Disclosure: I hold shares in ASML, Nvidia and other semiconductor companies.
On Tuesday I asked whether ASML's results would show customers still writing cheques. They did. Emphatically.
TSMC followed with a sixth straight record quarter. But in between these announcements, Korea's stock market had to be halted after an 8% fall, and the cost of insuring against default on an AI company's debt hit a record. It has been quite a week for the AI trade.
What happened?
ASML reported on Wednesday: record third-quarter sales of €11.7bn, a 57.5% gross margin and up to €15.2bn guided for the fourth quarter. The line that mattered came from chief executive Christophe Fouquet: supply "as currently planned for both 2027 and 2028 will not fully meet the demand of our customers", with some customers already asking about 2029 and 2030.
TSMC reported on Thursday: net profit up 69% to about $23.6bn, revenue of $46.7bn, AI-heavy high-performance computing at 68% of sales, and fourth-quarter guidance implying growth of up to 61%. Even so, CEO C.C. Wei struck a more cautious note on the call, warning that rising component prices and economic uncertainty were weighing on demand in some parts of the market.
By any normal standard, those are extraordinary numbers from both companies.
Why the shrug?
The likely reason: the market was busy panicking about something else. US consumer and producer inflation both came in hot, driven by fuel, traders started pricing another Fed rate rise, and the Nasdaq fell 1.9% on Wednesday, its worst day since July. The selling hit Asia on Thursday: the Kospi triggered a circuit breaker after an 8% fall and the Nikkei lost almost 5%. The Philadelphia chip index fell for a seventh straight session.
The sharper signal came from credit. CoreWeave is the poster child of the "neoclouds": companies whose data centres do nothing but run AI, built on Nvidia chips and rented out to the likes of OpenAI, Microsoft, Meta and Anthropic. It has around $52bn of debt and lease obligations, and in its latest quarter made an operating loss of $49m while paying $640m in net interest. Its 2032 bonds, issued just four months ago, slipped below 80 cents on the dollar, which would be considered distressed territory, giving them a yield of around 15%.
On Thursday, the cost of insuring its debt for five years rose to almost 1,200 basis points, The Times reports, more than triple the early-October level and a clear sign of how much more risk lenders now see. XTB's Kathleen Brooks called it "the AI trade's canary in the coal mine".
It isn't alone. Default insurance on Oracle, also borrowing tens of billions for AI, hit a record too, as did spreads for Amazon, Alphabet, Meta and Nvidia, albeit from far lower levels and much more slowly.
The defence was robust. CoreWeave's chief executive Mike Intrator told staff that markets can "misunderstand and act irrationally", and Jensen Huang told Bloomberg that "CoreWeave's strength is improving every day" and that Nvidia had no need to backstop it. The shares bounced 6% on Thursday. It's worth remembering that Nvidia is both CoreWeave's chip supplier and one of its shareholders, so Huang's confidence, however genuine, isn't independent evidence that the financing is safe.
By Friday, calm had returned, helped by Fed chair Kevin Warsh praising a "massive investment surge" in AI, which cut the odds of an October rate rise. Asian markets rebounded, though the MSCI Asia Pacific index still ended the week down around 3%.
What I think it means
The AI trade is no longer one trade. It has now split into three layers.
In the bottom layer, ASML and TSMC are sold out, with pricing power and order books stretching into the next decade. Their risk is cyclical: some orders may represent a place in the queue rather than a firm commitment.
At the top, the model makers are generating real revenue. Anthropic reportedly told prospective IPO investors that its revenue for the third quarter was $16.8bn, up 45% on the second. That is a remarkable number, yet some commentators suggest it represents a slowdown, because revenue had more than doubled the quarter before, which says something about how high expectations have become.
Steve Eisman, of Big Short fame, says his bigger worry is concentration: much of the build-out ultimately rests on two customers, OpenAI and Anthropic. He thinks the era of companies telling staff to use AI regardless of cost is ending and that cheaper open-weight models could start a price war. If that happens, he argues, the whole chain will feel it.
The stress seems to be in the middle, where companies are borrowing vast sums to buy from the bottom layer and build the infrastructure for the top layer: in essence, building data centres with expensive debt and betting customer contracts will pay it back before the hardware ages. Market strategist Neil Campling told Bloomberg that CoreWeave alone is expected to burn through more than $30bn of cash, far more than it brings in as revenue. That is where higher rates bite hardest,and where credit markets tend to price in trouble before equity markets do.
