AI’s power problem could keep the bubble at bay
Can an artificial intelligence financial bubble be kept in check? In this episode of All Day Digital, CoBank’s Jeff Johnston shares his insights into the AI infrastructure market. The electricity shortages and permitting delays constraining data center construction may be what puts an upper bound on how much capital can ultimately be deployed by the hyperscalers.
Jeff Johnston: Hello, everybody, and welcome to the All Day Digital podcast, where we talk to industry executives and thought leaders to get their perspective on a wide range of factors shaping the digital infrastructure market. This podcast is brought to you by CoBank’s Knowledge Exchange group. I am your host, Jeff Johnston.
Instead of me bringing on a guest today, we’re going to do something a little bit different. What I want to do is provide some perspective around what’s going on in the AI infrastructure market.
Obviously, there’s a lot of reports out there regarding potential slowdowns in the market, given the results of the recent midterm election and talk of overspending and this bubble narrative being tossed around. What I wanted to do is try and distill what’s going on into three key messages
The first key message is I think we are beginning to see evidence of widespread enterprise AI adoption. This is really important to keep this whole AI party going.
The second key message is that the lack of electrons, or the lack of electricity, and permitting delays at the community level, they’re going to put an upper bound on how much capital can ultimately be deployed by the hyperscalers. This is important because I think this upper bound is probably going to prevent us from overbuilding capacity and resulting in this proverbial AI bubble that a lot of people like to talk about. That’s the second key message.
Then the last message, which is tied to the second one, is that I think that AI tokens or AI applications are going to get more expensive over time because of this supply-demand imbalance in the compute market. This is going to benefit the hyperscalers primarily. Think about Amazon, Microsoft, Meta, Google and SpaceX because they own a lot of the compute. Those are the three key messages that I want to get across. Of course, there’s a lot of detail behind all that. Let’s spend some time talking about that.
All right, let’s get into the whole AI adoption at the enterprise level. Why I think this is really important is because, in order to justify the amount of capital that’s being spent, over $750 billion this year and probably well over $1 trillion next year, we need more than just consumers using AI to get recipes and to create images and ask for relationship advice and things like that.
What we really need is companies using AI to improve productivity or to generate top-line growth. Integrating these applications into their operation is really, really important in order for the spend to continue and this whole AI party to continue.
It’s been difficult to get perspective into that up until recently. It’s largely because of Anthropic and their plans to go public. Anthropic is a U.S.-based AI model developer, application developer, but the important thing is Anthropic is targeting the enterprise market. They’re developing technologies, AI technologies, and tools and applications that they’re selling into the enterprise segment.
Now that they’ve going to go public and filed their S-1, which, of course, gives all sorts of information around their financial performance and risks and their business model so investors can assess whether or not they want to buy their stock when it goes public, it gives you a great insight into what’s happening, I would argue, at the enterprise level. Just to be quite frank, what was included in their S-1 is nothing short of spectacular, truly spectacular.
There’s a lot of nuggets in there, but I want to pull out a couple that I think are important to this conversation.
The first one has to do with how they’ve been able to expand their customer base. so check this out. At the end of 2023, Anthropic had about 20 customers spending at least $100,000 per year on Anthropic applications or technologies.
A year ago- so actually, call it September or October of 2025, they had 950 customers paying them at least $100,000 per year. Then more recently, in the last month, so call it 12 months later- they now have over 11,900 customers paying them at least $100,000 per year, and they have over 1,900 customers paying them over $1 million a year. Again, because these are largely enterprise customers, this feeds into the narrative that we’re seeing widespread enterprise adoption of AI, which, again, is critically important.
All right, so if that wasn’t impressive enough, here’s a couple of more nuggets for you. They generated over $33 billion in revenue during the first nine months of 2026, and that’s up 1,200% from the same period last year. Insane, okay? This company is only six years old, and they did $33 billion in revenue in nine months. No company in the history of the world has ever grown at that speed. Remarkable.
One last thing I’ll share with you. It’s expected that they will go public with about a $1.5 trillion valuation. [. If you add up the market cap of Coke, Pepsi, Disney, McDonald’s, Nike, Starbucks, and Home Depot, you combine all of those market caps, you get to about $1.5 trillion, equivalent to what Anthropic is potentially going to be valued at when they go public. Insane, guys. Insane. It’s great news for the whole AI ecosystem. I think it will justify continued spend into the ecosystem. Of course, that has huge implications for the economy and for the stock market and so forth, so really good news on that front.
That was the first key message I wanted to get across. Now, let’s move on to this notion that I have that electrons, or the lack of electrons, and permitting delays are going to put an upper bound on how much capital the hyperscalers can ultimately deploy. Now, I think if you look at the commentary from some of the governor-elects, there’s certainly a lot of talk about how there’s going to be more guardrails or process around permitting data centers, be it transparency or electricity regulation, because there’s clearly a lot of pushback at the community level with respect to data centers being deployed in those communities.
