Semiconductors: Keep calm, it's another week in chips

The US semiconductor sell-off from 7 to 14 October was uneven. Micron (-16.0%) and AMD (-11.6%) led the losses, while Nvidia (-6.8%) held up best. Most of the damage came on 8 October, after OpenAI's revenue disclosure, and 14 October, after the US inflation print. Even after the fall, SMH remains up 57% YTD.

Figure 1: Semiconductors’ YTD performance

 

Higher yields triggered a rotation out of a crowded trade

The trigger was the bond market. September CPI rose 3.7% YoY, above the 3.6% expected, lifting the odds of a hike at the 27–28 October FOMC meeting to 32.7% from 22.1% and pushing the 10-year Treasury yield to about 5.3%, above SMH’s FY27E earnings yield of about 5.2%. With bonds now offering a better yield without equity risk, we have witnessed a swamp of profit taking activities out of equities, with semiconductors being the largest victims as investors took profits in semiconductors, one of the market’s most crowded trades and still up 57% YTD after the fall. The selling therefore followed positioning rather than balance sheets. Micron (-16.0%) and Lam Research (-12.9%), with little direct financing exposure, fell further than Nvidia (-6.8%), which carries the sector’s largest financing book.

Table 1: Four steps, one common link: the funding of AI compute

Date	Trigger
8 October 2026	OpenAI's annualised revenue reported at about USD50bn at end-September, below the USD68bn figure circulated a month earlier
10 October 2026	CXMT says it will launch server DRAM on its denser 4F² design by end-2026
12 October 2026	Broadcom CEO Hock Tan says capital "might be the biggest constraint" on AI Intel CEO Lip-Bu Tan points to "capital and the cost of capital" and investors wanting to "see the ROI"
14 October 2026	September CPI rose 3.7% YoY (3.4% prior, 3.6% expected) Core inflation rose 2.6% YoY (2.4% prior, 2.5% expected)
Source: CNBC, Seoul Economic Daily, iFAST compilations. Data as of 14 October 2026.

 

The other headlines put AI financing back in focus, starting with OpenAI, the largest private buyer of AI compute. The USD18bn gap reflects accounting rather than lost sales, as the USD68bn counted sales through partners such as Microsoft in full, while the USD50bn includes only OpenAI's share. The real concern is funding. OpenAI holds about USD1.4tn of compute deals and needs USD15–20bn of annual revenue per gigawatt added, so a USD50bn base leaves revenue further behind its commitments than investors assumed.

The Broadcom’s CEO comments confirmed the worry from the supply side. If capital is the binding constraint, the question becomes whether buyers can pay. Inflation then raised the price of that capital, lifting the odds of a hike at the 27–28 October FOMC meeting to 32.7% from 22.1% and leaving SMH's FY27E earnings yield of about 5.2% below the 10-year Treasury yield of about 5.3%.

Two competitive headlines were secondary. Nvidia (+1.0%) and Broadcom (+0.7%) rose after Microsoft said its Maia chips will become its main inference platform from 2027, and CXMT is a memory supply story rather than a financing one.

Why financing has been brought back in focus

Vendor financing itself is not new. What has changed is its cost. The same rise in yields that drew investors into bonds has made lenders more selective and borrowing far more expensive for the companies building AI capacity. This matters more than in past cycles because a growing share of AI spending is funded by vendors and debt rather than by buyers’ own cash, leaving the build-out more exposed to any cooling in lenders’ appetite.

Much of that exposure sits within the chip designers themselves. Nvidia held about USD99bn of equity stakes in private companies at end-July and up to USD108.5bn of conditional guarantees, USD105bn of them for OpenAI's Ohio campus, payable only if OpenAI walks away and the site cannot be re-let or sold. AMD has granted penny warrants to OpenAI and Meta, and Broadcom backs an Apollo-Blackstone financing vehicle with residual guarantees.

At the other end are borrowers such as Oracle, which carries about USD117bn of debt after free cash flow of almost -USD24bn in the year to May 2026. Suppliers sit in the opposite position. TSMC, Micron, with USD32bn of take-or-pay commitments, and equipment makers are all paid by their customers.

Figure 2: How financing today can create the demand of tomorrow

Source: IDC, Seeking Alpha, iFAST Compilations.

Figure 3: The AI circular financing network: equity stakes and purchase commitments

Source: IDC. Data as of June 2026.

