CALVERELL PERSPECTIVES

Semiconductors: A Funding-Risk Correction, Not an Earnings Downturn

· · 24 min read

The Philadelphia Semiconductor Index (SOX) fell 12.4% over seven sessions to 11,581 on 15 October, leaving it 21.0% below its June record.

Conceptual silicon wafer and optical glass on warm ivory limestone in soft natural light.

Semiconductor fabrication — illustrative image.

Why record results failed to stop a $1.6 trillion selloff, and how we would position for the next twelve months

Research reference date: 15 October 2026, US market close (4:00 p.m. ET). Valuation and scenario figures are measured against this date.
Market data updated through: 15 October 2026, US market close, including Korean, Taiwanese and European sessions that day.

Market data and valuation update: 16 October 2026

Executive Summary

The Philadelphia Semiconductor Index (SOX) fell 12.4% over seven sessions to 11,581 on 15 October, leaving it 21.0% below its June record. In the same window TSMC, ASML and Samsung reported or pre-announced record quarters, and TSMC and ASML raised their full-year outlooks. Reported earnings show no downturn. In our view the market has repriced two things: the cost of capital for the companies that buy AI hardware, and how long record memory margins can last. The financing stress is concentrated in debt-funded neoclouds and loss-making AI labs; the large hyperscalers, which still fund most of their capex from operating cash flow, have not cut their plans.

A firmer-than-expected September CPI print sped the move up through front-end rates and leveraged positioning, mainly in Korea. Over the next one to three months we expect a volatile basing process rather than a quick recovery, with a further leg lower if credit stress spreads beyond the neoclouds. Over six to twelve months we see modest upside for the sector, with better odds in foundry, lithography and Nvidia than in memory. The main risk to this view is a cut to 2027 hyperscaler capex plans.

What Triggered the Selloff

The decline was fast and narrow. From the 6 October close to the 15 October close, the SOX fell from 13,217.82 to 11,581.22 over seven consecutive losing sessions, with the heaviest falls on 8 October (−3.4%) and 14 October (−4.3%). Reuters put the loss of US chip market value over the first six sessions at $1.64 trillion. On 15 October the index fell as much as 4.4% intraday, to 11,171.53, before recovering to close 0.9% lower. The Nasdaq 100 lost 4.1% over the same seven sessions and closed almost flat on 15 October. This was a sector repricing rather than a broad de-rating of growth equities.

Table 1: Selected market moves, 15 October 2026
InstrumentCloseLatest sessionFrom 52-week high
SOX index11,581.22−0.9%−21.0% (14,655 on 22 Jun)
SMH (VanEck Semiconductor ETF)$558.84−1.0%−16.8%
Nasdaq 10029,935.59−0.05%−4.5% (record on 6 Oct)
Nvidia$217.23−1.8%−10.8%
Broadcom$344.33−0.6%−30.4%
AMD$562.23−1.5%−14.6%
Applied Materials$430.01−3.7%−41.9%
SK Hynix (15 Oct, KRW)1,262,000−11.1%c. −58% from June peak
Samsung Electronics (15 Oct, KRW)210,000−9.5%c. −20% since 8 Oct guidance

Scroll horizontally to view all columns.

All prices and latest-session changes are for the 15 October session: US-listed instruments at the US close, SK Hynix and Samsung in Seoul. Percentage changes from highs are Ocean Capital Research calculations using intraday highs where available.

We rank the likely causes below. The ranking is our judgement, based on the sequence of events and on which assets moved most. The timing is clear; the relative weight of each driver is not, and several of them reinforced one another.

The price of financing for AI buyers (primary, in our view). On 8 October OpenAI was reported to have told investors its annualised revenue was about $50 billion at the end of September, against the $68–70 billion figure that had circulated. The gap reflected a difference in accounting method rather than lost sales, but it reset expectations for the most important customer in the chain, and the SOX fell 3.4% that day. On 12 October Broadcom's Hock Tan said capital, not chips or power, may become the industry's main constraint. On 14 October CoreWeave's 2032 bonds yielded about 14.8% (FactSet data, as reported by CNBC) and its shares fell 15.5%. On 15 October Reuters reported its 2031 bonds yielding more than 16% at one point, on roughly $50 billion of debt. Five-year CDS spreads remained above 1,000 basis points on 15 October, despite a rebound in the company's shares. So far the stress sits with debt-funded neoclouds and the labs that buy from them, not with the hyperscalers. The market's question has moved from whether demand exists to who pays for it, and at what rate.

