Global Investment Outlook 2024 — The Capex Turn

If 2023 established artificial intelligence as an investable theme, 2024 revealed what kind of investment it actually was. The distinction matters more than it first appears.

The initial framing treated AI as a software opportunity — and software, as an asset class, has a specific and highly attractive shape: high gross margins, negligible marginal cost of serving an additional customer, and modest capital requirements relative to revenue. That profile is why software dominated venture returns for two decades.

Through 2024 it became clear the AI build-out did not have that shape. Training and serving large models requires data centres, specialised semiconductors, and — increasingly the binding constraint — electrical power generation and transmission. These are long-lived physical assets with heavy upfront capital requirements and long payback periods. They are, in economic character, infrastructure.

That reclassification had consequences beyond terminology. It changed which investors could sensibly participate, since a venture fund is not structured to finance a power station. It changed the return profile, since infrastructure returns are lower and slower than venture returns but more predictable. And it created a genuine question that remained open through the year: whether revenue would arrive at the scale required to justify capital commitments being made years in advance.

Meanwhile the interest-rate cycle turned. Policy easing began in several major economies. For public markets this mattered. For private markets it mattered less than expected, because by 2024 the binding constraint was not the cost of capital but the absence of exits.

From software economics to industrial economics

The difference between a software business and an infrastructure business is not a matter of degree. They are different asset classes with different financing, different risk, and different natural owners.

A software business requires modest capital, scales at near-zero marginal cost, and can grow revenue far faster than costs. Venture capital exists because this profile produces outcomes where a small number of investments return the entire fund.

An infrastructure business requires very large capital before any revenue, produces a relatively predictable return over a long life, and cannot scale faster than it can build. It is financed with debt and long-duration equity, and owned by investors matching long-dated liabilities.

The AI build-out required both simultaneously. The model layer and application layer retained software characteristics. The compute layer beneath them did not.

The question stopped being whether AI would generate value and became who owns which layer of it — because the layers have entirely different economics.

This produced a financing pattern with few clean precedents in prior technology cycles. Physical build-out was funded substantially by the balance sheets of a small number of very large, highly profitable technology companies, supplemented by infrastructure funds and private credit. Venture capital participated mainly at the model and application layers.

For allocators the practical consequence was that AI exposure required a decision about which layer, and those decisions had almost nothing in common with one another in terms of risk, duration or return.

The power constraint

An underappreciated development of 2024 was that the binding constraint on AI capacity began to shift from semiconductors to electricity.

Chips can be manufactured relatively quickly once fabrication capacity exists. Generation and transmission cannot. Building new power generation, and particularly new transmission, involves permitting, land, grid interconnection and construction timelines measured in years — in some jurisdictions, over a decade.

This gave the theme an unusual characteristic: a physical bottleneck that capital alone cannot relieve within the investment horizon of most funds. Money does not shorten an interconnection queue.

Its investment implications ran in several directions. Existing generation assets with grid connection — particularly firm, dispatchable capacity — became more valuable, since they held something that could not be quickly reproduced. Locations with power availability gained a durable advantage in attracting data centre investment. And the timeline for AI capacity growth became partly a function of infrastructure policy rather than technology or capital.

It also introduced a risk that is easy to miss when analysing the theme purely as a technology story: if capacity growth is constrained by physical build-out, then revenue forecasts predicated on rapid capacity expansion inherit that constraint.

Rate cuts mattered less than expected

Policy easing began in 2024 after the most aggressive tightening cycle in decades. A reasonable prior would have been that this substantially relieved private markets, since the 2022 difficulties were fundamentally a rate story.

The relief was more modest, for a reason worth understanding.

By 2024, the binding constraint in private markets was not the cost of capital. It was the absence of exits. Institutions were not short of willingness to commit; they were short of realised capital to commit with. That is a liquidity problem, and it does not resolve simply because borrowing becomes cheaper.

Cheaper debt does help transaction volume at the margin — buyout maths improves, and the bid-ask between buyers and sellers narrows. But the deeper problem was the gap between carrying values set in 2021 and prices buyers would pay in 2024. That gap closes through either time, marking down, or an improvement in the underlying businesses. None of those is accelerated by a rate cut.

The general principle: when the constraint has migrated, relieving the original cause does not restore the original condition. A market can be released from the force that stopped it and remain stopped, because a different mechanism now binds.

Private credit grew faster than it was tested

Private credit continued expanding rapidly through 2024, extending the trend that began when banks retreated from leveraged lending and accelerated when rising rates made floating-rate lending attractive.

The growth was well-founded in several respects. Bank retrenchment was structural rather than cyclical, driven by capital requirements. Borrowers valued speed and certainty of execution. And floating-rate structures performed exactly as designed through the rate rise.

The open question was not whether the asset class had a reason to exist — it plainly did — but whether its central claim had been verified. That claim is that direct lenders underwrite better than syndicated markets, hold to maturity, and work out problems more effectively because they hold concentrated positions and have direct relationships with borrowers.

