Capital Markets Update #4

A plethora of topics exist with sufficient gravity to anchor a New Years post.  We suppose AI is one of them, so that’s what we’ll write about.  Admittedly, our views on AI are mixed.  It’s obviously an incredibly powerful productivity tool.  However, we think society will need to learn to use it effectively as an advisor rather than an answer machine.  There’s real value in challenging yourself to do the work, find the data and determine right versus wrong.  We often find that navigating this unpredictable, time-consuming maze of mental obstacles on the way to a thesis ultimately produces a more sophisticated end result, as opposed to accepting what immediately presents itself at face value.  Additionally, while the sum-total of the internet’s popular viewpoints may solve to an answer, there’s no guarantee its actually correct.  That said, we’re long AI-based productivity growth, we wish it the best.  Hopefully we all don’t end up in the matrix.  We do, however, question the foreseeable economics around AI for no other reason than an utter lack of data and an entire market resting on the outcome.  There are plenty of potential limiting agents to scaled AI, but the ultimate revenue model could be deterministic.  We see profitability as being a real sink-or-swim reckoning for the industry. 

Before we delve into AI profitability, it makes sense to play devil’s advocate. First, if you consider America’s AI dominance to be a national security issue, you can shoehorn the business into some amalgamation of a defense and infrastructure play. At that point, there’s nothing to really stop the capital infusion when you start throwing around loan guarantees and direct US government intervention.  Second, there are plenty of examples where the market leaned into a big story, choked down some volatility and re-upped.  Investors can point to the Teslas or Palantirs of the world which have, so far, successfully capitalized on a long-term growth outlook.  The old saying that the market can stay irrational longer than you can stay solvent became an old saying for a reason.    

However, its logical that the present market-wide optimism will shudder under a chill unless there’s a real value proposition on offer – and by that, we mean tangible productivity growth which turns into AI company EBITDA.  Nearly every company in the S&P 500 benefits, to some degree, from an AI-based lift to its PE multiple.  With that much at stake, here’s the math, presented a few ways.   

According to Bain, global AI companies will need to generate approximately $2.0T in annual revenue by 2030 to fund the computing power to meet expected use cases.  Bain thinks the most a global market of “AI customers” can likely afford to pay is approximately $1.2T in annual revenue, obviously representing a critical shortfall.  Ironically, we’ve been making a similar argument for a while, but come at it a little differently.  Note – these are all estimates meant to basically box a thesis; however, though not exact, we’ve tried to be as accurate as possible given limited data.  2026 AI spending in the US is expected to be approximately $525B (Goldman Sachs).  Assuming that’s about 80% levered through the whole system (banker quote), that’s approximately $100B/year in total equity being committed by domestic companies to fund the AI complex. For reference, different phases of the AI buildout are levered at different advance rates, we’ve assumed 80% across the supply chain. Let’s assume companies want to profitably underwrite to a 2-year payback on CapEx, which is accretive compared to a Google / Amazon / Oracle ~30x enterprise PE multiple and justified given they’re taking on so much leverage.  Generally speaking, CIO’s are reticent to pursue a CapEx project if it underwrites to a lower multiple than the present traded enterprise multiple.  Otherwise, you just buy your own stock, like Apple.  So, assuming 50% margins (various online estimates), companies may need to earn approximately $100B/year in revenue to manufacture $50B/year in EBITDA.  But then companies must pay interest on debt and reserve cash to fund replacement of their AI chips with an estimated 3-year useful life.  For reference, Meta, Amazon and Alphabet previously reported 3 – 4 year useful lives for their equipment for years and have only just started extending depreciation schedules out to 5 – 6 years as CapEx swelled, in order to save their P&L.  Nvidia thinks it’s a ~3-year CapEx cycle.  Assuming companies maintain 80% leverage, that’s another $50B in annual cash needs or another $100B in revenue.  Added up, you need to earn about $200B in annual revenue in order to make $500B/year recurring CapEx spend work.   If you up that aggregate CapEx commitment to $1.0T/yr, you’re now talking about $400B in annual total revenue.   To put that into perspective, OpenAI allegedly has a ~$25B recurring revenue run rate.  Meta, Alphabet and Microsoft do not report distinct AI revenue line-items – likely because they’re tough to calculate but generally thought to be insignificant.  Then you need to sit there and say to yourself where are these dollars coming from?  In theory, the Amazon / Microsoft / Meta / Alphabet / Nvidia contingent won’t be sources of revenue for themselves so you can likely exclude 5 of the top 30 revenue producing companies from the pool of customers.  If you were to ask yourself how much would a bunch of global individuals pay for what seems to be a widely adopted monthly service, you could look at streaming.  Collectively, Disney and Netflix direct-to-consumer services account for about $70B in annual revenue.  Now you’re probably getting into the substitution realm or a theoretical on how many services will the customer pay for in total.  Regardless, sticking with $500B annual CapEx requiring $200B in annual revenue, we can guestimate that you need about $130B of that coming from corporate consumers.  By comparison, three of the largest software as a service companies in the world (Salesforce, Adobe and Servicenow) collectively account for about $75B in annual revenue.   Assuming you add these together, you get about $70B coming from individuals and perhaps another $75B coming from corporations, you’re still about $55B short.    

In summary, if any of these assumptions are directionally correct on a stabilized basis, these models really better change the world because, at some point, someone needs to pay for all of this.