structural rather than incremental. Doubling the interface width to 2,048 bits roughly doubles bandwidth without forcing per-pin speeds into a region where signal integrity and power degrade sharply. Its base die also moves from a memory process to a foundry logic process, which makes memory suppliers dependent on foundry partners for a component they used to make themselves.

3. Who controls supply

Three companies make HBM at volume. On Q2 2025 shipment share, Counterpoint put SK hynix at 62%, Micron at 21% and Samsung at 17%. No fourth supplier is near qualification, and the barrier is not lithography but packaging yield — stacking sixteen dies with thousands of vias each, without a single fatal defect, is a manufacturing problem few can solve economically.

That concentration is the strategic fact. When one supplier holds roughly three-fifths of a component that gates AI accelerator performance, its capacity decisions propagate into the availability and price of AI compute worldwide.

Two caveats on the numbers. Shipment-share estimates come from a single research firm and differ from revenue-based rankings, since generation mix affects average selling price. And no supplier publishes HBM pricing, so any cost comparison across generations rests on market research rather than disclosure.