Memory had a driver of its own. Micron fell -3.9% and -3.6% in the two sessions after CXMT's announcement, but pricing does not support the fear. DRAM contract prices are still expected to rise 10–15% QoQ in 4Q26, CXMT prices only 5–10% below the Korean leaders, and it has disclosed no yields or launch date and has no qualified HBM. We therefore expect limited impact towards memory chip shortages over the near term.

Credit is getting dearer and pickier, but has not dried up

The pressure shows first among the most leveraged borrowers. AI-related debt issuance fell to USD23bn in September from a record USD113bn in June, according to Morgan Stanley, and about USD18bn of loans tied to Oracle's Project Jupiter campus trade near 85 cents on the dollar.

Figure 4: AI-linked debt issuance halved in September

Source: Morgan Stanley, Financial Times.

Neoclouds are where the strain is most visible. CoreWeave's five-year credit default swap spread rose above 1,000 basis points this week, while its unsecured bonds yielded 15–16%, reflecting heightened investor concerns. However, its USD8.5bn Meta-backed facility secured financing at SOFR + 2.25%, compared with 4.5–5.5 percentage points on other facilities, showing that strong customer contracts still support cheaper borrowing. Nevertheless, heavy capex, negative free cash flow and rising financing costs could delay CoreWeave's path to sustainable profitability. For now, credit stress appears concentrated among leveraged infrastructure developers, with broader semiconductor demand at risk only if financing constraints begin delaying data-centre projects and chip orders.

Figure 5: CoreWeave's borrowing costs are rising again after its investment-grade deal

Source: Bernstein, Financial Times.

Banks are tightening at the margin rather than withdrawing. JPMorgan's Jamie Dimon said the bank is “tightening up a little bit on credit”, and on a USD5bn loan for neocloud Volta, banks underwrote only USD1.3bn, with the rest pitched at about 11%. Yet the loan still drew more than USD5bn of orders, JPMorgan's average loans rose 10% YoY and the Fed's July survey found business-loan standards basically unchanged. This favours suppliers paid by cash-rich customers over the borrowers. Tighter credit hits neoclouds and Oracle first, while foundry, memory and equipment earnings rest on hyperscalers that fund from cash flow.

Not yet a repeat of 2000

The AI build-out rhymes with the 1997–2000 telecom boom in scale and vendor financing but differs on what caused the bust. Telecom vendors lent to customers with little revenue because banks would not, and the cycle broke when bank credit tightened.

Table 2: Where the AI build-out differs from the telecom boom

	Telecom boom, 1997–2000	AI build-out, 2026
What investors paid for	Traffic and "eyeballs", with little revenue	Revenue and earnings: Nvidia's 2Q FY27 revenue +106% YoY to USD96.2bn
Main buyers	Loss-making start-ups such as Winstar, bankrupt in 2001	Mainly profitable hyperscalers; AI labs and neoclouds at the margin
Why vendors financed	Banks would no longer fund smaller carriers	Banks still lend, and institutions are joining via Nvidia's planned USD500bn alliance
Vendor exposure	Direct customer credit: Lucent committed USD5.7bn in 2000	Mostly modifiable supply commitments (USD279bn) and staged guarantees
Bank credit	Over 40% of US banks tightening by November 2000	Standards basically unchanged in the Fed's July survey
Interest rates	Fed raised rates six times, from 4.75% to 6.5%, between June 1999 and May 2000, driven by an overheating economy	One 25bp hike in September, in response to oil-driven headline shock. We do not expect a sustained rate hike cycle
Where it rhymes	Nortel and Lucent held 53% of North American optical transport; telecom capex more than doubled to about USD120bn by 2000	Nvidia and Broadcom supply about 75% of AI server chip units; 2026 hyperscalers capex estimates exceed USD700bn
Source: Seeking Alpha, Bloomberg Intelligence, iFAST compilations.

Even those who draw the parallel place today nearer 1997 than 1999, when the boom had more than two years to run. The sell-off itself lowers bubble risk, as Nvidia's earnings have grown faster than its share price.

