Everything About AI
NextFin News - The semiconductor sector is staring at its longest losing streak in 18 years, and this time the selling is not about AI demand failing. It is about the bill for that demand arriving in a higher-rate world. The iShares Semiconductor ETF logged eight straight daily losses, its longest skid since October 2008, according to market data, as a hotter-than-expected inflation print collided with fresh doubts over who ultimately pays for the artificial-intelligence buildout. The Philadelphia Semiconductor Index fell 3.1% on Wednesday alone. The setup is a collision of two forces the market had kept in separate boxes: a rate shock that compresses valuations, and a cost squeeze that compresses margins.
The Sell-Off Ran Fast, and It Ran Wide
From the October 1 close of 576.33 to the October 16 close of 504.04, the iShares Semiconductor ETF fell 12.5% in roughly two and a half weeks. It now trades 23% below its 52-week high of 655.95 and below its 50-day average of 538.11. That is a lot of ground to lose in ten trading sessions, but the more telling detail is how uniform the damage is. Nvidia, the sector's largest company, was essentially flat on the day, yet AMD fell 3.4%, Broadcom lost 3%, Taiwan Semiconductor dropped 1.6%, Micron slid 3.4%, and Intel declined 3.6%. The Nasdaq Composite dropped 1.4% and the S&P 500 slipped 0.7%. When the leader holds still and everything else breaks, the problem is the sector, not the stock.
Every major name now sits at a discount to its own recent average, and the gaps quantify the de-rating. Broadcom trades 9% below its 50-day average, Micron 8.5% below, Intel 11% below, and Taiwan Semiconductor 4% below. Nvidia is the exception that proves the rule: it sits only 3% below its average, and AMD is actually above it. The market is not rejecting AI. It is repricing everything that depends on AI converting into profits without friction.
The memory segment has been hit hardest. The Roundhill Memory ETF entered a second bear market in 2026, falling more than 20% from its September rebound high and sitting roughly 40% below its June record close. Samsung Electronics, Micron, SK Hynix, and Sandisk, all tied to DRAM, have shed more than $700 billion in combined market value since October 6. A broader basket of 61 semiconductor and storage stocks has lost nearly $2.5 trillion. This is a repricing of the entire AI supply chain, not a correction in a single name.
Context matters, because the size of the move depends on where you start measuring. The sector did not get here from a standing start. From its 52-week low of 260.44 to its 52-week high of 655.95, the semiconductor ETF climbed 152%. What has been given back so far is 23% of that peak. The index is still up 94% from its low. A streak that reads like a crisis in isolation looks different against that backdrop: it is a trimming of an enormous rally, not the erasure of one. That does not make it painless for anyone who bought near the top, but it does frame the question correctly. The market is deciding how much of a 152% advance was justified by fundamentals and how much was justified by cheap money. The answer is landing somewhere in between.
The Rate Channel: Why Chips Take the First and Hardest Hit
The trigger was the September consumer-price report. Headline inflation rose 3.7% year over year, above the 3.6% forecast and August's 3.4%, with monthly CPI up 0.6%. Core prices, excluding food and energy, rose 2.6%, above the 2.5% estimate. Traders immediately repriced the Federal Reserve's path: the odds of a rate increase at the October 28 meeting jumped from 22% to 41.5%, and markets began pricing two more hikes in 2026.
That matters disproportionately for semiconductors. The sector's valuation rests on earnings expected years out, because the payoff from today's AI capital spending lands in 2027 and beyond. When the 10-year Treasury yield rises, the discount rate applied to those distant cash flows rises with it. The 10-year yield is trading at 5.13%, up from a 50-day average of 4.94%, and each basis point of additional yield pulls present value forward. A dollar of earnings expected in 2028 is worth materially less today when discounted at 5.1% than when discounted at 4%. Chips are long-duration assets wearing growth-stock multiples. When the risk-free rate moves, they move first and hardest.