Front-end rates after the CPI release (accelerant). September CPI rose 3.7% year on year and core CPI 2.6%, against expectations of 3.6% and 2.5%. The Fed raised its target range to 3.75–4.00% on 16 September, and the minutes showed most participants expecting another increase by year-end, with some citing the AI build-out itself as a source of demand pressure. On CME FedWatch, the implied probability of a quarter-point hike to 4.00–4.25% at the 28 October meeting roughly doubled from about 22% after the release, then eased later in the day, while fed funds futures continued to price at least one more increase by year-end. A much hotter September PPI on 15 October did not lift the October odds further. The 10-year Treasury yield fell to 5.10% on 15 October, from 5.19% the day before and 5.22% a week earlier. On this evidence the long-duration discount-rate story looks overstated; the policy path and credit spreads moved more.

Memory competition and peak margins (segment-specific). Reports that China's CXMT plans a denser server DRAM product by year-end sent Micron down 4.5% on 13 October and Samsung and SK Hynix down 7% and 10% in Seoul. The claim is unverified and commercial yields could be years away. With Micron's gross margin at 86.8%, the market does not need proof to cut the multiple.

Positioning and forced selling (transmission mechanism). The KOSPI fell 7.57% on 15 October, triggering a 20-minute circuit breaker, with foreign investors net sellers of 3.6 trillion won that day. Korean authorities had already acted to curb a boom in leveraged single-stock ETFs, and buybacks at Samsung and SK Hynix that had supported the market are winding down. SMH traded about 15 million shares on 15 October, after 11.7 million on 14 October, more than double its typical September turnover.

Deterioration in reported fundamentals. Not observed. We discuss this next.

What the Earnings Evidence Actually Shows

Every major semiconductor result of the past seven weeks delivered profit at or above expectations and came with a higher or reaffirmed outlook. ASML's sales were fractionally below consensus and Samsung's preliminary revenue slightly short, but in both cases profit and guidance were stronger than expected. That is not the pattern that has usually marked the start of an earnings downturn.

Table 2: Latest reported results and guidance
CompanyPeriodReportedForward signal
TSMCQ3 2026Revenue NT$1,494.2bn (US$46.7bn), +50.9% y/y;
net income NT$763.1bn, +68.7%;
gross margin 66.4%
Q4 revenue US$52.3–54.3bn;
2026 growth raised to about 44% in US$;
2026 capex US$60–64bn
ASMLQ3 2026Sales €11.7bn;
gross margin 57.5%;
net income €5.0bn (incl. c. €0.5bn Mistral AI revaluation)
Q4 sales €14.2–15.2bn;
2026 €44–45bn;
planned 2027–28 supply short of demand
Samsung ElectronicsQ3 2026 (preliminary)Operating profit c. KRW107.4tn on sales of c. KRW195tnFull results due 29 October
MicronFQ4 2026 (to 3 Sep)Revenue $54.2bn;
GAAP gross margin 86.8%
FQ1 2027 revenue guide c. $61.5bn
NvidiaQ2 FY2027 (to 26 Jul)Revenue $96.2bn;
data centre $89.0bn;
gross margin 75.0%
Q3 revenue $108.0bn ±2%
BroadcomFQ3 2026AI semiconductor revenue $16.7bn, +221% y/yFQ4 AI revenue c. $21.7bn;
FY2027 AI target c. $115bn

Scroll horizontally to view all columns.

TSMC's detail matters more than the headline. Advanced nodes of 7nm and below made up 79% of wafer revenue, and 2nm rose to 8% from 3% in the second quarter. High-performance computing reached 68% of revenue. Gross margin fell 1.3 points from the second quarter because of 2nm dilution, and fourth-quarter guidance of 65–67% implies no recovery yet. Management said demand remains above capacity even after checking whether customers have the financing and supporting components to deploy what they order, and it struck a cautious note on consumer and price-sensitive markets in 2027.

