Thursday, November 19, 2026

Show Me The Money

Source
Michael B. Kelley's 'Yikes': CoreWeave stock tanks on bond yield woes, CEO blames market acting 'irrationally' illustrates the AI industry's hunger for investment grade debt:
Co-founder Brannin McBee, who leads CoreWeave's financing, told the FT that investment-grade ratings on more of its chip-backed debt would let it borrow from insurers, whose capital "is unlimited. Like, that's all the money in the world that sits in the insurance tranche."
CoreWeave is one of the bitcoin miners who "pivoted to AI". As I write it trades at 36% of its 52-week high,

Toby Nangle's The AI megatrade’s mounting portfolio concentration risk problem is a fascinating discussion of the looming conflict between:
  • the plans of the AI industry (hyperscalers, neoclouds like coreWeave, and AI labs) to borrow gigantic suums in the bond market to fund their data center buildout
  • and the constraints of the institutions such as pension funds that lend gigantic sums in the boind markets,
One of these constraints is widely known. Institutions such as pension funds will only lend to borrowers whose debt is rated "investment grade". And one problem for the AI industry is that the neoclouds are not, and the AI labs will apparently not be investment grade. The hyperscalers "special purpose vehicles" (SPVs) that get debt off their balance sheets are a ploy that isn't convincing either the rating agencies or bond market investors. Oracle's "investment grade" debt is trading as junk.

But that isn't the constraint that Nangle discusses. Instead, he focuses on "concentration risk". Follow me below the fold as I discuss Nangle's post and another from Bryce Elders.

Source
Nangle tries to estimate the demand for debt just from the hyperscalers:
S&P reckon that cumulative capex across the biggest six hyperscalers alone will exceed $7tn 2025-2030, and that by 2029 their ‘reported debt’ will be close to $1.2tn, up from $330bn at the end of 2025. Meanwhile, their ‘adjusted debt’ will be close to $2tn, up from just a bit over $200bn at the end of 2025.
Source
"Adjusted debt" includes debt from their SPVs. How does an increase of $1.8T in two years, or $750B/month, compare to the size of the market?
Because bonds regularly spew cash to their holders in the form of coupons, and (hopefully) disappear into a puff of money at maturity, bond portfolio managers are in a near-constant state of shopping.

As such, it’s completely normal for the investment-grade debt market to print hundreds of billions of fresh debt each month without generating front-page headlines.

Moreover, with investment-grade credit spreads — the amount of yield that investors demand to switch from government bonds — close to historic lows, the market is arguably begging for much more corporate bond issuance.

All this means that while $2-3tn sounds like quite a lot of money, it’s not a mad amount of debt for the market to absorb over the next few years. Not even close.
But it turns out that even an investment grade compnay may not be able to get their hands on all the money in the world":
But $2tn from a handful of individual companies operating out of a single sector? That’s a different matter.
Institutions lending in the debt market typically have very detailed policies expressed as "investment management agreements" (IMAs):
Contained within an IMA will be things like the benchmark that the asset manager has to beat, maybe how much tracking error they’re allowed to take, any issuers or sectors the manager must avoid, the minimum credit rating of any bond held, and what precisely to do in the event that an existing holding is downgraded below this level.

The IMA will almost always also contain some maximum permitted share of the portfolio that can be committed to bonds from a single issuer. This number tends to vary according to the security’s credit rating and often also the sector in which its issuer is based, but is usually low.
What does low mean?
Multiple investors Alphaville spoke to told us that their large institutional mandates typically have maximum individual issuer exposure 1-2 per cent, and that portfolios managed for insurance companies — which have a particular penchant for investment-grade credit, owning somewhere between a quarter and a third of the market — can sometimes be even more frugal with their maximum issuer limits.
Source
Suppose that a hyperscaler could borrow 1-2% of the entire market, how much would it be?
The US investment-grade corporate bond market weighs in with a face value of around $10.2tn, and 1-2 per cent of this is ca $100-200bn. Today, a small number of issuers are on the hook for more $100bn. None are yet on the hook for $200bn.
...
And using S&P’s narrow “reported debt” projections (and assuming this debt comes in dollars), Amazon, Oracle, Meta and SpaceX are all on track to displace the big banks as the largest bond issuers. In fact, all of them are on track to hit or breach the 2 per cent limit that most bond funds seem to stick to.
But is "reported debt" the criterion portfolio managers will use?
Investors that Alphaville spoke to told us that slapping different issuer labels on AI infrastructure debt that is accrued off-balance sheet via leases to data centre developers doesn’t really help with the concentration problem.