The mechanism is mechanical, not psychological. A growth stock's value is the sum of cash flows stretched far into the future, each one divided by a discount factor that compounds with time. Raise the rate and the farthest cash flows, the ones that carry the AI story, shrink the most in present-value terms. A mature utility with next year's dividend does not feel it. A semiconductor company whose investment thesis depends on 2028 data-center demand feels it in full. This is why the sector can post record revenue and still fall: the numerator, earnings, is growing, but the denominator, the discount rate, is growing faster in the market's estimation.
This is not the first time the mechanism has played out. In 2022, the last time the Fed hiked aggressively, the semiconductor ETF fell about 40% on multiple compression even as semiconductor revenue grew. The current drawdown from the 52-week high is 23%. The 2008 comparison that gives this streak its name belongs to a different world: then, demand itself collapsed and the index fell roughly 70% to 80% from peak to trough. Today's earnings are expanding, not contracting. That is the difference between a cyclical de-rating and a structural breakdown, and it is the difference the market is struggling to price.
The cyclical call is that this is rate-driven multiple compression, not a demand collapse. The evidence is in the order books. ASML, the world's largest supplier of chipmaking equipment, raised its full-year sales outlook for the third consecutive quarter, to between 44 billion and 45 billion euros, after reporting record third-quarter sales of 11.66 billion euros, up 55% from a year earlier. Net income rose to 4.96 billion euros and gross margin improved to 57.5%. Taiwan Semiconductor's September revenue was up 54.6% year over year, at 511.86 billion New Taiwan dollars. The demand is not breaking. The multiple is.
The Second-Order Squeeze: The Pass-Through Loop That Defeats Itself
The first-order story is rates. The second-order story is margins, and it is the one the market has not finished pricing. Apple raised prices on some products to offset surging memory costs, and Microsoft raised prices on its current-generation Xbox consoles. That is the transmission mechanism in plain sight: memory makers' pricing power is being passed through to consumers, and the market is asking what happens to demand when it is.
Here is the tension. Strong memory demand is good for Micron, SK Hynix, and Samsung. But if higher DRAM and NAND prices choke off handset, PC, and console demand, the same pricing power that lifts memory revenue today can erode the unit volumes that sustain it tomorrow. The Roundhill Memory ETF's 40% drop from June's record is the market starting to ask that question. The irony is sharp: the AI boom pulled memory supply into data centers, tightening supply for everything else, and now the resulting shortage threatens the consumer-electronics demand that still makes up the bulk of the industry. It is a self-defeating loop, and it is why a sector with record order books can still lose a quarter of its value.
The demand side has not rolled over, which is what makes this a de-rating rather than a recession. Industry forecasts put 2026 semiconductor revenue up 94.1% year over year, with memory expected to account for more than half of total industry revenue. The U.S. high-bandwidth memory market alone is projected to grow from $1.43 billion in 2026 to $4.91 billion by 2031, a 27.98% annual pace. Bottlenecks in HBM and advanced packaging are expected to persist until at least 2027. In other words, the physical shortage is real and it is not ending soon. The stock market is simply deciding that scarcity at the component level does not guarantee profit at the shareholder level when the customer is being priced out.
There is a third layer to the margin question that investors are only beginning to price. If Apple and Microsoft can pass costs through, their customers pay. If they cannot, the chipmakers absorb the hit. Either way, someone's margin shrinks, and the market is rotating out of the names with the least ability to defend theirs. That is a subtler read than "AI is over," and it is more useful. It explains why the selling is broad but not blind.
The Fragile Link: Circular Finance
The most uncomfortable question is who pays for the AI buildout, and with whose money. Analyst William Mortonson put it directly: "I think the street is tiring of the Bank of Nvidia, the Bank of AMD and the Bank of Broadcom." He was describing the scrutiny now falling on chipmakers that effectively finance their customers, selling chips on credit-like terms to cloud and AI startups whose free cash flow cannot yet cover the bill.