The signals that usually precede a semiconductor earnings downturn are guidance cuts, inventory build at customers, shrinking lead times and lower customer capex. We see none of them in compute, foundry or lithography. On our aggregation of company guidance given at July results, combined 2026 capex for the four largest hyperscalers is roughly $720–745 billion, after upward revisions. Fiscal years differ, and so do capex definitions: the totals include finance leases at some companies and spending outside AI infrastructure at all of them, so the figure is indicative rather than a like-for-like measure of AI investment. These companies still fund that spending largely from operating cash flow, although it now absorbs most of it. The debt-dependent part of the build-out sits elsewhere: in neoclouds, AI labs and the vendor-financing structures around them. That is the part the credit market has started to price. Current guidance provides reasonable visibility into earnings over the next two to three quarters. The greater uncertainty concerns their durability beyond 2027.

Why TSMC's Record Quarter Did Not Stop the Selling

TSMC's sixth consecutive record profit arrived on the day the KOSPI fell 7.6%, and its Taiwan-listed shares closed 2.0% lower. We see three reasons for the disconnect.

First, the results answer a question the market has stopped asking. TSMC's revenue reflects wafer orders placed six to twelve months ago. The current doubt concerns whether the end buyers, particularly loss-making AI labs and debt-funded neoclouds, can finance the 2027–28 round of orders on acceptable terms. TSMC's remark that it now checks customers' financing before adding capacity shows management takes the risk seriously; it does not remove it.

Second, a beat was expected and the margin direction was slightly negative. Net income exceeded the LSEG SmartEstimate by about 2%, a modest surprise for a stock that had led the year's rally, and gross margin moved down rather than up.

Third, the marginal seller was not reading income statements. Leveraged Korean retail products, foreign outflows from Seoul and cross-asset risk reduction were setting prices. BlackRock's research arm cut US equities to neutral on 12 October, citing volatility around the AI build-out. Selling of that kind tends to end on exhaustion or a change in the macro or credit backdrop, not on one more good set of numbers.

Strong prints from Samsung, SK Hynix and Nvidia over the next six weeks are necessary for a recovery but probably not sufficient. We think the trigger is more likely to come from credit, through stabilising neocloud spreads and successful refinancing, and from a clearer Fed path.

Illustration of a data-centre aisle lined with server racks
AI compute infrastructure — illustrative image.

Valuation, Positioning and the Cost of Capital

The correction has so far been a multiple compression. On the consensus, holdings-weighted next-twelve-month series we use (compiled by History of Market from terminal data), the SOX forward price-to-earnings ratio was 18.3× on 13 October, down from 21.7× on 5 August, against a post-2011 average of 17.5× and a median of 15.7×. Holding estimates constant, the 15 October close of 11,581 takes the multiple to about 17.4×, in line with its long-run average; the intraday low of 11,172 briefly took it to about 16.8×. On the same basis the Nasdaq 100 trades on about 21×.

An average multiple is not the same as cheap. In semiconductors a low multiple on peak earnings is a late-cycle signal, and the market is applying that logic unevenly. Memory carries the lowest multiples, with Micron at about 12× and SK Hynix at about 6× trailing earnings, because investors doubt that 80%-plus margins will last. A KB Securities analyst described the market as reverting to valuing chipmakers as commodity cyclicals. Compute and equipment still carry growth multiples: Nvidia about 27× trailing earnings, Broadcom about 20× forward and Applied Materials about 36× forward.

The cost-of-capital picture is more specific than "higher yields". The 10-year Treasury yield stood at 5.10% on 15 October, about 10 basis points higher than a month earlier but below its early-October peak, and the 10-year real yield was 2.78%. Yet the SOX rose 18.7% between its 14 September close and 6 October while long yields were climbing. Duration alone does not explain this selloff. The more likely change is in the risk premium attached to AI cash flows that depend on external financing, visible in CoreWeave's bonds and in the reaction to OpenAI's revenue figures. In September the Fed chair noted that competition for capital tied to the AI build-out was one reason Treasury yields had risen. In that sense the build-out is starting to raise its own cost of funding.

