The Weekend Brief: AI Capex Meets Its Credit Wall

The week every AI conversation was waiting for finally arrived, and it split cleanly in half: the demand numbers kept getting stronger, while the money needed to build on that demand started getting dramatically more expensive. Taiwan Semiconductor printed record profit up 68.7% year over year and lifted its 2026 outlook on AI demand. ASML said AI equipment orders now stretch to 2030. Anthropic hit 16.8 billion in third-quarter revenue with its first profit, a number that quietly set up a public listing. In the same window, CoreWeave's co-founders sold 3.3 billion in shares, its bonds traded toward 15% yields, and the entire global chip complex sold off on the fear that the debt behind the buildout had become too expensive to roll. When a hot September print for producer prices and a 10-year Treasury near 5.1% walked in behind that, the October Fed meeting became a live risk again. The question the market is asking is no longer whether AI demand is real. It is who funds the next leg now that the capital stack is repricing.

The cleanest way to read the week is to stop treating the AI narrative as one story and see the two stories that are actually running. The first is demand, and it is as strong as it has been at any point in the cycle. The second is financing, and it had a genuinely bad week. Both are true at once, which is exactly why the market behaved the way it did: records in the chip complex on one hand, a quiet re-rating of the pure-play debt stories on the other.

Start with the demand side, because the numbers are the load-bearing wall. TSMC, the company that effectively manufactures the entire frontier-AI pipeline, reported third-quarter earnings on 15 October with profit up 68.7% year over year and revenue to match, and management raised the full-year 2026 outlook, attributing the upgrade to AI demand (our TSMC piece). One day earlier, ASML guided that AI demand now exceeds its planned tool supply, with orders stretching through 2030, and pointed to record-equipment sales in its fourth-quarter outlook (ASML Q4 outlook). These are not projections from a start-up with a pitch deck. They are orders, fabrication capacity and revenue from the two most important names in the physical supply chain. When the companies that actually build the machines are this far past their own supply expectations, the demand story is real.

Then there is Anthropic, which on 16 October presented the proof-of-concept the skeptics have spent two years demanding. The company reported 16.8 billion in third-quarter revenue, up about 45% from the prior quarter, and its first operating profit, in a memo aimed at prospective investors that makes clear an IPO is the endgame (Anthropic, 16 October). A frontier lab that actually covers its enormous compute bill changes the debate. It is no longer a land grab to be measured by funding rounds and capex. It is a business with revenue, a margin and a path to a public market.

Now the other half. CoreWeave is the most transparent ledger of what has changed. Its co-founders, led by chief executive Michael Intrator, sold a combined 3.3 billion in shares, 2.4 billion of it this year, according to filings and reports this week, and the company's bonds repriced so violently that its yield touched 14.8% (CoreWeave sell-down). The equity fell a further 16% in the same window (details here). When the people who built the largest neocloud and funded its debt stack choose to be paid in cash before the credit market makes up its mind, the market notices. It is not that the technology is wrong. It is that the cost of the money that pays for it just arrived at the door.

The financing story did not stop at CoreWeave. Record credit-default-swap spreads on hyperscaler and neocloud debt, a KOSPI circuit breaker during the Asian session on 15 October, and a Gartner estimate that AI-infrastructure impairment could reach 3.5 trillion all hit in the same trading day, and global chip stocks sold off on exactly that fear (global chip selloff). Then the macro backdrop made the timing worse. September producer prices rose 1.3% month over month, the biggest monthly jump since March 2022, after September CPI came in at 3.7% headline and 2.6% core, hot enough to push October Federal Reserve hike odds back to about 54% (October Fed analysis) ahead of the 27 to 28 October meeting.

Into that, the Treasury market told its own story. The 10-year spent the week near 5.1%, and a 30-year auction still cleared at 5.618% on 14 October, a level that says the long end has stopped cooperating with anyone who borrowed cheap in 2024 or 2025. Fed Governor Christopher Waller, speaking at the Bloomberg New Economy Forum on 13 October, put the whole squeeze into one phrase: higher yields are doing some of the Fed's work. For the AI capital stack, that sentence is the entire problem. The Fed's tightening is being replaced by a market that is tightening for it, and the entities in the middle, the neoclouds and the balance sheets that financed them, are the first to feel it.

