Global Investment Outlook 2026 — The Return Question

This is an outlook rather than a retrospective, and its statements should be read as scenarios and questions to monitor rather than as forecasts.

The dominant fact entering 2026 is the scale of capital committed to artificial intelligence infrastructure over the preceding three years. That commitment was made on a thesis about future demand. Three years is long enough that the thesis begins to be testable.

The question is not whether AI produces value — that is largely settled. It is whether the value produced accrues to the parties who financed the build-out, at a rate justifying capital committed years in advance against assets with long lives and heavy depreciation.

Several things follow. If revenue growth validates the capex, the investment case extends and capital continues flowing. If it disappoints, the correction would concentrate in exactly the assets that have driven index returns — which, given how concentrated those indices have become, would transmit to the broad market regardless of what the rest of the economy is doing.

Beneath that headline question sit three slower-moving issues with their own timelines: a thinning early-stage pipeline whose cost is only now becoming visible, a private credit market that has grown to systemic importance without a stress test, and an exit environment that has improved without normalising.

The depreciation schedule

This is the most underappreciated mechanism heading into 2026, and it is worth setting out precisely because it is an accounting matter with real consequences.

When a company builds a data centre, it does not expense the cost immediately. The asset is capitalised and depreciated across its useful life — commonly several years for computing equipment, longer for buildings. Cash leaves immediately; the earnings charge arrives gradually.

This produces a specific and predictable pattern:

• In the build phase, reported earnings are flattered. Enormous cash outflow occurs while only a fraction hits the income statement.
• As assets mature, depreciation accumulates. Each year's spending adds a new layer of charges on top of prior years'.
• The charges arrive regardless of revenue. Depreciation is a function of the asset and its assumed life, not of whether it is productively used.

So a company spending heavily on infrastructure looks more profitable during the build than the economics justify, and less profitable afterwards — with the crossover arriving on a schedule set years earlier.

The capex decision was made on a demand forecast. The depreciation charge arrives on a calendar. Only one of those is contingent.

Two variables deserve close attention. Useful-life assumptions determine how quickly charges accumulate; extending assumed life reduces annual depreciation and raises reported earnings without any change in economics, and such changes are disclosed in filings. And utilisation determines whether the assets generate revenue — depreciation on underused capacity is a pure drag.

This does not predict a bad outcome. It identifies where evidence will appear first, and in a form that is publicly disclosed rather than inferred.

The early-stage gap arrives

The cohort gap described in the 2025 report begins producing visible effects in 2026.

The mechanism is simple and slow. Seed formation outside favoured themes was constrained from roughly 2022 onward. Companies funded at seed in a given year typically reach Series A two to three years later, Series B four to five. A shortfall at seed therefore appears as thin Series A supply after a two-to-three year delay.

The consequences run in two directions at once, which is what makes the situation awkward rather than simply bad:

• For growth-stage investors: capital is available but the population of companies that have reached the relevant milestones is smaller than usual. Too much capital chasing too few qualifying assets raises prices for the companies that do qualify — which looks like a strong market while actually reflecting a supply shortage.
• For the companies that were funded: conditions are unusually favourable, since they face less competition for attention and capital than a normal cohort would.

The distortion is that strong pricing at Series A in 2026 should not be read as evidence of a healthy pipeline. It may be evidence of the opposite. Distinguishing the two requires looking at deal count alongside deal value — a thin market with rising prices looks very different from a broad one.

Private credit meets its test

Private credit enters 2026 at systemic scale, having grown through a period with no meaningful default cycle. Its central claim — that direct lenders underwrite better, hold through problems, and work out difficulties more effectively than syndicated markets — remains untested by the event it most needs to be tested by.

If corporate defaults normalise toward long-run averages, several things become observable that currently are not:

• Whether mark-to-model valuations converge with realisable value. Direct lending marks are largely model-based. A default cycle forces the comparison.
• Whether payment-in-kind arrangements have deferred visible stress. PIK lets a borrower add interest to principal rather than paying cash. Widespread use can postpone the appearance of distress without changing the underlying condition. BDC filings disclose PIK income and are public.
• How concentrated positions behave in workout. The claim is that concentration improves outcomes because the lender has both incentive and authority to act. The counter-argument is that concentration removes the option to exit. Both are plausible; only a cycle distinguishes them.

The honest position is that this is genuinely unresolved. Private credit's growth rests on a real structural change — bank retrenchment from leveraged lending is regulatory and permanent — but its risk statistics describe a benign period. The asset class is sound in rationale and unproven in stress.

Exits determine everything else

Of all the variables in 2026, the one with the widest downstream effect is the least discussed: whether distributions to limited partners normalise.

