Chip demand hasn't cracked. Confidence in who pays for it has

Our answer

Not because demand for chips has cracked. The demand figures published around the slide were strong, and the chipmakers' own chief executives say they still cannot supply everything they are asked for. What cracked is confidence in the money that pays for that demand: how much the biggest AI customers actually earn, how much of the build-out now runs on borrowed money, and what that money costs while the Federal Reserve is raising rates again. Chipmakers sit at the end of that chain, and some of them now help finance it, so they are where the doubt lands. Memory makers, whose profits lean hardest on prices, are the most exposed if the doubt turns into a slowdown.

We read this as a repricing of financing risk, not the end of the cycle. But it is not noise either: the financing problem it exposed will still be there when sentiment turns.

How far, and why "again"

The PHLX Semiconductor Index closed at 13,217.82 on 6 October and at 12,236.79 on 12 October ^[1], a fall of 7.4% in four sessions. The worst day was 8 October, when it closed at 12,623.71 against 13,066.15 the day before ^[1], down 3.4%. The slide deepened in Asia on 13 October, where SK Hynix fell more than 8% and Samsung Electronics 7.9%, both to their lowest intraday levels since August ^[2].

This is the second slide since the summer. The index set a record close of 14,634.72 on 22 June and fell to 10,447.49 by 29 July ^[1], a drop of 28.6%. In July alone it lost 21%, its worst month since October 2008 ^[3]. It then recovered 26.5% to the 6 October close ^[1], and the 12 October close leaves it 16.4% below the June record ^[1].

The first slide had the same ingredients. When it began on 23 June, Reuters cited an investment manager who blamed a tougher interest-rate backdrop and the scale of capital needed for the next phase of AI, and noted that rate-sensitive technology stocks were being hurt by expectations of tighter policy under Fed Chair Kevin Warsh ^[4]. By August, Bloomberg put much of the volatility down to "greater scrutiny being placed on big-tech capital expenditure plans" ^[3]. In late July, CXMT's blockbuster debut in Shanghai renewed worries that its expansion could weigh on memory prices ^[5].

Rates, the capital bill, doubts about who pays, and Chinese memory supply are all back this month. What has changed is where the doubt sits.

The trigger: the customer's revenue

In the summer, the question was whether the hyperscalers would keep spending. In October it moved one link further along the chain, to the AI labs that sign some of the largest orders.

On 8 October CNBC confirmed that OpenAI had told investors its annualized revenue was roughly $50 billion at the end of September, below the $68 billion figure widely reported late in September. That day Nvidia fell 3%, Oracle nearly 6%, CoreWeave nearly 8%, AMD and Broadcom 4% and Intel 5% ^[6]. Reuters tied that day's 3.4% fall in chipmakers to the Financial Times report that OpenAI's annualized revenue was $20 billion less than the company had signalled ^[7].

Part of the gap is presentation rather than collapse. A person familiar with the matter told CNBC the $68 billion figure included gross revenue from OpenAI's partners ^[6]. That evening, New York time, Bloomberg reported that OpenAI expects to reach or exceed $70 billion in annualized revenue by the end of the year ^[8].

The reaction matters more than the number. One customer's run-rate moved the whole chip index because some of the largest orders now come from AI labs ^[10][11], and OpenAI, by its supplier's account, is still losing money. Broadcom's chief executive said of OpenAI: "For now it's still losing money because they're spending a lot on training" ^[9].

The real story: the build-out now runs on debt

The second link is how those orders are paid for.

On 7 October the Wall Street Journal reported three debt deals under negotiation, as relayed by Investing.com ^[10]:

• Broadcom is working to assemble a package of more than $50 billion to finance custom AI chips for OpenAI.
• Oracle is negotiating a structure in which private investors fund an off-balance-sheet entity that buys chips and leases them back to Oracle.
• SpaceX has approached lenders about a $40 billion debt structure to buy Nvidia processors.

Two days earlier, banks had begun syndicating $42 billion of senior secured loans backed by Broadcom's A-minus credit rating. The loans fund Anthropic's chip orders due for delivery in 2027, and the lease payments start only after the hardware is deployed ^[11].

Read those together and the chipmaker is no longer only the supplier. Broadcom's rating backs the $42 billion senior part of the Anthropic financing ^[11], and Broadcom is raising money for OpenAI's chips ^[10]. If a customer's revenue disappoints, in our reading, the doubt flows back to the vendor.

The credit market has noticed how much the vendors themselves are borrowing. On 8 October Broadcom's five-year credit default swaps widened to a record 136 basis points, and the cost of insuring Oracle's debt closed at a record 261 basis points. Bloomberg tied the rise to a wave of potential new financings from SpaceX and Broadcom ^[12].

The scale keeps growing:

• Bloomberg counts nearly half a trillion dollars of new debt priced this year to pay for AI infrastructure ^[12].
• Morgan Stanley data put AI-linked debt raised in 2026 so far at an estimated $466 billion, against $101 billion in the same period last year ^[13].
• Bloomberg calculates that Broadcom alone may raise about $600 billion to finance computing power in the coming years ^[12].

Issuance has slowed. Tech companies raised $23 billion of AI-related debt in September, less than half August's total, and AI-related bond issuance in the US investment-grade market stopped completely that month ^[13]. Morgan Stanley put the drop down to the large amount of debt already raised earlier in the year ^[13]. Appetite is being tested too: on 9 October Firmus, an Nvidia-backed Australian data-centre operator, scrapped what would have been the second-largest new share sale in Australia's history after lukewarm demand ^[14].