I think it’s fair to assume that the cadence at which we’ve been moving to deploy data centers in this country is going to slow down. How much? Nobody knows, but I think it’s going to slow down. Now, first of all, I don’t think this is a 2027 thing, I think the capex and the plans for 2027 are not going to change as a result of the midterm elections. When we get into 2028, I think that’s when we could start to potentially see the industry hitting some headwinds around how quickly they can stand up data centers.
Again, that supply-demand imbalance on the compute side, which is prevalent today, we don’t have enough compute today, that’s probably only going to get worse as we get into 2028 and beyond, given more regulation, or at least what it sounds like, regulation and process around how these data centers get approved. That’s something to watch out for. Now, I also believe that these hyperscalers, they’re not one to sit around and just accept the future status quo. They’re going to look for ways to address this compute shortage, Where I’m going with this is I think what potentially could happen is you could see these hyperscalers start to target other countries, to where they might start deploying data centers if they can’t do what they need to do in the United States.
The reality is, the industry has realized that moving gigabits of data around the planet is actually not that hard anymore compared to what it used to be. permitting and getting access to electrons, that’s a much bigger issue.
Depending on what sort of AI applications they’re doing, for example, if they’re doing training as opposed to inference, they can do training anywhere. That’s why Elon Musk is talking about putting data centers in space because you don’t have latency requirements when you’re doing training versus when you’re doing inference. I think in Southeast Asia, that’s an area to watch out for, so Singapore and Malaysia.
You hear talk of the Middle East. I don’t know the geopolitical risks there. Maybe that’s not as desirable as it used to be. I think Latin America is another market where hyperscalers might start looking at deploying data centers there. Mexico and Chile, those areas you might see some deployments. I think there’s always going to be a plan B with these companies. They’re big, sophisticated companies with enormous access to capital.
But this all takes time, right? I still think that even if they were to do that, I still think that we’re going to have this supply-demand imbalance on the compute side, which is going to benefit the hyperscalers more than anybody else.
Just to be clear, this is not a demand issue whatsoever, right? The demand for AI, the demand for compute, is very powerful, I’ve shared with you, the Anthropic numbers, which were very encouraging. I’ll give you one more reference point to reinforce this point that any slowdown in capex is a supply issue, not a demand issue.
If you look at what Amazon Web Services reported recently, 39.3% revenue growth, that’s the sixth consecutive quarter of accelerated growth in that business. They did $46 billion in revenue in Q3 of this year through Amazon Web Services, which is where their AI business resides. They did $125.8 billion in the first nine months in revenue this year in AWS, and that’s up from roughly $93 billion in the first nine months of last year. Remarkable growth. What they’re guiding to is equally remarkable. They’re talking about AWS growing 40% year-over-year in ‘27 and 40% year-over-year in 2028.
The CEO of Amazon is talking about Amazon Web Services potentially being a trillion-dollar business in 5 to 10 years from now. Again, any slowdown in capex spending and any weakness in the supply chain is not because there’s not demand for AI. It’s really around the supply, and the ability to stand up data centers is what is going to potentially put pressure on some of these capex budgets and spending. Again, we’ll put that upper bound on how much capital can be deployed, which, optimistically, hopefully, that prevents us from overbuilding like we did back in the dot-com bust.
All right, so let’s move on to my last key point. It’s tied to the second one, as I mentioned earlier. I think because of this supply-demand imbalance that we’re continuing to see and expect to get even worse in AI compute infrastructure, I think what’s ultimately going to happen is AI is going to get more expensive.
Access to data center compute and processing tokens for the major model developers, like OpenAI and Anthropic for these companies, it’s going to get more expensive. These AI applications that enterprise customers are using could get more expensive. Will that dampen demand? I don’t know. Remains to be seen, but my main point here is, and it’s kind of ironic, but I think, again, ultimately, the beneficiaries of this are going to be the hyperscalers that have existing compute infrastructure to serve up these models.
Here’s the ironic part of all of it. If you’ve been following the markets lately and some of the nervousness around capital spend at the hyperscaler level, you’re probably hearing a lot about these guys. Google and others are actually cash flow-negative because they’re using a lot of their cash flow to build data centers. That makes people nervous because Google’s never generated negative cash flow. Then, of course, they’re issuing debt at levels they’ve never issued before. They’re doing a lot of off-balance sheet stuff.