The chain runs mechanically. Private funds return capital through sales and listings. Institutions use those distributions to fund new commitments. When distributions fall, new commitments fall — not from any change of view about the asset class, but because the money has not come back.

Four years of constrained exits have produced a backlog of unrealised assets held well past their underwritten horizons. That backlog matters in ways beyond the obvious:

• It suppresses fundraising, including for managers whose performance is strong.
• It depresses reported IRRs even where multiples are intact, because return is time-sensitive.
• It sustains the gap between carrying value and realisable value, since positions marked but never sold cannot be validated.
• It keeps secondaries structural rather than opportunistic.

A genuine normalisation of exits would relieve all four at once. That is why exit conditions are the single highest-leverage variable to monitor in 2026 — more consequential than rates, and more consequential than any thematic allocation decision.

What would change the picture

Rather than forecast, it is more useful to name what would constitute evidence — the observations that should update a view.

Evidence the AI capital cycle is working: revenue growth at the application layer accelerating rather than plateauing; utilisation rates on deployed compute staying high; capex guidance rising alongside disclosed returns rather than ahead of them.

Evidence it is not: useful-life assumptions being extended in filings; capex growth decelerating while depreciation continues climbing; large customers renegotiating rather than expanding commitments.

Evidence the early-stage gap is resolving: seed deal count rising, not just seed deal value. Value alone can rise on concentration.

Evidence private credit is holding up: default rates rising while realised recoveries stay near modelled assumptions; PIK income share stable rather than climbing.

Evidence exits are normalising: distributions as a percentage of net asset value returning toward long-run averages — a more meaningful indicator than IPO counts, because it measures capital actually returned.

What the archive suggests about this moment

An archive covering eleven prior years is worth interrogating for patterns, and three recur with enough consistency to be worth stating as priors rather than as predictions.

Deferrals compound. The 2019 report describes a correction interrupted by policy support and arriving three years later with a larger accumulated gap. The 2023 US venture report describes bridge rounds deferring adjustments that cost more when they expired. In both cases the deferral was individually rational and collectively expensive, because the interval consumed the resources that would have made the adjustment survivable.

The 2026 application: the AI capital cycle's test is being deferred by continued capital availability. That is not evidence the test will be failed. It is a reason to expect that when the test arrives, it arrives against a larger position than it would have three years earlier.

Averages stop describing anything under concentration. The 2025 reports establish this at global and US level; the 2025 Asia-Pacific report shows the same distribution arriving with a different dividing line. In every case the market-level statistic became a weighted average of two conditions and described neither.

The 2026 application: any statement about "the market" this year should be treated as a composite until the segments are checked. Deal count alongside deal value, segment-level rather than aggregate data, medians rather than means.

A pattern is a claim about its precondition. The 2019 report makes this point about policy asymmetry — a decade of evidence that rested on an unstated condition, which held until it did not. The 2016 report makes it about the zero lower bound. The 2017 report makes it about measured volatility.

The 2026 application: the strongest current patterns — that capital remains available for AI, that private credit performs well, that concentration continues — each rest on conditions that are nameable. Naming them is the analytical work. A pattern whose precondition you cannot state is a pattern you cannot know is still operating.

Positioning for an unresolved question

An outlook that identifies an open question owes the reader something about how to hold a portfolio while it resolves, without pretending to know the answer.

Size to the branch you would not want to be wrong about. The AI return question has two broad outcomes with materially different consequences. A position sized to the expected value is over-sized relative to the adverse branch. This is the same discipline the 2016 report argues for around binary political outcomes, and it applies more strongly here because the adverse branch concentrates in exactly the assets that have driven index returns.

Distinguish exposure to the constraint from exposure to the thesis. The 2024 and 2025 Asia-Pacific reports develop this: a supplier whose revenue depends on the activity occurring is differently exposed from a company whose value depends on a particular layer capturing the economics. Both are AI exposure. They resolve differently, and a portfolio holding both is holding two things rather than one concentrated bet.

Watch the disclosed evidence rather than the funding announcements. Capex guidance, depreciation, useful-life assumptions and utilisation are published quarterly and are free. Funding rounds tell you what investors believe. Filings tell you what customers paid. The second updates a view; the first mostly confirms one.

Treat exit normalisation as the highest-leverage variable and measure it correctly. Distributions as a percentage of net asset value measures capital returned. IPO counts measure transactions announced. These diverge substantially — a listing with a six-month lock-up produces no distribution in the year it occurs — and only the first relieves the constraint that determines commitment pacing.

Accept lower deployment when prices are high. This is the archive's most consistently recurring finding and its least comfortable one. The 2021 US venture report establishes that entry price is the dominant variable and is fixed at the decision point. The 2026 India report describes the same discipline. The years that feel best to deploy in are, mechanically, the years with the highest prices — and the only available lever is the willingness to deploy less in them.