CoreWeave's bonds yield 14.8%; ASML's customers ask about 2030

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

CoreWeave's 2032 notes yielded 14.8% in a session that took 16% off its shares, and the Financial Times, as summarised by Startup Fortune, reports that four large lenders are weighing risk transfers on their AI data-centre loans. At the other end of the stack, ASML says some customers are expressing interest in orders for 2029 and even 2030. Further down: AI shows up on the cost line at a bank, a grocer and a chatbot company; builders consolidate around closed models while Berkshire backs a software buyer; and a court, a parliament and a group of scientists argue over the terms.

The debt behind AI compute gets dearer

CoreWeave's shares fell 16% in Wednesday's US session, CNBC reports, and the yield to maturity on its 9.625% notes due 2032 stood at 14.8%, with the 2031 notes at about 14.9%. Those are two separate facts, and the bond figure is the more informative: equity can swing on sentiment, while on our reading a near-15% yield on a company with a large contracted backlog says lenders will no longer fund it on anything like the terms of earlier this year. Chief executive Mike Intrator told staff, in a memo CNBC obtained, that the market sometimes "misunderstands and acts irrationally" and that the backlog "more than supports all of our debt obligations"; Paul Meeks of Freedom Capital Markets called CoreWeave's balance sheet "relatively bad" against neocloud peers while saying he does not share the solvency concern. The banks that financed the build-out are looking to share the load: JPMorgan, Morgan Stanley, SMBC and MUFG have been exploring significant risk transfers on loans tied to AI data centres, according to a Financial Times report summarised by Startup Fortune, with the FT reporting that the banks are approaching internal limits on how much AI-linked exposure any one lender can hold. That is a statement about concentration, not necessarily about expected losses. Startup Fortune, citing Bloomberg, adds that a group including JPMorgan and MUFG has spent more than six months trying to distribute US$38bn of construction debt for a data-centre project leased to Oracle. A Gartner model, reported by The Register, frames the longer arithmetic: Gartner estimates AI operators would need US$8.5trn of revenue to 2030 to cover capital spending and financing costs at a 12% return on invested capital, against no more than US$3.6trn it forecasts AI data centres will generate, and it expects 80% of neocloud operators not to survive as independent companies beyond 2030. Those are Gartner's figures; the revenue side of any such model is the soft number, and missing a 12% hurdle means some capital earns less than its owners wanted, which is not the same as a loss.

Suppliers at the bottom of the stack still see demand

ASML chief executive Christophe Fouquet said "some customers are expressing an interest in placing orders with us for 2029 and even 2030", driven, he said, by their views on the size and duration of AI-related investment. ASML says supply planned for 2027 and 2028 will not fully meet customer demand and that it is investigating ways to lift 2028 capacity beyond the 30% increase it had previously forecast; it reported third-quarter net profit of €4.96bn against a €4.1bn consensus, on sales slightly below expectations, and said Samsung and TSMC committed to its High NA EUV tools during the quarter. Those are ASML's own comments, and an interest in ordering is not an order. Mark's response to the CoreWeave and Character.ai stories was "meanwhile, ASML", and the two pictures need not conflict: ASML's named customers are the world's two largest chipmakers, while the visible stress sits among leveraged companies renting out the finished compute. The build-out also shows up, in part, in China's trade figures: China's September exports rose 36.3% year on year, official data showed, beating a Reuters poll, and CNBC's own calculation from the official figures puts the value of integrated-circuit board exports up 136% on only 1% more units, which implies a sharp change in price or mix rather than volume. CNBC attributes part of the export strength to demand for China-made parts for AI data centres, alongside stabilising relations with the US. One of the largest sources of AI compute demand is changing who runs it: Sachin Katti, OpenAI's infrastructure and compute lead, is leaving the role after about a year and will advise the company, CNBC reports, with compute chief technology officer Uday Ruddarraju taking over the compute organisation at a company targeting roughly US$600bn in total compute spend by 2030. Nothing in the reporting suggests a change of plan.