In our judgement the market now prices one more Fed hike, a return of sector multiples to their long-run average and a partial normalisation of memory pricing. It does not price a cut to 2027 hyperscaler capex plans, a funding failure at a major AI lab or neocloud large enough to affect order books, or a policy rate path well above 4.25%. The second risk would reach the large suppliers directly. Nvidia holds about 10% of CoreWeave and has agreed to buy any unsold CoreWeave capacity through April 2032. That agreement is a contingent purchase commitment, not a guarantee of CoreWeave's debt, and public disclosures do not allow us to estimate what it would cost Nvidia in a stress case.

Not One Trade: How the Segments Differ

AI accelerators. Nvidia guides to $108 billion of third-quarter revenue and sells to a broad base of hyperscalers, sovereign buyers and labs. Its share price is only 10.8% below its high, and in our view that relative resilience is justified. AMD is more exposed to a single customer through its multi-gigawatt MI450 commitments tied to OpenAI and trades on a far higher trailing multiple. We prefer Nvidia.

Custom silicon and networking. Broadcom has the steepest growth path, with AI revenue targets of about $58 billion in fiscal 2026, $115 billion in 2027 and $230 billion in 2028, but nearly all of it comes from about six customers, and Broadcom now helps finance the largest of them. According to Reuters, which has seen Anthropic's confidentially submitted draft IPO prospectus, Broadcom has agreed to lend Anthropic up to $42 billion in convertible notes, which it may place with a financing partner; Anthropic does not expect any notes to be sold before its IPO. The prospectus has not been publicly filed, but Broadcom's latest quarterly filing describes a facility of this kind without naming the customer.

Separately, banks have begun syndicating a $60 billion package to fund Anthropic's TPU leases: a $42 billion senior secured tranche reported to carry a partial residual-value guarantee from Broadcom, and an $18 billion junior tranche, led by Blackstone, without it. Reports differ on whether the senior tranche and the convertible facility are the same money, so we do not add them.

For OpenAI, Broadcom is reported to be in early talks to arrange more than $50 billion; nothing is signed. Most of this debt would sit with lenders rather than on Broadcom's balance sheet, and a residual-value guarantee supports the value of the leased hardware rather than the loans themselves. Broadcom's direct exposure cannot be sized until terms are published. Its revenue is particularly sensitive to the financing conditions facing its largest AI customers. After a 30% fall from its June high and on about 20× forward earnings, the stock already discounts a good deal of this; it will move furthest in either direction as funding news emerges.

Leading-edge foundry. TSMC combines the best visibility, the strongest pricing and the least dependence on any one customer's financing. It is capacity-constrained at 2nm and in advanced packaging, and management expects 2026–28 capex to be significantly higher than in the previous three years. It is our preferred way to hold AI semiconductor exposure through a funding scare.

Lithography and wafer-fab equipment. ASML says customers are asking about 2029 and 2030 slots and that its planned 2027–28 supply will fall short of demand. Its shares still fell about 5% in Europe on 15 October. Applied Materials is 41.9% below its high after declining to give an industry spending forecast at its investor day. Equipment is late-cycle when capex turns, but today it is supply-constrained. We favour ASML and would wait for 2027 spending commentary from Lam Research (21 October), KLA (28 October) and Applied Materials (12 November) before adding to the broader group.

HBM and conventional memory. Earnings are at records: Samsung's preliminary third-quarter operating margin was about 55% and Micron guides to further growth. Nothing in the reported numbers yet shows a peak. The risk is the classic one. Margins at these levels attract capacity from Micron's new fabs, from Korean expansion and potentially from CXMT, and some analysts already flag oversupply risk later in the decade. SK Hynix's fall of close to 60% from its June peak shows how quickly the market can reprice that possibility. Our caution concerns the durability of the margin rather than evidence of a turn, and low multiples do not offset the risk until contract-price data and fourth-quarter guidance show pricing holding.

