Dear Valued Clients and Friends,
The AI story is more or less the biggest story in investment markets, in the broader economy, and in the political world, all at once. There are a lot of different angles to that “AI investment” story, but no one can turn on financial television without concluding that AI or some AI adjacent angle is the obsession of the investing moment. The broad economic ramifications are equally massive if for no other reason than the fact that we basically are getting all of our economic growth right now (roughly 2% after inflation) from AI capex, data center construction, and the broad AI story. And then the political aspect is equally obvious as politicians on the left and right eagerly jump in the fray to demonstrate their bona fides, either as a champion of resisting data centers or as a champion for innovation. As heated as that last point has gotten, I suspect the divisiveness of AI politically is in the 3rd or 4th inning of where it will eventually go.
In the noise of these three different categories we are at risk of missing a very important point that actually impacts all three. I assure you all with no ambiguity that if and when the investment thesis of AI changes, the economic impact will be felt, and the political ramifications will adjust (and yes, I think that all happens in that order). The “important point” is this thing that, every now and then, investors, the media, and the general public decides is not important – that “this time” can be discarded. There is always a reason that it doesn’t matter “this time” – and that reason is usually reduced to some version: “Dude, you just don’t understand how big this is!” (I may be over-intellectualizing it). But that “thing” is the finances of it all, and I want to suggest that glorifying apathy about AI’s finances is one of the most arrogant and dangerous things I have ever seen. The historical track record for people who have said, “business models, P&L’s, capital structure don’t matter with this” is something near 0%. The historical track record for those who have insisted that they do matter is near 100%. I will let you play the odds.
So let’s jump into the Dividend Cafe where today we are not talking about public appetite for data centers, we are not talking about Malthusian or Luddite economic fears, and we are not talking about which LLM is going to destroy civilization or which one a high school senior will use to write their paper for them. Today, we are talking about the moneys being spent, the moneys being earned, the moneys making up the difference, and the moneys that need to come out of all this in the future. It is the thing that matters. And to that end we work …
| Subscribe on |
Our Vast Coverage of the AI Story Has Been Missing Something
Whether your interest is investment, economic, or political, the AI story is unavoidable, it is important, and it is often muddy. I do my very best to bring light and not heat around this subject (and all others, now that you mention it). I recognize that there are too many things going on at once to unpack the whole AI discussion in a given week. I have covered the question of AI and jobs. We have looked at the bubble risk in AI and how to avoid/mitigate it. We have analyzed whether or not AI will prove to be inflationary or deflationary. We have jumped in the discussion of data centers (though have more to do there). We have pointed out the unanswered questions about promised productivity from AI. We have applied the lessons of history regarding bubbles and manias to AI. We have told the truth about “disruption” in AI. And lest this list gets too tediously comprehensive, I will remind readers that our #1 theme for 2026 in my annual Year Ahead paper was that “AI vulnerabilities will become much more evident to the markets.”
If you re-read what I wrote at the beginning of the year there was little discussion of the financials in the AI story, very little discussion of capital structure, and very little discussion of how all these things intersect. I focused on valuation excess, of a low regard for risk, of over-inflated sentiment, of circularity in the revenue model between hyper-scalers, AI labs, and infrastructure companies, of the unlikelihood a “all will be winners” outcome, of cultural and political pushback, of under-appreciated competition from China, and so forth and so on. I don’t mind saying here in the first week of September – I think this whole thing is going to age well. Now, I did make a half-sentence reference to questions about debt vs. equity financing, and I did mention the risk of capex assumptions coming down, but my basket of “vulnerabilities” were not significantly focused on the “financial business.” And that is issue numero uno now.
Dollars and Sense
In the last few months all of these things have happened:
- Nvidia announced a $500 billion debt facility in conjunction with a consortium of Wall Street asset managers
- Google announced a $85 billion common stock offering
- Google also announced an unprecedented “one hundred year” bond offering!
- Intel announced a $15 billion common stock offering to fund additional capex needs for AI orders
- Oracle announced a plan to raise $20 billion in new equity capital and an additional $25 billion in a senior debt offering
- Amazon announced a six-part bond deal at the beginning of this year, raising $25 billion in new debt
- Meta announced a $30 billion bond deal late last year, their first bond offering in years
- Please note that this list is purposely skipping the massive amount of deals that are specific to particular data center build-outs and projects
But can I ask a very simple question, please? All of these stocks are up huge because “they are all making more money than ever” – or so I thought. And if they are all making more money than ever, why do they need to raise more money than ever? Well the known and highly public answer is that free cash flow is actually collapsing, if not going significantly negative. Revenues are increasing and cash flow is declining – it is that simple. And the reason, for good or for bad, is that the cost of making all this “stuff” (i.e. the infrastructure to build out the existence and application of Artificial Intelligence) is very, very high. These companies obviously do not need capital for general corporate use. It is the very specific and very expensive cost of building this AI compute thing out … But what we really say more and more is that a significant amount of the money needed for AI infrastructure is actually to fund the very customers doing the building.