That claim had not been tested by a default cycle. The asset class grew to its current scale during a period of unusually low corporate defaults. Its performance record therefore measures a benign environment, not underwriting quality.

Two structural features deserve attention regardless of one's view. Valuations are largely mark-to-model rather than mark-to-market, which reintroduces the smoothing problem discussed in the 2022 report — measured volatility understates economic volatility. And the growth of payment-in-kind arrangements, where a borrower defers cash interest by adding it to principal, can postpone the visible appearance of stress without removing it.

None of this is an argument that private credit is unsound. It is an argument that its risk statistics describe a period that has not yet included the event they most need to describe.

Concentration deepened

The narrow leadership that characterised 2023 did not resolve in 2024. It intensified.

The mechanism is described in the 2023 report and unchanged: capitalisation weighting means the largest constituents dominate index movement. What changed was the degree, and the fact that a second consecutive year made it a structural feature rather than an anomaly.

For allocators this raised a question that is uncomfortable precisely because the answer is not obvious. If a small number of companies constitute a very large share of an index, then a passive index allocation is a concentrated position. But the alternative — deliberately underweighting the largest companies — is an active bet against businesses that have been, by most measures, exceptionally profitable and well-positioned.

There is no analytically clean resolution. The honest framing is that index investing in a concentrated market involves an implicit view, and the choice is whether to hold that view knowingly or unknowingly.

Diagnosing which constraint is binding

The central analytical failure of 2024 was applying a framework built for one constraint to a situation governed by another. That failure is general enough to be worth setting out as a method.

A system under stress usually has several potential constraints and one that is actually binding. Relieving a constraint that is not binding produces no effect, which is why the correct diagnosis matters more than the strength of the remedy.

In 2022 the binding constraint was the cost of capital. Rates rose, long-duration assets repriced, and the transmission was direct. A framework built around the discount rate explained nearly everything, as the 2022 report describes.

By 2024 the binding constraint had migrated to the availability of capital at the LP level. Distributions had not recovered, so commitments could not be made, so funds could not deploy. Rate cuts address the cost of capital. They do not make an institution's private portfolio return cash.

The diagnostic questions that distinguish them:

• Is the party who wants to transact unable to afford it, or unable to obtain it? The first is a price constraint; the second is an availability constraint. They respond to entirely different remedies.
• Would the situation resolve if prices fell? If yes, the constraint is price. If a lower price would not help because the counterparty is absent, it is availability.
• Where in the chain does the blockage sit? Tracing the chain — exits to distributions to commitments to deployment to companies — identifies which link is broken. Applying a remedy to the wrong link produces nothing.
• What happened when the obvious remedy was applied? Rate cuts arrived in 2024 and exits did not reopen. That is diagnostic information, and it was available by observing the non-response.

When a familiar remedy produces no effect, the useful conclusion is not that more of it is needed. It is that the constraint has moved, and the framework that identified the old one is now pointing at the wrong variable.

The same error appears elsewhere in this archive. The 2019 report describes a recession indicator that was correct about direction and useless on the timeline that mattered. The 2016 report describes markets pricing an event correctly and the trade incorrectly. In each case the framework was sound and was applied to the wrong question.

What an allocator could act on

Measure distributions, not exits. Distributions as a percentage of net asset value measures capital actually returned. Listing counts and announced transaction values measure activity. These diverge substantially — a listing with a lock-up produces no distribution in the year it occurs — and only the first relieves the constraint that determines commitment pacing. Bain publishes this figure free.

Separate the capital-intensity question from the thematic question. AI's reclassification from software to infrastructure changed the capital required per company by roughly an order of magnitude. That is a question about what a venture fund can fund, and it is separate from whether the theme is attractive. A fund whose size was set for software economics cannot participate at infrastructure scale, regardless of its view.

Look for the constraint that has migrated, not the one that is familiar. Physical constraints — power, land, grid connection, advanced manufacturing capacity — re-entered technology investing in 2024 after two decades of irrelevance. The frameworks most investors carried had no place for them, which is precisely why they were under-priced.

Hold index concentration as a knowing position. A cap-weighted index in a concentrated market embeds a view about the concentrated names. There is no analytically clean resolution to this — the choice is between holding that view knowingly and holding it unknowingly, and only the first can be sized.

Assess private credit by vintage rather than in aggregate. The 2022–2023 origination vintage carries the asset class's best terms; 2021 and earlier carry the risk. Aggregate portfolio statistics blend them and describe neither. Asking a manager for the split by origination year is a reasonable request, and reluctance to provide it is itself informative.

What 2024 established

• AI became an infrastructure question, not only a software one, with all that implies for duration, financing and ownership.
• Physical constraints re-entered technology investing after two decades in which software's freedom from them was the defining feature.
• The limits of monetary policy as a private-market cure became clear: rate cuts did not restore an exit market.
• Private credit reached systemic scale while its central underwriting claim remained unverified.
• Concentration was accepted as a persistent condition rather than a temporary distortion.