Analog and power. These businesses lack a comparable AI tailwind and follow the industrial and auto cycle; consensus expects ON Semiconductor's 2026 revenue to grow about 9%. They are less exposed to AI funding risk but not a hedge against higher policy rates.

How This Compares With Past Downturns

Measured by year-end levels, the SOX fell 59% between 1999 and 2002, 48% in 2008 and 36% in 2022. Each of those declines came with falling industry earnings, which is not the case today. The closest template is 2018, when a memory-price peak coincided with Fed tightening while reported earnings were still strong; the index ended that year down about 8%, with a deeper intra-year drawdown. Today's memory segment fits that pattern most closely.

The 2000 comparison is relevant through one channel only: credit-financed buyers. Then it was telecom carriers funded by vendor loans and high-yield debt; now it is neoclouds and AI labs funded by private credit, bonds and supplier support. The differences are large. The biggest buyers today still fund most of their spending from operating cash flow, and the leading suppliers earn operating margins above 50%. These comparisons rest on a handful of episodes and an index whose composition has changed, so we use them to identify mechanisms, not to set targets.

On this evidence the sector is in the valuation phase of a correction, with multiples falling while estimates still rise. The more damaging phase historically begins when estimates turn down. We see no sign of that in compute, foundry or lithography. Memory is the segment most likely to enter it first.

Semiconductor Outlook: The Next Three to Twelve Months

One to three months: basing, with downside risk. We do not expect a durable recovery before three tests are passed: the FOMC decision on 28 October, hyperscaler third-quarter results in late October, and Nvidia's results in late November. On 15 October the index tested the mid-September lows of 11,100–11,200 intraday, reaching 11,171.53, and closed more than 400 points above that level at 11,581.22. CoreWeave's shares also rebounded that day while its default-swap spreads stayed above 1,000 basis points, so we read the recovery as short-covering and exhaustion of forced selling rather than a change in the funding picture. It is a possible sign of stabilisation, not evidence that a low is in place. Further short-covering rallies are likely in a market this stretched, and we would treat them as technical until credit spreads confirm. A sustained close below the tested zone would most likely require the bear-case triggers below. We do not set a point target for this horizon because the dominant drivers, credit headlines and forced selling, cannot be modelled with useful precision.

Six to twelve months: modest recovery, led by compute and foundry. Our scenarios apply an assumed change in an implied earnings base of about 667 index points, together with a forward multiple. That base is the 13 October close of 12,208.79 divided by the 18.3× consensus forward multiple, a valuation scale rather than a reported index EPS, and we hold it unchanged because no new estimate data were available at the 15 October close. Returns are measured from the 15 October close of 11,581.22. The figures are research-team estimates, not consensus forecasts, and the probabilities are our judgement. The base case carries the most weight because guidance is still rising in compute and foundry. The bear case carries twice the weight of the bull case because credit stress is already visible, whereas the bull case needs capex, refinancing and the Fed to turn favourable together.

Table 3: Twelve-month SOX scenarios (Ocean Capital Research estimates)
ScenarioProbabilityAssumed change in implied earnings baseForward P/ESOX rangevs 15 Oct close
Base55%+10%17–19×12,500–13,900+8% to +20%
Bull15%+20%19–21×15,200–16,800+31% to +45%
Bear30%−15%14–16×7,900–9,100−32% to −21%
Probability-weighted midpointc. 12,210c. +5.4%

Scroll horizontally to view all columns.

The implied earnings base is the SOX level divided by its consensus forward P/E, not a bottom-up index EPS; ranges are rounded to the nearest 100 points. The probability-weighted figure uses range midpoints of 13,200, 16,000 and 8,500; returns are measured from the 15 October close of 11,581.22. The expected value is modestly positive, but the spread of outcomes is wide and skewed to the downside in the near term. That argues for selective exposure rather than a broad sector bet.

Base case. The Fed raises rates once more, in October or December, and then pauses around 4.00–4.25%. Neocloud spreads stay wide but there is no default. Hyperscalers reaffirm 2026 plans and point to higher 2027 spending. Compute and foundry estimates keep rising while memory estimates flatten during 2027. The multiple holds near its long-run average.