Bloomberg's own reading is that the issue "is less about actual deterioration in credit quality than about market saturation" ^[12]. We agree. Saturation is still a problem for a build-out whose financing needs are rising.

The industry's leaders said much the same in Singapore on 12 October ^[15]:

• Broadcom's Hock Tan: "[What] I think might be the biggest constraint is capital".
• Intel's Lip-Bu Tan said the front-line challenge is "capital and the cost of capital", and that in the longer run "people want to see the ROI".

Both remain bullish on near-term demand ^[15]. Hock Tan said demand for computing infrastructure exceeds supply and that the imbalance could persist through 2029 and potentially into 2030 ^[15]. Lip-Bu Tan said hardware will remain in short supply for the next two to three years ^[9].

Vital Knowledge's Adam Crisafulli said Thursday's sell-off reflected extreme positioning with further to unwind, and pointed to "growing signs that markets are pushing back on the tidal wave of AI-linked debt and equity being thrown at them" ^[16].

The backdrop: money costs more again

The third link is the price of the money. The Fed raised its target range by a quarter point, to 3.75% to 4%, on 16 September ^[17]. The minutes released on 7 October said most participants judged that "another increase in the target range for the federal funds rate would likely be appropriate by year end" ^[18]. On 8 October Governor Waller said 16 of the 18 participants anticipated at least one more hike this year ^[19].

The minutes tie AI to this directly:

• Market commentary relayed in them pointed to "competition for capital from heavy private debt issuance to finance the development of artificial intelligence (AI) infrastructure" as one of the factors behind higher term premiums and Treasury yields ^[18].
• Participants said surging AI-related investment was adding to inflation pressures ^[18].
• Waller counted evidence that the AI build-out "was significantly driving up high-tech consumer prices" among his reasons for supporting hikes ^[19].

The boom has become one of the reasons money costs it more.

On 7 October the 10-year Treasury yield traded as high as 5.365%, its highest since April 2002, and CNBC reported that tech stocks fell on worries that higher borrowing costs would limit the AI build-out ^[20]. But rates were not the trigger. On 8 October, the worst day for chips, the Treasury's own end-of-day curve shows the 10-year yield falling to 5.22% from 5.28% ^[21]. Rates make a financing doubt more expensive; they did not start it.

Why memory is the most exposed

Memory is where we think the doubt matters most, because memory profits lean hardest on prices staying high.

At an industry conference on 10 October, CXMT's president outlined plans to launch a new generation of DDR5 server memory modules by the end of 2026, built on a denser 4F² DRAM design ^[22]. When markets reopened on 12 October, Micron fell 3.9% in New York, and the Korean memory makers fell sharply the next morning ^[2]. The strategists CNBC quoted played down the near-term threat. UBP's Vey-Sern Ling said investors "are jittery and looking for excuses to take profit" on AI stocks, especially memory. Rayliant's Phillip Wool read it as "a sign of how vulnerable AI sentiment is at this moment" ^[2].

The demand evidence did not move:

• Samsung guided third-quarter operating profit to about 107.40 trillion won, against 12.17 trillion won a year earlier ^[23]. That beat the brokerage consensus by 1.9%, and the stock still fell 2.42% on the day ^[24].
• On 12 October, the day of the CXMT sell-off, Nanya said it expects DRAM prices and gross margins to keep rising in the fourth quarter because the market remains severely undersupplied ^[25].

The worry is how long it lasts, and here the evidence points at price:

• SanDisk's management noted, as Trefis reported, that higher prices drove about two-thirds of its revenue growth in fiscal Q4 2026 over the previous quarter ^[26].
• Apple has told some of its suppliers to cut production of components for the iPhone 18 Pro and Pro Max, after soaring memory costs forced price increases that dampened demand ^[27].
• Intel's Lip-Bu Tan said memory could be among the first parts of the market to lose some steam ^[9].

A cycle driven by price is the one most exposed to new supply. That helps explain why a Chinese product plan for the end of 2026 weighed on the group this month.

What would tell us which way this goes

• Inflation and the Fed. September CPI is due on 14 October at 8:30 a.m. Eastern ^[28], and the Fed next meets on 27-28 October ^[18]. A hot print makes every financing structure above more expensive.
• The chip reports. ASML reports third-quarter results on 14 October ^[29] and TSMC on 15 October ^[30]. TSMC's quiet period runs from 5 to 14 October ^[30], so its management has not spoken to investors during the slide. Its only release during the slide was the September revenue report on 8 October, which showed revenue up 54.6% from a year earlier ^[31]. Global chip sales in August were $159.7 billion, up 144.3% ^[32]. What these calls need to answer is not whether demand exists, but what customers will order for 2027 and how they will pay.
• Arm's court case. The Delaware jury in Qualcomm's licensing dispute with Arm ended Friday without a verdict and was due back on 13 October. Before trial, Qualcomm had asked the judge to rule that a clause in the agreements would let it stop paying royalties for five years if the contract were breached ^[33].
• The debt market. Watch whether AI-linked issuance reopens after September's drop ^[13], and where Broadcom's and Oracle's credit default swaps go from their records ^[12].

Our verdict

This slump is a repricing of who pays for the AI build-out, and at what cost. It is not a sign that the build-out is slowing. The chip and memory makers report demand ahead of supply; it is their customers' finances that are being questioned.

That distinction tells you what would end it. A strong quarter from TSMC answers a demand question, and the demand evidence is already strong. What would end it is evidence that the money is there: AI debt selling again without record spreads, and the largest labs' revenue growing into the orders they have signed.

Until then, the most exposed companies are of two kinds. Some both supply the chips and underwrite the customers. Others rest their profits most heavily on price, and memory comes first among those.