Market discussion

In last quarter's note we set out six reasons why I don't think the planned AI capex can earn a decent return, and therefore why I doubt it will all actually get spent over the next 5-10 years.

It is fairly straightforward to predict that this outcome would have a very detrimental impact on global markets, given AI stocks have been ‘the only game in town’ for several years now and have driven index returns – and earnings – extremely strongly over recent history.

But when the reckoning does come, how will it impact the European markets, which have a very different construction to global or US markets, no hyperscalers, and more limited (but still material) ‘direct’ AI exposure? What about a European portfolio that owns no ‘AI stocks’: is that enough? As I see it, there are four clear channels through which European markets could be impacted. This has important investment implications regarding areas to avoid and, contrariwise, areas to invest behind.

Before discussing those four channels, it is worth recapping the scale of the issue. Hyperscaler AI capex has gone from $154bn in 2023 to an estimated $794bn this year, and consensus has $1.1tn pencilled in for 2027 (Figure 1). Further growth is anticipated for several years to come. For many of these companies, this capex is no longer covered by operating cash flow.

The forecast earnings charts are where the aggregate market impact of all this investment really shows up. In the US, consensus now has S&P 500 earnings of about $420 for 2027, 52% above where 2025 came in, and the index has more or less tracked that estimate upwards (Figure 2). Year-over-year growth for earnings this year is +30%. Let that sink in. 0-10% was a good number post GFC. Nominal global GDP grows 4-6%, even in this more inflationary environment, with real growth typically 2-3%.  30% in a year is self-evidently not sustainable.

A lot of that growth is simply the other side of AI capex (one company's spending is another's revenue), which is fine for as long as the spending keeps rising. Don’t dismiss the fact that a lot of this spending is financed by the very beneficiaries who turn it into these earnings, via capital raises (often via bonds and sometimes via equity in private markets so it doesn’t show up in earnings per share). In turn these companies spend themselves, as well as expand their profit margins, so this capex impact is multiplied into aggregate profit many times over. That’s before we start discussing some of the off balance-sheet contingent liabilities (e.g. revenue guarantees) that go along with this. The debt and equity being raised pours fuel on the fire. Europe is a more moderate version, but STOXX 600 earnings are still expected to grow 28% over the same two years (Figure 3), still far better than Europe’s 0-5% post GFC trend.

In summary, any slowdown of AI spending would clearly have a huge aggregate market earnings influence. So, we need to consider the secondary and tertiary impacts as well as the obvious ones.

The Four Horsemen of the Apocalypse

When this tide goes out then, what are the areas to avoid? I think this is a very relevant discussion for clients not least as the AI trade has blurred value and growth styles (Micron is a value stock that has 1000x (!) its cash flow over last 3 years, as are banks and telcos that have benefitted from the volume growth – and obviously many large funds are growth style and tech), so clients are almost certainly not as diversified as they think they are.

So, how will the four channels referred to impact European markets specifically ?

First, the direct supply chain. In Europe for example ASML, BE Semiconductor and ASM International make the tools that make AI chips, and their order books move more or less one for one with hyperscaler and foundry capex. They are good businesses. The problem is that the share prices assume the capex cycle carries on uninterrupted. Even a pause in spending, never mind a cut, would hit orders first and the multiples second.

Second, the tangential beneficiaries. This is a bigger group than people tend to think, and it has rather crept up on the market. To pick on a few European mid-caps: Aalberts has its semiconductor systems business, Huber+Suhner sells the fibre connectivity, Munters does data-centre cooling (it booked a record SEK 2.1bn order recently), Prysmian signed a cable supply deal with Molex in July worth up to €5.5bn, and Siemens took nearly €6bn of data-centre orders in the first nine months of its fiscal year, more than double the year before. Grid investment, power, electrification and other related areas have all seen a demand boost.

So unsurprisingly, one way to win the performance game over the last year has been  to try and identify the next bottleneck and find the company exposed to it. Fine if you get it right, and fine while the music keeps playing, but a disaster when the supply demand dynamics rebalance and investment slows or is cut. Meanwhile, many of these stocks are owned as 'quality industrials' rather than as AI plays. I'm not sure all their shareholders appreciate how much of the incremental growth, and the multiple, now rests on one set of customers. Cement and steel also benefit at the edges from data-centre and grid construction, though in Europe I'd say that link is real but small, and the bigger risk to those sectors comes through the wider economy.

Third, the financial plumbing. In our Q1 note we said the 'big one' would come when the software sell-off spread to the 'mag-7 + friends', and that it could go a lot deeper into the financial system than listed private equity, through the banks and the insurers who have been handed the private credit and private equity exposures as well as benefitting from the fee income of this boom. The ECB's May Financial Stability Review puts some numbers on that: Euro-area insurers hold about €211bn of private credit, for example.

In the event software rallied and hyperscalers continued to invest. But Q1 was really a dry run.

The hyperscalers themselves are borrowing hand over fist. Around 45% of their capex over the last year has been funded with new debt. For some companies the investment is consuming over 100% of the cash flow of these behemoths, despite the contingent liability guarantees and circular financing shenanigans designed to reduce the cash flow burden. Interestingly the hyperscaler five-year CDS has blown out to around 100bp, roughly double the banks' CDS, having sat below them for most of the last five years (Figure 4). As is so often the case, the credit market has started to price the risk first, even if the equity market hasn't yet.