Bull case. Hyperscalers lift 2027 capex and fund it mainly from operating cash flow, CoreWeave and its peers refinance successfully, and the Fed signals a pause. Earnings growth and a partial multiple recovery both contribute.

Bear case. Two or more of the largest hyperscalers hold 2027 capex flat or cut it, a large AI lab's funding round fails or is repriced, or credit stress spreads to investment-grade technology issuers. Memory contract prices roll over, and the multiple falls to around its post-2011 median on lower earnings.

Evidence that a low is forming: the sector multiple is back at its long-run average while estimates are still rising; Korea has shown capitulation-style signs, including a circuit breaker, heavy foreign selling and sharp losses in leveraged products; the SOX has fallen 21% from its peak without any downgrade to guidance; and on 15 October the index reversed from an intraday test of the mid-September lows to close well off its low. Evidence that the decline is not over: CoreWeave's default-swap spreads remained above 1,000 basis points on 15 October even as its shares rebounded; futures still price at least one more Fed hike by year-end; leadership remains narrow; and memory multiples have room to fall further if pricing data soften.

Events that could change the call: Lam Research (21 October); Intel (22 October); the FOMC (27–28 October); KLA and SK Hynix (around 28 October); Samsung's full results (29 October); hyperscaler results in late October; AMD (early November); October CPI (10 November); Applied Materials (12 November); Nvidia (late November); Broadcom (December).

Investment Implications and Conditions That Would Change Our View

Our recommended sector weighting is neutral. A probability-weighted return of about 5%, with a 30% bear case implying a fall of more than 20%, does not justify an overweight on the sector as a whole. The risk is unevenly spread, however. The bear case falls hardest on memory and on companies whose revenue depends on externally funded customers, while TSMC, ASML and Nvidia should keep raising estimates in the base case and are the suppliers paid first and hardest to replace. We therefore rate TSMC and ASML overweight, where visibility runs into 2028 and customer funding risk is spread most widely, and keep Nvidia as a core overweight.

We are neutral on Broadcom. We would move higher once the Anthropic package is fully placed and the OpenAI financing is signed, provided the terms leave those customers with a funding cost their revenue can carry and Broadcom's own exposure, through the notes and the residual-value guarantee, is disclosed and limited.

We are neutral on AMD and on wafer-fab equipment other than ASML until the 2027 spending outlook is clearer. We are underweight memory, meaning Micron, SK Hynix and Samsung, until fourth-quarter guidance and contract pricing show margins holding.

We separate a technical rebound from a trend recovery. A rebound needs only a pause in forced selling. A trend recovery needs three conditions: estimates for compute and foundry still rising after the third-quarter reporting season, neocloud credit spreads stabilising, and fed funds futures pricing a peak policy rate no higher than 4.25%.

A sector downgrade would need evidence that the capex cycle as a whole is turning, not that one company is adjusting its timing. We would move to underweight if two or more of the four largest hyperscalers guided 2027 capex flat or lower, if TSMC or ASML cut capacity plans or reported weaker orders, if credit stress reached investment-grade AI borrowers through a downgrade or a failed bond sale, or if Nvidia guided fourth-quarter revenue below its third-quarter result. A cut by a single hyperscaler would lead us to review exposure to its main suppliers rather than the sector call. We would move to overweight if at least two hyperscalers guided 2027 capex higher with funding mainly from operating cash flow, CoreWeave's bond yields fell back below 12%, and the Fed signalled a pause.

Our judgement: the pressure on the sector starts with how customers are financed, while the chipmakers' own earnings are still rising. That favours TSMC, ASML and Nvidia over memory, where we would wait for pricing evidence before buying on low multiples.

This article is for professional investors and is provided for information only. It does not constitute investment advice or an offer to buy or sell any security. Overweight, neutral and underweight are research recommendations relative to a sector benchmark and do not describe actual holdings. Scenario ranges and probabilities are estimates by Ocean Capital Research and are subject to change.

Originally published at Ocean Capital Insights.