The Factories Are Sold Out. The Loan Window Is Closing.

What Needs Your Attention Today

Watching means normal, warning means look closer, danger means act. The word always carries the meaning — never the colour alone.

The 60-Second Picture

Read the weekend's two headlines side by side and the pattern is the whole issue. On the demand side, Seoul Economic Daily reported Sunday that Samsung will use arch-rival TSMC's 3-nanometre process to make base dies — the foundation layer of the stacked memory chips (called HBM) that sit next to AI processors and feed them data — for Nvidia's next-generation custom memory, the first time the two rivals have cooperated at this stage of memory production. Rivals do not share factories in a weak market. On the financing side, Bloomberg reported that lenders are pulling back from exactly the kind of borrower that needs those chips most. Demand is not the constraint anymore. Credit is.

The Macro Vector: A Boom Meets a Narrowing Loan Window

Think of the build-out as two different booms sharing one name. The first belongs to the giants — Microsoft, Alphabet, Amazon and Meta — who guide to roughly $725 billion of combined spending this year (about $220 billion Amazon, $195–205 billion Alphabet, about $175 billion Microsoft, $130–145 billion Meta) and can still borrow cheaply because lenders trust their existing businesses. The second belongs to the specialist renters — the "neoclouds" that buy chips, rent out computing power, and fund almost all of it with borrowed money. That second boom is where the window is closing.

The evidence arrived in layers this week. Global AI-related debt issuance fell from $113 billion in June to $23 billion in September, according to figures reported by the Financial Times — an 80% drop in the flow of new loans while announced spending kept climbing. Bloomberg put total AI-related debt fundraising this year at nearly $500 billion, and reported that the average price of high-yield data-center bonds has slid from $101 in June to about $95 — bonds are priced as a share of their face value, so lenders who bought in June are already sitting on a loss — with the broader junk market's average yield reaching 8.41% on Wednesday, an 18-month high. One live deal shows the new terms: the $5 billion leveraged loan for Volta Infrastructure was restructured, ending at a yield above the 11% bankers first discussed. And the stress is spreading from prices to projects — roughly $18 billion of bank loans tied to an Oracle data-center campus have entered challenging territory, and at least 47 major US data-center projects have been delayed, withdrawn or blocked this year.

The mechanism is simple, and it is not "AI is a bubble." It is arithmetic. A borrower paying 5% can survive a slow quarter; a borrower paying 15% cannot miss at all. Goldman Sachs data cited by Bloomberg show companies have already issued $106 billion of project-finance bonds this year — nearly triple the $37 billion issued in all of 2025 — with perhaps $300 billion more coming globally next year. That wall of new supply is itself the problem: each new deal competes with the last for the same lenders, so prices fall and yields rise even when nothing about demand changes. The International Monetary Fund warned on October 13 that exactly this combination — heavy borrowing, high valuations, and returns that must arrive on schedule — could become a source of instability if the revenue disappoints.

What is priced in: consensus already assumes the giants spend more than $1.1 trillion in 2027, and that the loan window reopens for everyone else. What would prove this wrong: a neocloud refinancing at single-digit yields, or AI borrower spreads tightening back toward September levels while issuance recovers. Neither happened this week.

Winners and Losers Split: The Balance-Sheet Trade

Friday's chip sell-off looked indiscriminate until you sort it by who needs lenders. Nvidia — cash-rich, and the seller everyone else borrows to pay — fell just 1.43% to $214.13, beating the Nasdaq. The falls steepened as balance sheets weakened: Broadcom −4.52%, Micron −5.63%, AMD −5.95%, and TSMC −2.28% despite posting the strongest numbers of the season. Credit told the same story with a week's head start: between October 8 and October 15, the median borrowing premium across eight AI infrastructure borrowers widened from 334 to about 413 basis points, while the median ordinary junk bond moved 1 basis point. CoreWeave's median spread reached 934 basis points. When one corner of the credit market reprices and the rest shrugs, the market is not scared of the economy — it is scared of a business model.

CoreWeave remains the stress gauge. Its second quarter showed the paradox in full: revenue of $2.575 billion, up 112%, a contract backlog of about $104 billion — and a $626 million net loss with $640 million of interest expense in a single quarter, against roughly $50 billion of total debt and lease obligations. Much of that stack is older, secured or lease-based borrowing at lower rates; the 15% is the price of the newest unsecured money — which is precisely the next dollar CoreWeave needs to keep building. Its bonds, issued in June at a 9.675% yield, traded near a 15% yield last week, and Bloomberg reported it is now exploring the investment-grade market — borrowing from insurers under stricter rules — because the junk market has become too expensive. Management insists the backlog more than covers the debt, and Nvidia's Jensen Huang publicly backed the company this week. Both can be true and the bonds can still be right: the backlog is real, but it pays out over years, while the interest is due every quarter.

An analogy: this is a property boom where the builders with cash keep building, and the builders with mortgages discover the bank has quietly changed its mind. It breaks in one place — buildings do not halve in value every two years, and AI chips do, as each new generation lands. That is why lenders charge the renters 15%: the collateral ages faster than the loan.

One quieter lens, from the research shelf: a new McKinsey and Semiconductor Industry Association study maps the hidden upstream dependencies behind US chip production — suppliers in the US, Europe and Japan account for about 95% of the value of equipment and materials, but the physical risks sit several layers deeper, in critical minerals and sub-components. And McKinsey's explainer on AI costs notes that moving a tool from pilot to production can nearly quadruple what it costs to run — one bank's projected AI bill jumped from $3 million to more than $25 million once every employee had every tool. The build-out's bill, in other words, keeps growing after the ribbon is cut. See McKinsey on semiconductor supply chains and “What does AI really cost?”

What to Watch This Week

• Big Tech earnings — Microsoft and Alphabet confirmed for October 28, Meta and Amazon expected the same week. One question matters: the 2027 spending guide. Reaffirmed or raised guides, paired with strong cloud growth, are the cleanest way this sell-off ends. A trim from any one of the four hits every supplier and every lender at once.
• CoreWeave's next financing move, not its stock. If it places investment-grade or secured debt at tolerable cost, the stress story softens. If the only bids are at 15%, expect consolidation: research firm Gartner forecasts that more than 80% of specialist AI cloud providers will cease operating independently by 2030. The listed names carrying the heaviest borrowing loads — and therefore the most refinancing risk — are CoreWeave (CRWV), Nebius (NBIS), Core Scientific (CORZ) and Applied Digital (APLD); the cash-rich end of the trade is TSMC (TSM) and Nvidia (NVDA), which collect from all of them.
• Issuance, weekly. September's $23 billion is the number to beat. A recovery toward June's pace says lenders are back; another month near $23 billion says the second boom is being rationed.
• The 10-year yield at 5.10%. Every AI loan is priced off it. Below 5% loosens everything on this page; a push back toward recent highs tightens it all again.
• If you work at — or sell to — a builder: two numbers now protect your budget. The first is cost per unit of computing: teams that can show each unit of AI getting cheaper to run keep their funding when credit tightens. The second is your own department's AI bill: McKinsey's researchers found that moving an AI tool from pilot to full rollout can nearly quadruple its cost, and one bank's projected spend jumped from $3 million to more than $25 million once everyone had every tool with no guardrails. Audit your team's tools, switch off the duplicates, and bring both numbers to your next review — before finance asks for them.