2006-7 is a decent analogue. It is hard to pinpoint the exact mechanism by which the bubble bursting percolates into the financial system but the imbalances and leverage are obvious, and it is almost certain this will impact financial stocks in a negative way.

The fourth and final Horseman is the wider economy. For Europe a key question is whether the boom has crowded other activity in or out, and whether it has dragged investment offshore. My view is mostly out: it has soaked up capital, power, grid capacity and credit. It is the same point we made last quarter about AI sucking capital away from the rest of the stock market, just applied to the real economy. When the spending stops those resources get freed up, but the more immediate effect is just to initially lower economic growth given the scale of what we have seen, compressed into such a short period (figure 6 below). Europe does have its own engines in German fiscal spending and defense, which have little to do with AI and should help cushion things, but they won't offset a US slowdown, a negative wealth effect and tighter credit if those all arrive together.

In conclusion these four channels are the key areas to be wary of in Europe when the market dynamics change: Direct AI capex stocks, tangential industrial beneficiaries which have seen new markets and avenues of growth open up, financial stocks exposed to the leverage (which are hard to identify in advance) and wider macro sensitivity.

Portfolio implications

BofA puts the AI Big 10 at 42% of US market cap, which is about where the Nifty Fifty, Japan in 1989 and TMT in 2000 all peaked (Figure 5). Europe is again lower, but still significant. Obviously, the key point is that, at 42% of the US index, any downturn could become reflexive and harm the broad global economy even more violently than the raw economic impact would suggest. So one obvious portfolio implication is to mind one’s eye on overall portfolio cyclicality. Only the railroads were significantly larger, and they kicked off the “Long Depression” which lasted more than five years. Japan’s asset bubble deflation famously caused a “lost decade” (which actually turned out more like a lost three decades).

The railroads also have something to say about timing, which is the uncomfortable bit for anyone running money. US railroad stocks went up threefold in the 1860s, roughly halved, and then went up another 2.5x into 1881 (Figure 7). There were about six years between the last top of the first bubble and the peak of the second. Six years is a very long time in fund management, easily long enough for a sceptic to be proved right but lose all their clients on the way. 

Incidentally, we've just lived through a shorter version of this. The FAANG stocks outperformed for many years before peaking in 2021 on the back of the pandemic, then fell hard in 2022 when inflation came back and rates went up. Within two years, largely the same companies, with Nvidia added and renamed the Magnificent 7 (and now, it seems, the AI Big 10), were taking the market to new highs.

On the flip side many fantastic track records (and careers!) were made through avoiding the storm: being zero weighted in TMT in 2001 or Japan in 1989. We can only hope our positioning, based on our analysis, leads us to this outcome and not the former one!

A final thought regarding the railroad bubble. In the end it was finished off by freight rates deflation which then led to a credit crisis for those caught up in it. They fell by roughly 5% a year for two decades as capacity came on (Figure 8). The track got built, the economy got cheap transport, and the people who had paid for it never got their return. The obvious modern equivalent is the price of a token, which is falling fast (on a like-for-like same model basis. And that is in a temporary era of compute shortage). History tells us two things. Being early can look a lot like being wrong for a long time. And in the end, the gravity of economic reality wins out.

So where to hide? One obvious answer is the classically defensive areas of healthcare, defense and probably staples, and in an AI unwind they should do the job. However there is probably more nuance that might normally be possible in the end of a cycle – a large number of stocks have not seen the supernormality that the broad AI cohort have.  In addition, all three of these defensive areas are vulnerable to certain configurations of the downturn, such as bond-vigilantism driven government spending cuts.

So we are also on the look-out for idiosyncratic businesses well away from the centre of the storm, whose earnings run on their own cycle rather than on AI capex, private credit or US economic growth.

We debate the merits of the ‘AI losers’. No doubt some of these will escape unscathed. Some will even become beneficiaries. However knowing this prospectively is unclear, and while we see the capital flow and investment into the bubble unwinding, we don’t doubt the efficacy of the technology to disrupt many business models. In our view this isn’t a contradictory statement.

Conclusions

We have discussed positioning in detail in several previous notes so don’t intend to rehash again here.

Suffice it to say that what will become desirable when these large compartments of the market which have driven all the index performance and now account for a very large portion of the overall index weight start to struggle, is the opposite of what has been desirable up to now – namely low correlation to said index. This is a self-evident statement: obviously in falling markets low correlation and a low beta is extremely valuable.  This characteristic – low beta and low correlation – is seemingly lacking in client portfolios, given the large funds in the sector do not tend to exhibit it. With this in mind it is gratifying to note our correlation to the index has been falling since the Iran war started and now sits near the lower level of funds in the sector at 53% when measured daily since inception (rolling 30 days). Other funds with similarly low levels of correlation tend to have very skewed positioning in deep value – banks et al, or extreme growth, which seems vulnerable when the tide goes out as discussed. Ours comes from a broad exposure by sector, country and type of stock, but an exposure that avoids the areas we perceive as vulnerable to the bubble bursting as discussed above. We cannot be confident of the timing. But we are confident that our positioning leaves the fund as a reliable exposure, if, and when, the markets dynamics change.