Here is the structural read the week earns. This is not 1999 in new clothes. The 1999 crash happened because demand arrived late and over-capacity arrived early, and a capital market that refused to discipline either obscured both. The AI cycle has run in the opposite order. The demand is here now, it is profitable, and it is being confirmed by the biggest names in the physical supply chain. What arrived this week is the discipline, and it is coming from the credit market, not from the demand side. That is why the S&P 500's rotation out of AI names, with technology down 4.8% while healthcare, staples and materials led, reads less like a bubble bursting than like a market deciding which leg of the trade carries the value (our rotation analysis). The machines are wanted. The question is who pays for them, at what rate, and on what schedule.

The bear case still deserves its airtime, and it is not flimsy. That Gartner impairment figure, 3.5 trillion, hangs over every hyperscaler depreciation schedule in the sector, and a number that large does not get to 3.5 trillion without some portion materializing. CoreWeave's 14.8% yield is the most honest market price available for what lenders now think a neocloud earns relative to its debt. If the 10-year holds above 5% and the Fed actually hikes later in the quarter, the discount rate on every future AI dollar rises, and the companies that borrowed at 7% to build capacity against 2030 demand are the ones that feel it first and hardest.

The bull case is just as concrete. ASML is not a start-up, and its order book stretching to 2030 is not a pitch deck; if AI tool demand genuinely exceeds planned supply, semiconductor capacity is scarce and the dominant fabricators have pricing power for years. TSMC is not a conceptual AI company, it is the toll booth, and a 68.7% profit print with a raised outlook is earnings backing the narrative, not the other way around. Anthropic's first operating profit is the single best counter to the idea that the entire buildout is a round-tripping of capital. And the rotation out of the highest-multiple names is, in a perverse way, healthy: it is the market pricing risk properly rather than pretending the whole complex has no cost of borrowing.

So the mature version of the question is not whether AI demand is real. This week answered that as emphatically as it has at any point this year. It is who funds the next leg now that the capital stack is repricing. The companies that can build with cash, or with equity at a price the credit market accepts, will keep building. The ones that needed the debt market at whatever yield it now demands face a genuinely different future than the one their 2025 financings assumed, and both TSMC's records and CoreWeave's credit spreads are pointing at the same trade from opposite ends.

AI and technology: the credit wall from both sides of it

No corner of the market illustrated the week's split better than the two halves of the AI trade judged side by side. The fabricators reported records: TSMC's 68.7% profit jump and ASML's guide that demand exceeds planned tool supply, with orders through 2030, both landed while the S&P 500 technology group fell 4.8% for the week. The market did not disagree with the numbers. It repriced the risk around them.

CoreWeave is the cleanest case study of that repricing, because it shows up in real securities instead of forecasts. Its insiders sold 3.3 billion of stock this week on top of an earlier 2.4 billion tranche in 2026, its bonds traded toward 15% yields, and the equity lost another 16%. Insider cash-outs at that scale are not a bullish technical signal. They are the people who built the company deciding to be paid before the credit market decides what the equity is worth. The Gartner impairment estimate and the record CDS spreads on hyperscaler and neocloud debt are the same signal from the buy side: the debt funding the buildout is now priced for a real risk of loss (the debt-repricing analysis).

The counterweight is that the demand data kept arriving all week. ASML's order book to 2030 and TSMC's raised outlook are backward-looking to real capacity and forward-looking to real purchase orders, not to hypotheticals. Anthropic's first profitable quarter gave the market a model lab that covers its compute bill. The week's message is less that AI is over than that the financing of AI has a new price, and the pure-play names built to borrow at the old one are the ones bleeding.

IPO and deals: a pipeline that moved through the scare

The deal calendar did not stop for the AI debt scare, and the shape of what moved is informative. Anthropic's first profitable quarter dovetailed with the reports that it is building toward an IPO, making it the most valuable private AI company openly arranging a public exit at the very moment its end-market is being questioned. A profitable frontier lab going public is the kind of event that either re-rates the sector or exposes it.