Not that there’s anything wrong with that, but it gets people a little bit nervous around just how much leverage these folks are taking on and how much cash they’re actually burning. If you fast-forward the next couple of years and if compute infrastructure gets more expensive, and then the ability to stand up data centers gets throttled because of the issues I’ve talked about already, well, that means that these hyperscalers may not be able to deploy as much capital as they had planned, which, again, back to the narrative of people getting nervous around how much money is being spent, those concerns might subside a little bit.
Then on the cash-flow side, if compute gets more expensive, these guys might start generating cash flow again while capex is moderating. They start to look really attractive all of a sudden. That’s the ironic thing about what could potentially happen around this growing supply-demand imbalance. I really think that the hyperscalers are in a great position to benefit from that. Then the likes of Anthropic and OpenAI, again, it’s going to get more expensive for those companies to develop and sell AI applications to enterprises.
Those prices are probably going to go up, and then who knows? It might result in application developers and companies to start to look at some of these much cheaper open-source models. If these applications from OpenAI and Anthropic are too expensive, you might start to see more computing being done in private data centers, with proprietary data and much smaller models. It’ll be interesting to see how this all plays out.
So those are the three key messages. Hopefully, that made sense. I know I threw a lot at you, but, of course, there’s always risks. There’s always risks in every situation. I’d be remiss if I didn’t talk about some of the risks that I think we need to pay attention to. I mentioned earlier, just the idea of open-source models becoming potentially a bigger part of the AI application story. I think that’s potentially a headwind for the closed-model companies: OpenAI and Anthropic.
They’re huge buyers of compute. If they start to lose share to some of these open-source models that are run on local data centers or private data centers, then that could potentially create some headwinds for the closed-model U.S. companies. The question then, of course, would be, do they not need as much compute because they’re losing share to the open-source model?
That’s one, potentially, area that I think that we need to pay attention to. Obviously, just how interconnected this industry has become between hyperscalers and chipset companies and model developers. There are massive, massive $100 billion-type deals that are being struck between these companies, whether it’s equity positions or vendor financing or purchase commitments, et cetera.
The concern, of course, is that when you’ve got such a small number of companies so interconnected with each other, the risk, is that if one of these companies doesn’t fulfill their end of the bargain, what happens? Is this a contagious event that spreads to other parts of the AI ecosystem? Do investors get really nervous? Is there a knee-jerk reaction? These companies have become incredibly reliant on each other, and I think that’s definitely something that we need to be mindful of and pay attention of as a risk area.
Then the last thing I would say is there’s a lot of comparisons between the dot-com bust and what’s happening today. I hear a lot of people say, “We’ve seen this movie before. I was around in the dot-com bust, and this feels like the same thing.” I’m not necessarily convinced that that’s entirely true. If you look back at what happened in the dot-com bust, there was a lot of optimism around how much data was going to be generated because of the internet. Companies went out and built an enormous amount of fiber infrastructure to transport all of this data.
The reality was the industry got ahead of itself. They massively overbuilt fiber infrastructure. They overestimated how much data the internet was going to generate back in the early 2000s. I’ve seen some reports saying there were as much as 70% or 80% of the fiber that was laid in the ground was dark, meaning it wasn’t being used. That, of course, led to this telecom bust that we saw back then because there was just too much capacity. That’s really what it boiled down to. Valuations were also a lot higher back then as well.
Well, if you look at where we are today, we don’t have any excess capacity, as we’ve talked about extensively here. There is no data center sitting around not doing anything. There’s not GPU sitting in a warehouse because nobody wants them. It’s the exact opposite. I think that’s important to keep in mind. Then from a valuation perspective, I don’t think valuations are nearly as stretched. Well, they’re not even stretched; like Nvidia is not even that expensive relative to, if you look at Cisco back in the early 2000s, I think at one point, they were trading at 60 times next year’s earnings, which was pretty frothy.
We’re just not seeing those stretched valuations in the market amongst the key AI ecosystem providers. You could make an argument that maybe Anthropic, it looks a little frothy, depending on what they come out with, but nobody’s ever grown anything like this ever before. Kudos to their management team. It’s not even close. No company has ever grown this fast at this scale ever. Maybe they deserve a high valuation, but we’ll see.
Anyway, those are the key points and some of the risks that I think we need to be aware of. Hopefully, that was helpful. Before I leave, I want to acknowledge my fellow CoBank associates, Christina Pope and Tyler Herron, because without them, there wouldn’t be an All Day Digital podcast. Thanks for listening, and watch out for our next episode.
Disclaimer: The information provided in this podcast is not intended to be investment, tax, or legal advice and should not be relied upon by listeners for such purposes. The information contained in this podcast has been compiled from what CoBank regards as reliable sources. However, CoBank does not make any representation or warranty regarding the content, and disclaims any responsibility for the information, materials, third-party opinions, and data included in this podcast. In no event will CoBank be liable for any decision made or actions taken by any person or persons relying on the information contained in this podcast.