Orders and guidance show AI demand is intact

While the slowdown in credit will bring near-term volatility, our conviction lies on the demand for AI which is still intact despite the ability of AI labs to finance their commitments are increasingly a concern. Anthropic's 3Q revenue rose +45% QoQ to USD16.8bn, slowing from +147% in 2Q, and despite an adjusted profit its infrastructure commitments exceed USD500bn, while OpenAI is expected to stay loss-making until around 2030. Slower growth against such commitments leaves the labs more dependent on outside capital, and a funding shortfall could delay new capacity and weigh on Nvidia and AMD orders. This does not yet change our outlook, as AI labs account for about a quarter of Nvidia's expected FY28 demand and hyperscalers with stronger cash generation provide much of the rest. What matters is whether customers are still committing money ahead of delivery and whether suppliers are cutting guidance. On both counts, fundamentals are intact.

TSMC's results on 15 October, the first major read after the sell-off, showed no orders being pulled. 3Q revenue rose +41.2% YoY to USD46.72bn, above guidance, net profit rose +68.7% YoY, ahead of consensus, and 4Q revenue is guided a further +14% QoQ higher at the midpoint. As TSMC makes both GPUs and custom accelerators, it gains whichever design wins.

Customers are also reserving capacity years ahead. ASML, whose machines are ordered two to three years before delivery, reported 3Q sales of EUR11.7bn, +55% YoY, raised the bottom of its 2026 guidance and expects 2027–28 supply to fall short of demand, with some customers seeking 2029–30 slots. Micron guided 1Q FY27 revenue to USD61.5bn, +13% QoQ.

Nvidia also expects about 70% revenue growth in FY28, against 45% consensus before the guidance and about 52% growth in US hyperscaler data-centre capex, and says supply, not demand, sets that ceiling.

Figure 6: Nvidia's growth guidance is limited by supply, not demand

A break in fundamentals would show up as order deferrals, guidance cuts, falling contract prices or a called guarantee. None has appeared.

The industry is already de-risking how AI is funded

A key concern earlier this year was hyperscalers' return on invested capital (ROIC), as weak returns would push them to cut capex. The ROIC visibility has improved as discussed in “Digital Economy (Internet) 4Q26 outlook: Sailing through the tides; Maintain 3.5 Stars”. Despite record capex, Google Cloud's operating margin rose to 35.6% in 2Q from 20.7% a year earlier, and AWS's rose by more than 500bp YoY to about 39%, helped by custom chips. Payback periods are also becoming visible, with Amazon's AI servers breaking even in just under three years against useful lives of five to six years, and Google citing a two-year payback.

The funding concern is real, but companies are already responding. First, risk is moving to stronger balance sheets. Lenders price the customer rather than the borrower, neoclouds have every incentive to secure hyperscaler contracts before they build, and hyperscalers have the backlog to stand behind them, with backlog of USD513.9bn at Google Cloud and USD496bn at AWS. CoreWeave's own backlog of over USD100bn is about two-thirds with investment-grade customers, and markets are now pricing that difference, refusing AI debt without operating cash flow behind it rather than AI demand itself.

Second, funding is broadening beyond chip vendors. OpenAI raised USD110bn this year, USD50bn of it from Amazon, and Nvidia's planned alliance with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR aims to mobilise more than USD500bn of third-party capital, with institutions providing most of the money. Vendors are also pulling back from the riskiest support. Reuters reported in August that Nvidia had paused parts of a plan to rent back unused capacity from smaller AI clouds.

Third, the cost of AI is falling and monetisation is broadening. Vera Rubin is designed to cut token costs well below Blackwell Ultra's, while ChatGPT's enterprise subscribers rose to about 9m in February from about 2m a year earlier, with advertising and AI agents to come. That is how the monetisation gap closes, and it answers the question Intel's CEO raised about when investors will see the return.

The market has priced in more risk than earnings show

The sell-off has gone further than fundamentals justify. Since our 30 September upgrade, SMH has fallen about -6% while our FY27E earnings estimate has edged higher, cutting its FY27E P/E to 19.4x from 21.0x on a 42% expected earnings growth in 2027.

The selling was also indiscriminate. Micron (-16.0%) and Lam Research (-12.9%), with little financing exposure, fell further than Nvidia (-6.8%), which carries the largest financing book. Investors cut crowded positions rather than repricing credit risk, leaving our preferred segments cheaper than their exposure warrants.

That does not mean the bottom is in, as another hot inflation print or a hawkish FOMC could push prices lower. On FY28E earnings, however, SMH's earnings yield rises to about 6.3%, so investors with a 12-month horizon are being paid to wait.