Not everyone thinks it ends quietly. Bill Blain, the veteran bond market commentator, compared CoreWeave this week to Northern Rock in 2007, the moment a bubble wakes up and smells the lack of money, and argued that a default or "troubling rescue" is on the cards. His wider worry: private credit funds have lent billions secured on GPUs, and second-hand chips may be worth far less than the loans assume. Stephen Innes adds that CoreWeave is also the tenant behind roughly $14bn of high-yield data-centre bonds, so its credit problem becomes its landlords' problem too. CoreWeave and Nvidia clearly disagree, but these are the right questions to ask about debt backed by hardware that ages fast.
Eisman has the view that this isn't 2008: the banks are fine, and a failing AI borrower is more like General Motors going bust than a run on the financial system. Painful, possibly enough to tip the US into recession, but not the end of the world.
There was one more twist, though perhaps a predictable one: a flight to safety. Treasuries rallied hard on Thursday despite hot producer prices, suggesting a 5%-plus yield on US government debt was too attractive to turn down when everything else was turning red.
The counterargument
A bond market tantrum over a handful of leveraged companies isn't really a verdict on AI. As ASML and TSMC showed, the backlogs are real, and their customers have some of the deepest pockets in the world. So this week's selloff may be more about inflation and the Fed than AI. Chip stocks have run very hard this year, and by Thursday afternoon US investors were already buying the dip. Altimeter's Brad Gerstner, who has cut his own exposure to its lowest in three years, still told CNBC he was more positive on 2027 and 2028, and that this kind of churn is healthy because "it avoids bubbles". If the hyperscalers keep spending, this week will look like an overreaction.
My read
The supply chain's verdict is clear: the build-out isn't slowing yet. But the risk seems to have moved. The chipmakers have visibility on demand, whereas the companies financing the infrastructure have to turn that demand into cash before their hardware ages and their debt comes due.
Higher rates make that test harder and less forgiving. I'd watch AI credit spreads more closely than the next set of chip results. Hock Tan's casino is still open. The question is more about who is playing with too much borrowed money.
The Comms Read
Two lessons. ASML stopped publishing quarterly bookings this year, so the market now reads every word, and "will not fully meet the demand" worked because it was specific: named years, a capacity figure, orders for 2029 and 2030.
CoreWeave is the opposite case. An internal memo is an external document the moment it's written, and telling staff the market is "irrational" rarely calms investors pricing your debt at 15%. What they want is boring and specific: debt maturities, contracted cash flows, and how quickly the backlog turns into cash. When your bonds start doing the talking, the response should focus on the numbers, not adjectives.
What to watch
AI credit spreads, especially CoreWeave and Oracle; the hyperscalers' results over the next fortnight; the Fed's October meeting; and whether Anthropic launches its IPO in November.
Time will tell. GLA & DYOR.
Disclaimer
This newsletter is published by James Lea / CorpCast for general information, educational and journalistic purposes only. Nothing in it constitutes financial, investment, legal or tax advice, or a personal recommendation, and it is not an offer, invitation or inducement to buy, sell or hold any security, cryptocurrency or other financial instrument.
Any views, opinions, forecasts or interpretations expressed are the author's own personal opinions at the time of publication, do not reflect the views of any business or entity the author is affiliated with, and may change without notice. No account is taken of any reader's personal circumstances, objectives or financial situation, and nothing here should be relied upon as the basis for any investment decision.
The author may from time to time hold long or short positions in securities, cryptocurrencies or other assets referenced in this newsletter, and may transact in them at any time without further notice. This is a standing disclosure covering all issues and is not repeated or itemised per asset.
Investing and trading involve substantial risk of loss, including the potential loss of your entire investment. Past performance is not a reliable indicator of future results. Figures, data and third-party information are believed reliable but are not guaranteed to be accurate, complete or up to date.
James Lea / CorpCast is not authorised or regulated by the Financial Conduct Authority. Always conduct your own independent research and consult a suitably qualified, FCA-regulated financial adviser before making any investment decision.
(The sources and standing disclaimer are in the file.)