AI shows up on the cost line

Bank of America brought in 1,000 outside contractors in the second quarter to help 3,000 internal staff patch software flaws exposed by advanced AI models including Anthropic's Mythos, chief executive Brian Moynihan told CNBC. He said the bank spent a "couple hundred million plus" in the quarter on software remediation together with anti-money-laundering compliance, and expects remediation costs to fall: "Once you patch, you don't have to patch as much." CNBC describes the figures as among the first from a major US company to put manpower and dollars on Mythos-driven defence; Moynihan also said the bank's roughly 130 internal AI projects should lift savings from US$300m this year to US$700m in 2027. A supermarket chain cites AI on the other side of the ledger: Kroger will lay off 3,000 corporate employees over three years as it expands its use of AI, the Associated Press reports. That is about 0.7% of a 409,000-strong workforce, concentrated in head-office roles, and the AP report does not separate AI-driven savings from ordinary restructuring under a chief executive who took over in February. At the consumer end, a company has set out why it failed: Character.ai's bankruptcy filing says enormous legal costs contributed to its failure, but that the business was wildly unsustainable even without them, according to 404 Media — it spent heavily to run its product, brought in little revenue, and cost-cutting changes did not close the gap. 404 also reports that Disney has offered Character.ai US$20m to license some of its intellectual property. Those details are 404's; we have not read the filings. The narrower reading concerns one business model, a cheap consumer chat product in which every message carries real inference cost, rather than the economics of AI as a whole.

Builders go closed; Berkshire backs a software buyer

Theory Ventures' survey of 455 technical builders finds teams leaning mostly or only on open-source models fell from 66% to 41% in a year, while closed-leaning teams rose from 25% to 55%. Use of the Model Context Protocol to connect chat clients to tools and data rose from 33% to 57%, Google's rival A2A fell from 17% to 8%, teams that have ever fine-tuned fell from 81% to 51%, and security rose from sixth to first among pain points. It is a survey of a venture firm's network of builders, not a random sample of enterprises, and some questions changed format between years, which Theory flags. Capital is making its own bet on where software value goes: Berkshire Hathaway has put another US$330m into Bending Spoons, the Italy-based serial acquirer of software businesses, taking its stake to 4%, according to a filing reported by CNBC. It is primary capital, 10 million newly issued shares at US$33 each, and Bending Spoons says the money will strengthen its ability to pursue acquisitions. Our interpretation, which neither company has offered, is that if AI makes seat-based software businesses cheaper to buy, an operator that can run them with far smaller teams gains a widening pool of targets; on that reading the bet is on the buyer of software rather than on software itself.

A court, a parliament and scientists argue the terms

Anthropic has asked the federal appeals court in Washington, DC, to rehear its challenge to the Pentagon's designation of the company as a national-security supply-chain risk, CNBC reports, after a divided panel upheld the label 2-1 in September. The petition argues the company was branded a risk for declining to let its models be used for mass surveillance of Americans or lethal autonomous warfare; in a parallel case, a San Francisco federal judge ruled in August that the designation amounted to illegal First Amendment retaliation. The Pentagon declined to comment on pending litigation. In Canberra, Prime Minister Anthony Albanese and Opposition Leader Angus Taylor are moving closer to a copyright deal that would clear the way for the largest AI firms to build training capacity in Australia, The Sydney Morning Herald reports, with tech firms offering to pay into a broader compensation pool rather than sign individual licences. There is no final agreement, and the Greens, independent senator David Pocock and the recording industry oppose a deal they say would shortchange creators. Anthropic's Jeff Bleich told the AFR's energy summit the company is coming regardless, "but we can only do inference at this stage, given legal uncertainty." Scientists, meanwhile, want containment made the default: an open letter whose signatories include Yoshua Bengio and Geoffrey Hinton urges organisations that develop, deploy or test cyber-capable AI models to default to airgapping them, and proposes an international body to standardise containment and record failures. The letter is advocacy, not regulation; separately, security experts quoted by CSO Online argue that the run of incidents since July reflects labs skipping decades-old security basics.

Read all 14 signals →