CoreWeave, Nebius, and Oracle sit near the bottom of the technology group on free cash flow, and all three are highly dependent on financing going forward. When the cost of capital rises, the circular structure wobbles. CoreWeave shares are down roughly 56% from their 52-week high of 141.79, and the stock fell as much as 9.5% in early trading on Wednesday before recovering. The point is not that AI demand has vanished. It is that the financing chain linking chip revenue to end-user demand is longer and more leveraged than the bull case assumed, and leverage is the first thing a higher-rate regime reprices.
The circular-finance concern is the bridge between the two halves of this story. Higher rates do not just compress multiples directly. They also raise the cost of funding the very customers whose chip purchases justify the multiples. A higher discount rate and a weaker customer balance sheet are the same shock arriving through two different doors. That is why the sell-off has been both fast and hard to buy: every dip buyer is asking whether the earnings supporting the price are real revenue or financed revenue.
The Counter-Thesis: A Buy-the-Dip, Not a Regime Break
The strongest argument against the bear case is simple: the demand is real. ASML raised guidance for the third consecutive quarter. Taiwan Semiconductor is running at record revenue. Memory demand is outstripping supply through at least 2027. Bank of America has called the AI buildout the midpoint of a decade-long transformation. The sector is being punished for a business breakdown that is not happening.
That argument is powerful, but it answers the wrong question. The question is not whether AI demand is real. It is whether today's prices already assumed a flawless execution path. They did. The sector had rallied to record highs on the assumption that capital spending would convert to earnings without friction, no rate shock, no margin squeeze, no financing gap. The streak is the market removing that assumption.
The bull case also understates how much good news was already carried in the price. When an index is up 152% from its low, perfection is not a high bar; it is the baseline. Any deviation, a hotter inflation print, a customer that cannot pay, a component shortage that kills unit volume, reads as a disappointment because the alternative was never priced. The market is not discovering that AI is fake. It is discovering that AI is expensive, and that discovery is a valuation event, not a technology event.
What Comes Next: Scenarios, Not a Single Line
The verdict is a cyclical de-rating inside a structural uptrend. The rate shock and the positioning unwind are mean-reverting forces; when inflation data cools or the Fed signals a pause, the multiple can expand again. The structural leg, AI infrastructure demand, ASML's backlog, and TSMC's pricing power, remains intact. But intact demand does not mean unchanged price. The two can coexist, and right now they point in opposite directions.
The base case is that inflation cools modestly, the Fed holds in October and signals one more move late in the year, and the streak ends with the multiple stabilizing while 2027 capital spending stays intact. The upside case is a clean inflation print that kills the hike odds; the sector re-rates quickly because the demand story was never broken, and leaders reclaim their averages. The downside case is a re-acceleration: a core PCE print at 2.8% or higher, or a second straight month of 0.4% CPI. Either outcome would turn multiple compression from cyclical into something more durable and open the door to a 2022-style drawdown.
The asymmetry is clear. The exposed are the leveraged neocloud providers and second-tier chip designers without pricing power, the companies whose revenue depends on financing that is getting more expensive. The beneficiaries, if the cyclical view is right, are the companies with backlog visibility and pricing power: the equipment makers, the foundry leader, and the memory suppliers with the most advanced nodes. That is an observation about relative resilience, not a recommendation.
What to watch, in order: the October CPI print; the Federal Reserve's October 28 decision; ASML's and TSMC's commentary on 2027 demand; and whether memory prices keep rising without breaking consumer demand. One signal would break the cyclical thesis: if ASML or TSMC roll back guidance, the demand story is no longer intact, and this streak is the start of something worse.
I think the street is tiring of the Bank of Nvidia, the Bank of AMD and the Bank of Broadcom.
The semiconductor sector is not being punished because AI demand has failed. It is being punished because the market finally noticed that the bill for that demand is coming due in a higher-rate world, and not everyone who ordered chips can pay it.
Explore more exclusive insights at nextfin.ai.
