The KOSPI Just Had One of the Fastest Bear Markets in History. Meta's Earnings Are the Bigger Warning.

Start with Korea, because the chart is almost unbelievable on its own. The KOSPI rallied 297% from April 2025 to its June 22 peak, the strongest bull run for the index since the 1982 to 1989 stretch. It has since crashed 34% in 25 trading days, a decline that Bespoke Investment Group already classifies as slightly worse than the average bear market drop of 33.5% for the index. The KOSPI fell 10.8% on Tuesday to 6,023.66, then dropped as much as 12.6% intraday on Wednesday before trimming losses, the second straight session of double digit swings. Samsung Electronics, the Korea listed chipmaker whose earnings kicked off this month's selling, fell 13.4% on Tuesday. $SKHY, the Nasdaq listed shares of its memory rival, dropped 14.7% the same day and closed below its own IPO price.

The U.S. chip complex caught the same fever. $NVDA alone has shed $238 billion in market value since Friday's close. SK hynix, Samsung and $MU lost $176 billion, $173 billion and $113 billion respectively over the same stretch, and $AMD is down roughly $110 billion, according to a CNBC analysis of FactSet data. The Philadelphia Semiconductor Index has gained 92% over the past twelve months and is still down only about 20% from its own peak, which tells you this is a fast unwind inside an enormous rally, not a collapse from a standing start.

The demand story doesn't explain this

If AI demand were actually breaking, you would expect the companies selling the physical chips to confirm it in their own numbers. They didn't. SK hynix reported second quarter revenue of 79.32 trillion won, with operating profit climbing to 60.54 trillion won from 9.2 trillion won a year earlier, a 557% increase, at a 76% operating margin. Both headline figures still missed Wall Street's estimates of roughly 84 trillion won of revenue and 64 trillion won of operating profit, and the stock sold off on the miss. Analysts put the shortfall down to HBM4 shipments landing below plan, which pushes revenue recognition into later periods rather than erasing it.

$INTC told the same story from the other side of the ledger, and told it louder. Second quarter revenue was $16.1 billion, up 25% year over year and the company's fastest growth in more than fifteen years, against the $14.42 billion analysts expected. Data Center and AI revenue climbed 59% to $6.3 billion. Adjusted earnings came in at 42 cents per share against a 21 cent estimate, and Intel guided third quarter revenue to $15.8 to $16.8 billion when the street had penciled in about $15.1 billion. A beat that size, on the segment that sells directly into the AI buildout, is not what a demand break looks like. The stock sold off anyway.

That gap between what the sellers of AI silicon reported and what their shares did is the tell. This is a repricing of expectations after an exceptionally strong rally, not a verdict on end demand, and the multiple dispersion inside the sector shows it plainly. Priced to the July 27 close, $NVDA trades at 30.1x trailing earnings after rising just 5.4% year to date. $AMD carries 162.4x while up 131% year to date, and $ARM sits at 314.7x. That is not one market view on AI. That is capital rotating out of the most stretched names and into the cheapest ones, inside the same sector, in the same week.

Screen the actual dispersion across the group instead of trusting the headline number: compare Nvidia, AMD, Broadcom and Arm on trailing P/E and year to date return.

The harder question landed tonight, and it wasn't about chips

Memory and logic chips are a real, physical, sold out product. The harder question this earnings season was never whether Korea can ship enough silicon. It's whether the hyperscalers spending hundreds of billions on that silicon can turn it into profit before the bill comes due. $META answered that question tonight, and the answer wasn't reassuring.

Revenue came in at $60.80 billion, up 28% year over year, beating the roughly $60.28 billion analysts expected. Advertising revenue rose 27% to $59.36 billion on 14% more ad impressions and 12% higher average price per ad. None of that mattered once the rest of the print came out. Diluted earnings per share landed at $6.18, missing the $7.22 consensus by more than 14%. Net income fell 14% to $15.85 billion as total costs and expenses jumped 55% to $42.03 billion, including a $2.40 billion legal charge and $1.18 billion in severance tied to the May 2026 layoffs. Operating margin collapsed to 31% from 43% a year earlier.

Then the two numbers that actually explain the reaction. Free cash flow, the line that measures whether the spending is being recouped rather than just depreciated on a schedule, cratered to $784 million for the quarter from $8.55 billion in the same quarter last year, even as quarterly capital expenditures hit $31.08 billion. And the outlook came in light: third quarter revenue was guided to $61 to $64 billion, a $62.5 billion midpoint against a street average closer to $63.2 billion. Management didn't blink on the spending. Full year 2026 capex guidance was narrowed to $130 to $145 billion, and the floor of that range went up, not down, from the prior $125 billion. Shares fell about 10% in after hours trading.

That's the distinction worth sitting with. SK hynix missed against an estimate that had gotten ahead of an already record quarter, on a product with sold out capacity and long term supply contracts. $META missed because spending outran the cash the business is actually generating, on infrastructure that doesn't get resold to anyone else the way a memory chip does. One is a valuation argument on real demand. The other is a cash flow statement saying the bill for AI infrastructure is starting to bite before the payoff shows up.

Run the capex against cash flow math across the hyperscalers on their latest full year figures, rather than taking either side's word for it: compare Meta, Alphabet, Microsoft and Amazon on capital expenditure and free cash flow.

And check where the memory makers' margins actually sit behind this month's chip selloff: compare Micron, SanDisk and Western Digital on gross margin and trailing P/E.

What actually resolves this

$MSFT reported the same night as $META, and its Azure numbers are the next data point on whether cloud capex is still converting into growth the market believes. $GOOGL's own 2027 capex commentary, still pending, is the one that tells you whether the company funding the TPU buildout believes its own bet enough to keep raising it. None of this gets settled by another day of ticker color out of Seoul or another circuit breaker on the KOSPI. It gets settled the way $META just settled its own quarter: in a free cash flow line that either holds up under the spending or doesn't.

Korea's crash is a leverage story unwinding inside a 297% rally. $META's quarter is a margin story that just started showing up in actual cash, not in a chart. Only one of those two is the one the rest of this earnings season needs to answer.