The sell-off strengthens our segment preference

Our thesis has not changed. We upgraded US semiconductors to 3.5 stars Attractive on 30 September on a longer earnings runway, sustained capex and persistent supply bottlenecks, and none of the past week's headlines has broken those pillars. What has changed is the price investors are willing to pay, which we see as a better entry point rather than a warning.

Higher rates and yields are a valuation headwind rather than a reason to change our view. We do not expect a sustained hiking cycle, the largest buyers fund AI from cash flow rather than debt and expected earnings growth of 42% in 2027 outweighs the rise in discount rates.

Confidence will return when investors can see returns, not just spending. The hyperscalers' 3Q results in late October will show whether cloud revenue and backlog keep pace with capex, and Nvidia's results in November should give fuller disclosure of its financing book. The 3 November midterm elections are another clearing event. Markets may stay cautious until the result is known, as the outcome could shape permitting, power and policy support for data-centre expansion, which matters directly to the pace of the build-out. Final agreements on Nvidia's financing alliance, orderly neocloud refinancing and no broad tightening in the Fed's November lending survey would show that capital remains available.

Read more: Semiconductor 4Q26: Longer runway, cheaper entry; Upgrade to 3.5 Stars Attractive

We remain our 3.5 stars "Attractive" star rating for Global Semiconductors. Our preferred exposure remains foundry, memory and wafer fabrication equipment, alongside selected fabless names in Nvidia and Broadcom. Unlike chip designers, which help fund some of their customers, foundry, memory and equipment makers are paid by the chipmakers and hyperscalers they sell to. They also supply both GPU and custom-chip programmes, so they benefit whichever design wins. That makes them less exposed to financing strain, although a broad spending slowdown would still reach them. Within equipment, we favour ASML and Lam Research. Memory's risk is the cycle rather than funding, and we see the next downturn as a 2028 rather than a 2026 risk.

Within fabless, we favour Nvidia and Broadcom, where stronger guidance has lifted our estimates and the sell-off has improved valuations. Nvidia stays a BUY, though we will track its financing book each quarter. AMD's elevated expectations and GPU warrants leave less upside, and Intel's recovery has yet to justify its valuation.

Valuation has become cheaper even as our estimates edge higher. We expect SMH's underlying earnings to grow 89% in 2026, 42% in 2027 and 22% in 2028. Applying an unchanged 24x to FY28E EPS, we raise our target slightly to USD850 from USD844, implying 51% upside.

Table 3: MVIS US Listed Semiconductor 25 Index valuation

	FY25	FY26E	FY27E	FY28E
P/E Ratio (x)		27.47	19.38	15.95
Earnings Per Share (EPS)	446.80	842.42	1193.97	1451.51
Expected Earnings Growth		89%	42%	22%
Fair P/E				24
Upside Potential				51%
Target Price (USD)				850
Source: Bloomberg Finance L.P., iFAST compilations. Data as of 14 October 2026.

Key risks and what would change our view

1. Funding stress could spread from neoclouds and AI labs to larger borrowers and hyperscalers budgets. Oracle's credit default swap spread rose about 10% in a week, and a downgrade from BBB could force sales by investment-grade-only funds.

2. Double ordering could overstate demand if adoption slows, given OpenAI's simultaneous capacity deals with Nvidia (10GW), AMD (6GW) and Broadcom (10GW).

3. Higher-for-longer rates, including a further Fed hike, would raise financing costs and weigh on valuations.

We would turn more cautious if hyperscalers cut capex guidance in the 3Q results season, a major AI lab failed to raise capital on reasonable terms, a chip designer had to honour a guarantee, or banks tightened standards in the Fed's November survey. Until then, we treat the pullback as an entry point rather than a turn in the cycle. Maintain 3.5 stars Attractive.

Declaration

For specific disclosure, at the time of publication of this report, IFPL (via its connected and associated entities) and the analyst who produced this report hold a NIL position in the abovementioned securities.

This research report was prepared with the assistance of artificial intelligence (AI) tools. iFAST Financial Pte Ltd does not rely exclusively on AI for content generation; the content of this report — including all investment theses, ratings, price targets and conclusions — has been independently reviewed and verified by the research analyst(s) to ensure accuracy and professional integrity.