Tackling the “Off Balance Sheet” Issue
I will add – I am purposely not including the so-called “off balance sheet” obligations (data center lease commitments are probably the biggest contributor to such but also GPU supply deals backed by the GPU’s being supplied). There are conservative estimates of over $1.5 trillion of various off balance sheet obligations out there. So why am I not including them in my discussion? First, I think the numbers on balance sheet are telling and perhaps daunting enough. Second, they are not really the same as a current liability that sits on a balance sheet (the cash outflow for the commitment is theoretically matched to future revenues). Third, many who report the off balance sheet number do it for shock and awe purposes, knowing that the net present value of these numbers that are payable over 15-20 years is actually much lower. I do not deny that substantial off balance sheet obligations exist; and I do not deny that a shortfall in AI demand or shortfall in revenue would enhance the vulnerability I am writing about today substantially. I just believe that my case in today’s Dividend Cafe can be made without the somewhat disingenuous melodrama that many feed into it, and I insist and doing so.
The Business Model in Question
From a pure business model standpoint, no one disputes that LLM’s are expensive to run and that their business models are, essentially, all the same as one another (they may have various product nuances but their underlying business models are all more or less very similar). What is most noteworthy, though, is how completely turned on its head the cost structure of LLM is versus what we have thought of as the cost structure of SaaS, software in general, and really technology in general, for at least 25 years. Usage costs go up with growing demand whereas the business models we are accustomed to thinking of in the software space saw usage costs barely move as demand for their product skyrocketed. This is not the case with the major AI labs. All indications right now are that heavier usage of AI makes the customer more expensive, not more profitable.
I am less critical of the current money losses than I am curious about the future ones. I get that the major AI labs currently lose a ton of money on even their largest paying customers, but I am willing (sort of) to just take for granted that they have some way to make money on them in the future. But I am also unaware as to what the plan is for increasing revenue and creating the scale we are used to in companies that become trillion dollar (and then some) enterprises.
I don’t think the most wildly optimistic AI bulls dispute that, right now, the AI infrastructure buildout – hundreds of billions of dollars going to the hyper-scaler capex needed for data centers, GPU’s, power, cloud capacity, networking, and more – is nowhere near being covered by what is a tiny fraction of that in actual generative AI revenue. We can all agree on this. The question, and I think it is a perfectly reasonable one, is what will the future revenue be that rationalizes all of it, where will it come from, and why will it come?
The Answer to that Question
I believe some AI critics will attempt to answer the above dilemma with some AI skeptical answer – it isn’t ever going to be able to do what it says it can do, the public will never allow all of this, the companies have over-promised and will under-deliver, etc. I want to be very clear that I am not, necessarily, saying any of this. I am limiting my scope here to the incremental profits of whatever AI does end up doing. I do not believe the current capabilities or even the potential promise of AI’s business use, adoption, application, and more are really the major concern. Yes, there is some additional risk about even those things, without financial consideration, under-delivering. But I am perfectly happy to bet that doesn’t happen, all the while asking the question: Are we sure those features and capabilities will come with the needed revenue growth and incremental profit growth to justify the investment?
To attempt to answer that question requires us to know several things we do not currently know. Even the one flashing across our screens most – the AI capex cost itself – appears to be nowhere near done. Paradoxically, the greatest AI bull argument continues to be, “but look how much everyone is spending on all this!” … I am not used to hearing that as a “bull case.” We clearly do not yet know the full cost of building this thing, but that lack of clarity pales in comparison to:
(a) What the actual customer cost will be (an economically sustainable one where the customer is paying full freight and the provider is not losing money). This number is a complete mystery right now.
(b) What future utilization will be (how many customers at what level of usage and what level of monetization will materialize). Many who are far more bearish than I am use very negative inputs here, and I am absolutely not suggesting they are wrong. But regardless of one’s inclination or suspicion as to where all this will go, there is no denying that it is a rather striking unknown right now.
(c) If the cost of A (above) comes in at an acceptable place, and the answer to B (above) outperforms expectations, will those two things happen in a timeline that is palatable? The assets involved to feed A and B are depreciating assets, right? Will the profitable utilization assumed in A and B (by the AI business optimists) happen before the technology involved is obsolete?
All of this leads us to the heart of the matter …
The Question Before Us Reduced to a Few Words
What we have here can basically be reduced to this old school financial question:
What will the return on invested capital be?