The rest of the week's deal flow leaned heavily into the AI and power theme the market was selling. IonQ agreed to buy Elisity in a 1.3 billion all-stock deal, a quantum-computing consolidation paid for in equity rather than cash (IonQ and Elisity). EverWind advanced a 10 billion AI data-centre campus in Nova Scotia, 420 MW at Point Tupper with 95 MW targeted by 2027, backed by a long-term lease with a frontier AI developer and a 2 billion wind portfolio (EverWind campus). GlobalFoundries won two fabs under the 21 billion US-Japan semiconductor initiative, a state-backed expansion that does not depend on the credit market's current mood (GlobalFoundries fabs). And UnitedHealth raised its 2026 profit outlook in a beat, a reminder that the AI trade is not the only earnings story in the tape (UnitedHealth outlook).

Read the mix the way a lender would. The deals funded with equity or by government moved forward without a wobble. The buildouts dependent on new debt priced for immediate risk. That is the capital-stack story showing up in real transactions, not just in the index.

Security and sanctions: infrastructure that flexed and retreated

The week's security news read like a sanctions whipsaw, and the money moved with each twist. President Trump issued an executive order suspending the 500% tariff on Russian goods, citing national-interest authority under the Graham Act, with carve-outs for India and China through the end of October (the waiver order). That followed weeks of rising pressure on Russian energy revenue and sits against a backdrop of the energy ceasefire between Russia and Ukraine having collapsed within days of its announcement, as strikes on refineries resumed.

In the financial-crime lane, Treasury's Operation Economic Outcast designated an Iraqi bank and shadow-financing networks as part of the pressure campaign on Iranian revenue, a designation that matters to lenders and to anyone clearing payments near that network (Treasury designation). The market transmission came through shipping: the easing of shadow-fleet freight sanctions pulled freight rates down and oil off its highs. And the Strait of Hormuz remains the single biggest risk factor in global energy, with two laden supertankers hitting mines and catching fire on 17 October while transits ran a third below baseline, after an earlier mine incident on 14 October (Hormuz mine strikes).

For investors the through-line is that policy moved both directions in one week. Sanctions relief on Russian goods and shadow-fleet freight eased supply fears on one side, while the Hormuz disruption and the collapse of two ceasefires kept war risk bid on the other. That wide range is why oil's moves this week were violent in both directions.

Jobs and sectors: a cooling payroll under a hot price level

The labor market kept cooling even as the inflation prints ran hot, and the two pulled against each other in a way that complicates the Fed's path. Real wages fell 0.5% in September, the second straight month of erosion, which is the sneaky cost of that hot CPI: nominal pay is still growing, but rising prices are eating the entire gain. Falling real wages alongside 3.7% CPI is exactly the kind of squeeze that makes a central bank cautious about hiking into a labor market that is already softening.

Layoffs continued to concentrate in telecom and plant closures. Verizon filed a WARN notice covering 282 workers at Basking Ridge as part of a batch of eight filings totaling 774 affected workers, the week's clearest signal of telecom restructuring against a labor market still printing initial jobless claims near 200,000 (the WARN roundup). Netflix also cut around 1,000 roles, its largest single round since 2022 (Netflix layoffs). The churn is real but narrow; claims near 200,000 and sliding continuing claims argue the labor market is loosening from a strong level, not falling off a cliff.

The big-bank earnings that opened third-quarter season captured the same cross-current. Goldman Sachs beat by a wide margin with 17.01 in earnings per share and equities revenue up 72%, and Citigroup beat as well. JPMorgan posted its first miss since 2023, and Bank of America missed while absorbing a 1.2 billion regulatory penalty. The banks are earning from trading and from AI-related financing demand, but the penalty and the misses are the precise loss and regulatory costs of the credit cycle this week put on watch (the earnings split).

The week ahead

The October FOMC, 27 to 28, is now the fulcrum for all of it. The inflation prints landed hot enough to re-arm hike odds near 54%, and a hike, or even a hawkish hold at the top of a tightening cycle, is the single biggest swing input for the AI capital stack and for the rotation out of the trade. Watch the Treasury auctions through the week for how much real money steps in to own duration at these yields, and the continuing bank earnings for how lenders are pricing the AI debt they have to roll. The demand data will keep coming. The question is who pays for the next leg, and at what price.

This content is for informational purposes only and does not constitute financial advice. Past performance is not indicative of future results. Consult a qualified financial advisor before making investment decisions.