Lenders put a price on the AI build-out as its chokepoints posted records

AI² — AI Investor Sydney · Friday 11 December, The Mint. One widely cited forecast puts cumulative AI and data-centre capital spending at about US$11 trillion by 2029. One day, six evidence-led briefings on where that capital is creating and destroying value — no panels, no pitches. Every ticket includes a year of NOOPS Pro. A$995 until 30 October → Book your seat

The week's argument is that the AI build-out met its price in the credit market rather than the chip market. CoreWeave's bonds traded at yields near 16 per cent and its credit default swaps passed the level an S&P Global analyst calls distress; four large banks were reported to be exploring transfers of AI-loan risk; Gartner put a number on the distance between spending and revenue, and the IMF flagged the risks of circular financing and a correction. In the same days TSMC reported a sixth straight record quarter and ASML said customers were asking about orders for 2029 and 2030. The suppliers say demand still exceeds supply. The lenders' question is who can afford to wait for that demand to turn into cash.

Lenders found a price

Broadcom's Hock Tan told a Singapore conference on Monday that capital "might be the biggest constraint" on AI expansion, ranking it above chips, memory and power, and Intel's Lip-Bu Tan agreed; Mark argued in the thread that memory remains the tighter constraint. By Thursday the market had answered for one borrower. CoreWeave's five-year credit default swaps last stood at 1,088 basis points, up 33 per cent on the week, with its 2031 bond yielding 15.78 per cent. S&P Global Market Intelligence's Gavan Nolan, who treats 1,000bp as the distress line, put the doubt precisely: not whether demand is real, but whether the cash arrives in time to justify the capex. Oracle's spreads widened too.

The banks are reaching a similar conclusion. The Financial Times reported, in a summary we could not check against the original, that JPMorgan, Morgan Stanley, SMBC and MUFG have explored significant risk transfers on data-centre loans as they approach internal limits on AI exposure. That is a statement about concentration rather than expected losses, but whoever buys the first-loss slices will want to be paid for it. Firmus showed what happens when the structure is wrong for the moment: it withdrew its A$43.7bn ASX float on Friday 9 October, with 46MW of its 912MW of contracted capacity running.

The arithmetic under the backlog

Jefferies estimates that about half of the roughly US$2.3trn contracted backlog at Microsoft, Alphabet, Amazon and Oracle comes from OpenAI and Anthropic, two private customers whose pricing, it argues, cheap Chinese and open-weight models threaten to undercut; it is one broker's view. Gartner calculates that operators would need US$8.5trn of revenue to 2030 to cover planned spending at a 12 per cent return, against the US$3.6trn it forecasts, and expects 80 per cent of neocloud operators not to survive independently beyond 2030. The IMF named an AI share correction among three salient risks to financial stability, without timing it. Missing a 12% return hurdle means returns below target, not necessarily losses; Gartner's separate forecast on which neoclouds survive is its sharper claim.

The revenue under those backlogs arrived as reports rather than filings. Anthropic reportedly told investors it made US$16.8bn in the third quarter, up about 46 per cent, with an 11 per cent adjusted operating margin and US$1.70 of revenue per dollar of compute. Revenue had more than doubled the quarter before, so 46 per cent is a deceleration, and statutory losses remain large. A frontier lab now reports adjusted profitability; whether the measures survive a prospectus is the test.

The chokepoints posted records

TSMC reported a sixth straight record quarter on Thursday, with net profit up 68.7 per cent and a 58.7 per cent operating margin. It is also sending staff to inspect US data-centre construction and power before adding capacity, and C.C. Wei still judges demand higher than supply: a supplier pacing itself to its customers' slowest input rather than to order books alone. ASML said some customers want to place orders for 2029 and even 2030, interest not yet booked. Mark's "meanwhile, ASML", posted beside the CoreWeave and Character.ai stories, carries the contrast: stress among leveraged renters of finished compute, forward demand at the maker of irreplaceable tools.