We do not need to question the usefulness of AI or the transformative wonder of it. We have to question whether the usefulness and transformation generates a return on the capital invested. And anyone who tells you that question doesn’t matter is not offering an investment opinion; they are asking you to join a cult.
The Pushback
Many would say that this concern is a mere present-tense observation of the mismatch between revenue and cost and that the future tense alignment of revenue to cost is going to be rosy. The pushback is that it is too early to assume costs don’t come down as users and user depth increase. Based on the outperformance of the technology so far, why not live a little and assume the monetization will outperform in the future, too?
And of course, some push back by saying that even if this does all prove to be a massive over-spend, no one ever got poor being invested in the companies that receive the over-spending! In other words, one man’s bubble is another man’s meal ticket (this is an argument to avoid those spending the money and go long those receiving it – the classic case for investing in the pick-and-shovel companies of the AI moment). This argument does, of course, presuppose an exit timing that I would suggest has not been historically discoverable (if I am being kind). That the stock prices of the pick-and-shovel companies presuppose continued spend at massive growth rates and are not a mere reflection of the massive moneys already spent should be a basic investment understanding for all carrying on this way. In other words, there is huge risk even for those limiting their exposure to the customers of the big spend.
Unpaid for Business is not Business at All
Does this seem like a relevant data point to you ? Per the Nvidia 10-Q, three customers are 21%, 17%, and 16% of revenue (54%) and three customers are 30%, 18%. and 16% (64%) of receivables (h/t Porter Stansberry). And if Nvidia’s customer base is not diversified, what about the customer base of Nvidia’s customers?
The major AI labs have purchase orders outstanding to their suppliers for what amounts to well over 10x their current revenues. This is basically the point of today’s Dividend Cafe. Your investment in the AI ecosystem right now has almost nothing to do with how AI as a technology does, how it is utilized, and even to some degree, how it is monetized. The story has become financial and by that I mean whether or not capital markets allow it all to play out. Of course, that still presupposes that it plays out well in the end, and that is obviously not a foregone conclusion. But what we have right now by any objective analysis are companies with very low revenues (that are growing) who make less money the more their revenues grow who can finance their growth with debt and equity (for now) coming to them from their customers. The receivables are the story, and if capital markets pause, sigh, or hiccup, the orders on the books will mean nothing to those holding the bag. If lending costs increase, if the clear rate to get deals funded moves, if the equity terms are re-rated in the market – the domino effect seems utterly stupefying to me.
None of this requires the order flow to slow down, are revenues to underwhelm. That is why investors to derive their optimism in this story entirely from growing order flow to the pick-and-shovel companies are not making the argument they think they are. The far larger risk than a slowing of order for compute power is that (a) The access to the money to pay for it goes away, or (b) The access to the money to pay for it changes as far as the terms involved, and/or (c) The profits to be derived from these expenses (the revenues to pick-and-shovel companies are expenses to those paying them) do not materialize.
I do not write today to say that A will happen, or that B will happen, or that C will happen. I write to say that A, and B, and C, are not impossible, and in terms of historical but also common-sense analysis, their probability is much higher than many investors are currently assuming.
Conclusion
I have avoided delving into the subject of AI’s social and political ramifications today. I am not concerned, today, with the macroeconomic reality (jobs, productivity, GDP, etc.). I am not picking winners and losers (the hyperscalers vs. the AI labs vs. the pick-and-shovel vs. WHICH AI labs vs. ______). I am merely wanting Dividend Cafe readers to understand that at the core of the AI vulnerability right now is NOT a debate about the popularity of data centers or the efficacy of AI applications. For investors, the core vulnerability is whether or not profits are coming to rationalize the largest capex boom in history, and whether or not those profits are coming in time. And by “in time,” I mean by the timetable of all that matters in the history of capital markets: The timetable set by markets.
Chart of the Week
Two things are true at once … +25% is a LOT. And, +25% is NOT what you are being told.
Quote of the Week
“A banker is a fellow who lends you his umbrella when the sun is shining, but wants it back the minute it begins to rain.”
~ Mark Twain
More to Chew on
* * *
I wish you all a wonderful three-day weekend. I am not a big fan of three-day weekends, but I am a fan of this bridge into the fall, a time of year I feel especially excited for this year. So may the heavy humidity and heat of summer 2026 be done, may football season be upon us, and may we all enjoy a weekend with friends and family.
With regards,
David L. Bahnsen
Chief Investment Officer, Managing Partner
The Bahnsen Group
thebahnsengroup.com
This week’s Dividend Cafe features research from S&P, Baird, Barclays, Goldman Sachs, and the IRN research platform of FactSet
