Nvidia was back in the news early this week.
(I know, I know: “When was it not in news?”)
According to the ubiquitous “people familiar with the matter” who spoke to Bloomberg, Jensen Huang’s working with a veritable who’s who of high finance including Goldman and Blackstone on half a trillion in financing for “the buildout of AI infrastructure.”
Prior to the official announcement, the market wasn’t sure if the funding was related to the $750 billion in new, Nvidia-arranged deals which rekindled vendor financing concerns late last month. As Bloomberg put it, of Jensen Huang’s new investment push, “it wasn’t immediately clear which projects or companies the funding would back, the nature of the funding or whether the $500 billion represented new commitments or already existing ones.”
At this point I doubt it matters. Whether the new funding arrangement’s for new projects or preexisting deals, I mean. What’s another half a trillion? Is anyone even keeping track anymore? And if so, how? As critics rightfully complain, it often feels as though the hyper-scalers and AI companies more generally are blurring the line between AI revenue and capex, “shell-gaming one another’s billions back and forth,” as one strategist put it last year.
All of that to set up a few new hyper-scaler capex and free cash flow charts from Goldman. Because there’s no such thing as too many hyper-scaler capex / FCF burn charts. (“I gotta have more cowbell!”)
The figure on the left, below, shows the breakdown of historical and projected free cash flow for each hyper-scaler with the black dots being the net gain / burn. The point, obviously, is that FCF will be net negative for the group this year and next before inflecting in 2028, when consensus expects Oracle to be the only member of the group still deeply in the red.
The figure on the right shows you the growth rate both for cash flow from operations and capex for the group, with the latter outstripping the former by a huge margin until, again, 2028, when cash flow will grow faster than capex.
The upshot is that the hyper-scalers will lean on capital markets this year and next to offset some of the cash burn. That’s well-worn territory. One of the more important questions going forward for equity investors (and for the high-grade credit market too) is how far the big spenders are willing to push the issue in terms of leveraging the balance sheet before they resort to follow-on equity offerings.
I mentioned this last week, but now I have the charts and some additional color: Goldman expects the hyper-scalers to issue some $400 billion in new debt next year, which would cover about a third of estimated capex.
The figure on the left is the key chart. It shows estimated global debt issuance from Alphabet, Amazon, Meta, Microsoft and Oracle out to 2030 (the blue bars) as well as the share of projected AI outlays that issuance would cover (the grey dots).
“Our credit strategists expect the hyper-scalers to fund 35% of capex with debt in 2027,” Goldman said, adding that in addition to the $400 billion in debt issuance, the bank expects “$300 billion in project finance issuance” next year “to fund data centers and chips.” They’d know. As mentioned above, Goldman’s part of the consortium backing $500 billion in new (or “new”) Nvidia deals.
But debt markets will — and indeed already are — want concessions to fund this spending spree the more it escalates. For now, such concessions, to the extent they’re observable in any given deal, aren’t especially punitive. If the hyper-scalers want to keep it that way, they’ll mix up the funding, which means more equity raises.
As Goldman went on to say in the same note, “equity capital will also continue to play a role as companies enact multi-year strategic plans to invest in AI while maintaining the quality of their balance sheets.”
And the underwriters are already placing orders for their new Ferraris.
I couldn’t resist. He is even wearing a leather jacket like Jensen Huang:
Has anyone asked what will happen if the hyperscalers keep having to spend large amounts of money on capex in 2029 and beyond?
Oh look … there’s a new GPU, HBM, DRAM, SOC, ASIC, etc with twice the speed, half the power at two-thirds the price but I just bought this horse and buggy …
My oh my, what are us peons supposed to do?
Trump and team want to run it hot and furious into the sun.
If I recall properly, if was CFO of any company, then I would the CEO to pull forward all CapEx forward (that was a H-man article, if my memory serves me well).
Isn’t this the run into the sun, Apollo!
Enjoy the sunshine & tax deductions…I need to talk to my CPA…
Trump and team want to run it hot and furious into the sun.
If I recall properly, if was CFO of any company, then I would the CEO to pull forward all CapEx forward (that was a H-man article, if my memory serves me well).
Isn’t this the run into the sun, Apollo!
Enjoy the sunshine & tax deductions…I need to talk to my CPA…
Are hyperscalers like the railroad companies 150 years ago? Everyone rushing to build first? Seems like it.
I think more like the people building gold-rush towns in Alaska in 1897.
Someday we’ll probably all be living in the hollowed out former data centers. Hey, at least they’ve got good air conditioning.
Already started. Just saw a piece yesterday showing homeless hanging out the windows in unfinished data centers where they were squatting. Recall the pictures of all the overbuilt unfinished housing in China starting in the previous decade.
What surprises me is there are no converts or preferred stocks being issued yet.
Eerily so.
Except the railroads got lots of land including mineral rights. The hyperscalers are building to an uncertain destination the same but they do not get any land rights.