This is because most sophisticated investors have risk systems that look through the myriad of complexities typically packaged into a data centre financing deal to pick out tenant risk. After all, the only cash that can pay you back as a lender to a data centre project is going to come from the tenant — Google, Amazon, Microsoft, whomever.
On an adjusted debt basis the chart of hyperscaler debt above suggests that next year Amazon, Meta, Microsoft and Oracle will have exhausted their ability to borrow from the investment grade bond market. This casts a shadow over their plans for 2028 and beyond.

S&P tracks six hyperscalers, Alphabet, Amazon, Meta, Microsoft, Oracle and SpaceX. If they each borrowed 2% of the debt market the total would be around $1.2T. The same chart suggests that on an adjusted debt basis they would have borrowed $1.2T, or 12% of the total market, by the end of next year. That's another shadow.

$1.2T is a lot of money, but it isn't enough for the hyperscalers' plans. But there is another constraint that may put even $1.2T out of reach:
On top of all of this, most investors limit their collective exposure to common risk factors to a certain portion of their overall portfolio, on top of limits to individual companies.

For example, it wouldn’t be unreasonable for insurance portfolios to invest, say, up to a maximum of 5 per cent of their assets into AI financing structures, one senior investor told us. The general accounts of US life insurance hold assets of around $6tn, and property and casualty insurers oversee a further $2.7tn.

So 5 per cent of this comes to only $435bn, less than the total increase in hyperscaler adjusted debt next year, according to S&P Global forecasts, and only a smidgen more than Goldman’s 2027 direct hyperscaler issuance forecast.
Thus Nangle is very skeptical that the investment grade bond market can supply the funds the hyperscalers want for their build-out. What are the alternatives?

Source
There is the junk bond market, where companies that lack an investment grade rating borrow at significantly higher interest rates. Nangle observes that some hyperscalers already trade like junk:
Each AI-related issuer trades wide to its rating category average, shown by the black line we’ve drawn on each grid. Some of this might be because people don’t believe the ratings. But every asset manager we spoke to emphasised that a lot of this additional spread comes the from understanding that portfolios are running out of room.
There is a further problem, in that among the companies borrowing in the junk bond market will be Anthropic and OpenAI, which will not get an investment grade rating. So the huge demand from the hyperscalers would compete in the junk bond market with the huge demand from the frontier labs, their customers. Not to mention the demand from the frontier labs' customers as the VC funding that is currently keeping them afloat runs out. So interest rates on the hyperscalers' debt would increase, making them evben more cash-flow negative.

Or the hyperscalers could issue a lot more stock, diluting their existing shareholders. Meta already announced such a sale, as Dan Milmo and Graeme Wearden reported in Meta plans $95bn stock offering to help cover AI expansion. That is 2/3 of their 2027 spending plan and 6% of their market cap. It would be hard for Meta to keep increasing this in the out-years.

Source
Neither of these look good. Toby Nangle's Alphaville colleague Bryce Elders is equally skeptical in Financing for next year’s AI build-out is looking increasingly funky. He discusses the mismatch between projections from Goldman Sachs' equity and credit analysts:
The pale blue line, chip sales, is from Goldman’s equity research team. The dark blue line, data centre additions, is from Goldman’s credit strategy team. The gap between them by the end of 2028, 33GW, is approximately equivalent to the UK’s baseline electricity demand.

And it could’ve been wider! Goldman Sachs credit analysts led by Shamshad Ali this week raised the bank’s 2027 global AI project finance forecast to $545bn, which is more than double the previous estimate.
Elders cites a number of signs that neither Goldman's credit nor equity teams consider:
Even before Jensen Huang’s exit, Nvidia’s balance-sheet-as-service was already showing signs of approaching peak bloat. Bloomberg reported last week that a $500bn AI finance platform launched in August had delivered no deals, as lenders wanted bigger backstops. Banks including JPMorgan, Citigroup and Goldman have all been talking up their underwriting discipline and how they’re limiting exposure to the junk-rated issuers.
One result of Jensen's exit is new restrictions on sales to China:
Analysts at Jefferies estimate that the new restrictions put between $44bn and $62bn of Nvidia’s 2027 chip sales at risk, equivalent to 11 per cent and 16 per cent of its 2027 data centre revenue at the consensus level. The forecast includes approximately 600,000 Nvidia chips reported to be stranded in Hong Kong without import licences, worth about $18bn, which China could now plausibly return to sender in retaliation. Such a reduction in GPU sales would imply a gigawatt or two off ex-China global data centre capacity, all else equal.
All this makes it hard to undestand the credit teams exuberance:
The bank’s credit team had previously forecast $300bn in global AI project finance supply. It now splits the estimate into $225bn for “shells” (meaning warehouses, not necessarily plugged in) and another $320bn for “compute” (meaning chips, not necessarily plugged in).
That's a total of $545B, 20% more than the $435B available from the investment grade market under a 5% cap on sector exposure.

So where is all the money to come from?

No comments: