AI is the new exciting game in town - much like the internet was back in the 1990s - and the headlines boast that total AI capex spend is expected to reach a total of around $400bn this year. But if the internet triggered a tech bubble, does that mean we are in an AI bubble right now?
Much of this money is going into data centres, with the remainder being focused on the development of AI technologies - but if $400bn is being added on the cost side of the balance sheet, doesn't that mean that at least $400bn needs to be added on the income side, and surely we aren't seeing this money appear in the market?
Now big numbers can be challenging to put into perspective, but figures these large are sufficient for it to be possible to see a clear drop off in the net cash generation at the big US technology companies.
Indeed, if you load up the accounts for these companies and look through the cash flow statements you will in general see two things :
- A strong level of growth in cash flow from operations
- A bigger increase in the level of cash outflow on capex
Now in many ways this need not be a problem. The level of spend is exceptionally high, but the increase in cash flow from operations would imply that there is an increased inflow that could, and perhaps should, offset the outflow.
But with a lot of crazy figures being quoted in the press, can it really be sustainable?
If we look at Google, we get headlines along the lines of 'Google to Spend $75 Billion on AI Infrastructure', so can this be sustained and also is this a fair representation of what is happening at Google?
Infrastructure Spend on Google
Looking through the accounts, Google has historically spent $25bn on infrastructure. This climbed up to $52.5bn in 2024, and this figure is reportedly set to reach $75bn in 2025. If they do spend that $75bn in 2025, then in my mind this means that they are spending $25-30bn on business as usual, and $45-50bn on additional 'AI infrastructure'.
This tells me that those figures of $400bn that are thrown around are a bit of an exaggeration of the real story (since it includes new and pre-existing spending levels) - but at the same time, there is undeniably an increased spend on capex beyond what has been normal. So how big a problem really is this?
Well, let's continue to dig deeper through the cash flows at Google. Here we can get a better picture of what is happening at Google.
Whilst the level of cash burn on infrastructure is on a heady rise up to $75bn in 2025 - the flow of cash into Google has been incredibly impressive in recent years - climbing up from a 'mere $72bn' in 2020 to $132bn in 2024. Tracking forwards into 2025, $150bn feels like a sensible estimate for total cash inflow this year.
This incredible growth in cash generation acts to negate the impact of that cash flow out - and so despite that incredible growth in infrastructure spend, the level of cash flow after core items (including infrastructure), seems to have stayed fairly flat at around $50bn for the past five years.
Once you factor in $10bn in stock awards, $60bn in share buybacks and $7bn in dividends - this does mean that Google is now at risk of starting to burn down it's cash pile - but at the end of 2024 Google retained $96bn in cash and marketable securities, and remains a financial powerhouse even at these higher spending levels.
If it has to maintain these levels of capex spend, I do wonder if there is a bit of overconfidence in a $3tn market capitalisation (as much as I do love what Google has been doing lately) - but that is more of a valuation consideration than an operational one.
But what about the other guys in play?
Selling Shovels for the Gold Mine?
Well in terms of other key players we probably have Microsoft, OpenAI, Amazon and Nvidia to think about, and the easiest to get out the way is probably Nvidia - as they are literally just selling shovels to the gold mine, and have the most straight-forward revenue opportunity.
Indeed the accounts of Nvidia throw up no massive surprises - in [the year ending January] 2025 the net cash generated for the benefit of shareholders after core items was about $70bn, with this resulting from a stonking rise in revenues from $27bn in 2023 to $130bn in 2025.
This is a HUGE rise, and as we move through the chain of AI spend we'll have to address whether this spend level is sustainable, but for Nvidia operating costs are tiny at a mere $16bn, and their profits will essentially just flex in line with demand for their product.
So far in 2026 revenues have continued to soar - totaling $90bn in the first half of the year, with profits and net cash both sitting in the region of $45bn. Operating expenses are up, but not by much - and full year $200bn revenue and $100bn net profit targets would seem very achievable for the full year [ending January] 2026.
Personally, I'm not sure I would value the business at $4.5bn on these figures, as that implies quite a lot of confidence that these high revenue levels continue unabated - and indeed continue to grow unabated. But otherwise there seem to be no operational problems to overcome as a result of the AI boom.
What Applies for Google Must Apply for Amazon and Microsoft?
Continuing on through our list, for Amazon and Microsoft the story is pretty similar to that of Google - levels of infrastructure spend are rising and putting heavy pressure on cash flows - but actually the overall spend increase does not look quite so high. Indeed here the overestimation of AI infrastructure spend is more significant as the increase in spend is relatively small.
So once again, it is clear that the headline figures of AI infrastructure spend are heavily overstated by not deducting for the normal level of cash burn - and so by now I think we can completely discount that $400bn headline total.
Indeed in Microsoft's case the additional spend on infrastructure really isn't particularly large (what is a few billion after all?), and the bump up in capex spend fits within their overall businesses comfortably - supported by more growing incomes.
But What About the Pointy End of the AI Market?
So far we have looked at the guys who are selling services, but what about the guys in the trenches on the front line?
In my mind there are two different types of revenue within the AI space :
- an expansion of the growth out of services provided by Azure, AWS and GCP
- a newer expansion that has built up around the provision of Language Learning Models.
The side that sits around the expansion of data services we have already looked at. And here it looks like cash spend, although large, is being well managed - and against that you have a clear level of demand and supply, which do seem likely to align over time (whether demand goes up or down).
But a more complex consideration is the use of AI tools by end consumers. At the centre of that world sits ChatGPT as the consumer friendly first port of call for your new AI user.
Although ChatGPT is not a publicly listed company, we do have data that gives us insight on how OpenAI (ChatGPT) is doing financially.
Starting out with the obvious, ChatGPT is hugely successful - with 800 million weekly active users - but at this point in time it still has a somewhat incomplete business model. This is our first contrast to Nvidia, Google, Amazon and Microsoft - for whom the explosion of AI is fitting neatly into their already established business models.
Drilling down into those 800 million users, we have 40 million paying subscribers, contributing $13bn of annual revenues - or $325 per annum and per individual user. Of that $13bn, 70% comes through subscription fees, while the other 30% comes from API fees.
The problem here is a straight-forward one - 760 million of their users are freeloaders who use their platform, but do not pay anything towards its maintenance or development.
The price paid to accommodate these freeloaders in the first half of 2025 was an $8bn loss, which isn't ideal for a company currently holding a valuation of $500bn. A value that would put them about 15-17th on the S&P 500 if they were listed, sitting alongside Netflix and Mastercard.
But the logic of this free tier is undeniable, since it gives a ready pool of potential paying subscribers, who could be a relatively easy acquisition onto a paid tier.
But the challenge remains that this pool of potential subscribers pushes up the overall cost of the service - which by my estimates probably has an annual all-in-cost to cover of $20-25bn a year. As a result, to justify the a $500bn valuation, I think the company would ideally want around $50bn in annual revenues from this user base.
So is this possible?
In evaluating whether ChatGPT could get up to $50bn revenues on 800m users, let's focus on three angles :
- Could paying users pay more?
- What revenues would be required from free-tier users?
- Could a combination of both deliver sufficient value?
Starting with point (1), we have a very simple calculation to do, namely $50bn divided by 40 million users, which implies that these paying users would need to pay fees of $1,250 per year, or $100 per month, to deliver that $50bn revenue.
Now for B2B (business to business) those sorts of figures don't look crazy, but for individuals paying to increase their personal performance the fees are probably a step too far. A pivot to focus more towards businesses paying for the service might not be a bad move for ChatGPT, but these values look too high on their current approach to how they provide their service.
Moving onto point (2), there is a similar calculation to be done, taking $50bn and subtracting the $13bn in existing revenue, then dividing this missing revenue across 760 million users. This implies that the $50bn could be reached by earning $48.68 or $4 per month from these users.
Equally, this calculation implies that the revenue could also be achieved by converting an additional 20% of these free tier users to a paying subscription - and here it would not be outrageous to suggest that they could achieve this by creating a bit of additional friction to push users into the paid tier.
For example, YouTube users can use the free tier at the price of having adverts pollute their experience, and this both provides an income from free-tier users, as well as providing a ghastly enough service to encourage you to upgrade.
Something similar for ChatGPT seems feasible as they start to focus on building out their business model, and so in this sense the valuation feels fairly reasonable. And we have seen similar businesses achieve this such as YouTube, Spotify, etc etc - even in challenging market segments.
The Problem
The challenge for ChatGPT is achieving this is a very simple one - competition.
Even if they could in isolation persuade users to increase their spending from $0 to $20 or $20 a month to $100 a month, this does change the dynamics of an industry where it is common for people to subscribe to multiple services.
If overnight the industry were to implement a more mature revenue model with a greater focus on revenue acquisition, would that imply that users would change their behavior and use fewer services?
And here the whole situation gets rather complex, because it is rather hard to imagine how this would change the landscape.
In the cloud space the situation is more mature - and here you have a mature landscape with stable established market leaders - with Amazon having a 30% market share, Microsoft 20% and Google 13%.
All these make money and it is a fabulous business to be in - but what I can't get my head around is how the landscape of AI services would function in the same way.
Extracting some sort of income from the free tiers in the form of adverts or affiliate programs feels like it will be necessary to make the industry function sustainably, but I struggle a bit to see how this will be achieved in such a competitive landscape.
So Do We Have a Problem?
At the end of the day, I find myself coming back to that $400bn figure. A spend of $400bn feels unsustainable, and hard to recover through these businesses. After all, in the above calculations we were only looking to achieve revenues of $50bn from 800 million users for ChatGPT - and this would only cover 12.5% of a $400bn AI spend.
But we have already debunked that figure of $400bn, and that reduction in the true AI infrastructure spend makes the situation look a lot more sensible (although here we have to raise doubts about valuations that assume unstoppable growth in spend in this space - and Nvidia comes to mind here).
But as much as the AI guys like to talk up how much infrastructure is going to go in and how revolutionary it all is, the guys spending the bulk of the money do seem to be doing it in a pretty controlled manner.
And whilst on the pointy end of the industry I think the guys like OpenAI have a real challenge on their hands building sustainable businesses - it doesn't feel as far out as has been suggested by the more negative opinions around the industry.
So Is There An AI Bubble Or Not?
Despite some of the more lofty valuations at companies such as Nvidia - which in my mind are far too high - I don't see this as an AI bubble.
But, the reason for this is probably not the reason you are going to be expecting, because my reasoning is based around the valuation of companies like Costco.
Whilst I disagree that Nvidia should be valued at $183/share, as I just don't see revenues in the long-term holding up enough to justify that share price, I equally don't see any logic for Costco to sit at $936/share - and in terms of relative over-valuations, both seem as far out of step as the other to me.
In fact I'm pretty sure I could spend all day here going through all 500 S&P constituents and would find many to be extremely overvalued. Some of the valuations at companies such as HubSpot truly blow my mind and give off extremely strong WeWork vibes, and I could go on with Telsa, Palantir, Apple, Walmart, etc etc.
For me there is an overconfidence bubble, not an AI bubble - it is a broader overconfidence in valuations that extends through the US stock market, across much of the global private equity market, and even into asset classes such as housing and gold.
So certainty there is work to do getting sustainable revenue models in place in the AI space, but to me the valuations seem only as crazy as the rest of the US equities market. Some of the valuations are crazy, but not out of step with the other examples of crazy in the same equity space.
Comments
Post a Comment