Memory is where scarcity becomes someone else's cost. Counterpoint estimates third-quarter smartphone shipments fell 20 per cent to the lowest third quarter since 2012, blaming memory shortages, with Samsung and Apple gaining share while Xiaomi's shipments fell more than 30 per cent. The memory makers did not trade like winners: their shares fell hard in the 13 October session, days after CXMT said it had progressed a denser DRAM architecture, though the analysts CNBC quoted saw little near-term threat to the Korean makers' HBM lead.

The frontier is priced for scarcity

The model layer split along the same line. Google began the wide rollout of Gemini 4 Argon to paying users only, with an API launch price set to double after 30 November, which looks like rationing compute rather than racing prices down. Artificial Analysis scored Meta's Muse Wonder Full at 56 on its index, fourth of 699 models, and Meta sells it only on paid plans. On Thursday Meta released the weights of Muse Spark Open, which it calls the most capable US open-weight model, a claim awaiting independent testing.

Builders, for now, are moving towards the closed end. Theory Ventures found teams leaning on closed models rose from 25 to 55 per cent between its two annual surveys of builders, on a question whose format changed, and Artificial Analysis's new AA-3D benchmark has Claude Sonnet 5.5 leading a three-way statistical tie, with the best open-weight models more than 200 Elo points behind. Below the frontier the price keeps falling: one developer's month with DeepSeek 4.1 Flash on a US$10 subscription left him unable to tell it from Opus mid-session, a single subjective account. If the bottom keeps getting cheaper while the top is rationed, lab margins depend on staying at the top.

Control moved into the infrastructure

Microsoft's premise for its Windows execution containers for agents is that "an agent cannot be its own security authority", and the week's governance news shared it. Satya Nadella called for an AI "emergency brake" and for treating frontier models as insider risks, which routes governance into an existing security budget. Anthropic gave six METR staff employee-like access with publication rights it cannot veto. A NeurIPS paper argued that models can provably hide capabilities even from evaluators with full weight access, shifting assurance from testing the finished model towards verifying how it was produced.

The costs are already on balance sheets. Bank of America brought in 1,000 contractors to help patch flaws exposed by models including Anthropic's Mythos, and JPMorgan will not let outside agents move money in customer accounts; John's view is that "nope is not a policy or strategy". The FTC reportedly sent civil investigative demands to OpenAI and Anthropic on Wednesday, and eight states sued OpenAI. Both labs are preparing to list; all of this becomes risk-factor text.

Software's seat is being renegotiated

The watchlist's application names went the other way to the hardware, with Salesforce, ServiceNow, Workday and Atlassian each up between 5 and 8 per cent. The signals do not explain that, but they show what vendors are trying. Workday wants its agents to reach customers' labour and outsourcing spend, priced on consumption rather than seats: a seat-based vendor conceding that the seat is the wrong unit. Microsoft is making the Copilot app the entry point to Office for Frontier programme customers. Berkshire put another US$330m into Bending Spoons, an acquirer of software businesses; our reading, not theirs, is a bet on the buyer of cheaper software rather than on software itself.

The labour evidence is about budgets more than capability. Monday's top Hacker News essay, by a laid-off engineer, put it as "Fable was cheaper than my compensation", not better at the job. Anthropic says merged code changes rose from about 2,500 a day in May to more than 30,000 a day by September, enough that it replaced git for its monorepo, while QGIS banned AI-assisted contributions, citing maintainer overload. Generating code is cheap; the systems and people that absorb it are not.

What we'd hold

First, CoreWeave's 1,000bp line and its next issuance, and whether the Oracle-linked US$38bn of construction debt clears and at what discount. Second, GPU rental prices. SemiAnalysis counts 323 Nvidia GPU rental providers and Ornn's index has B200 spot prices more than doubled since March; Dan Ives says what would change his view is "a demand signal, not a rate signal", falling rental prices among them. If rental prices hold while credit sells off, the market is repricing balance sheets; if they slip, it is repricing demand. Third, Anthropic's prospectus, and whether US$1.70 of revenue per compute dollar survives standard accounting. Lenders have named their price; those three will say whether demand can pay it.

All 73 of this week's signals are on the feed →