Global Investment Outlook 2025 — The Bifurcated Market

The most accurate description of 2025 is that there was no single market. There were two, operating under one name, with sharply different conditions.

In one, capital was abundant. Companies with credible positioning in artificial intelligence — at the model layer, the infrastructure layer, or in applications with demonstrable adoption — raised large rounds quickly, often pre-emptively, at valuations that would have looked unremarkable in 2021. Diligence periods compressed. Competitive dynamics between investors returned.

In the other, conditions remained close to those of 2023. Companies without a thematic hook faced extended processes, sceptical diligence, close scrutiny of capital efficiency, and valuations well below their prior marks. Down rounds and structured terms remained routine.

Both descriptions are accurate. This creates a genuine analytical problem, because most market data is reported as an aggregate — and an aggregate drawn from a bimodal distribution describes a condition that few participants actually experienced. A statement that funding was flat year on year, in this environment, is arithmetically true and practically meaningless.

For allocators, 2025 was a year in which market-level statistics became substantially less useful than segment-level ones, and in which the definition of the segment — what counts as an AI company — did much of the analytical work.

Why averages stopped describing anything

This is worth setting out carefully because it affects how every other 2025 figure should be read.

A market average is informative when outcomes cluster around it. If most companies experience conditions near the mean, the mean is a reasonable summary.

When a distribution is bimodal — a cluster at each end with relatively little in the middle — the average sits in the gap between them. It describes a condition that comparatively few participants occupied.

2025's venture market was bimodal. A statement like "median valuations were roughly flat" could be simultaneously true and misleading, because it averaged a segment where valuations rose sharply with one where they fell.

The practical consequences were immediate:

• Founders benchmarking against market data received systematically wrong guidance, in one direction or the other depending on which segment they occupied.
• LPs assessing manager performance against benchmarks faced a similar problem: a manager's return said as much about their thematic exposure as their selection skill.
• Anyone using market data to time a raise was working from a number that did not describe their situation.

In a bimodal market, the average is the one value that describes almost nobody. Segment-level data stops being a refinement and becomes a requirement.

The AI question changed shape

Through 2023 and 2024 the central question about AI investment was whether the technology would prove commercially significant. By 2025 that question had largely been settled in the affirmative — adoption was measurable and revenue was real.

The question that replaced it was harder: where in the value chain does the value actually accrue?

This is a different kind of question, and it is underwritten differently. The first is technology risk: will this work? The second is competitive-position risk: given that it works, who captures the economics?

Several possible answers were live simultaneously, and they are not compatible with one another:

• Value accrues to the compute layer — semiconductors, data centres, power. This is the position with the clearest revenue but the heaviest capital requirement and the most obvious competitive response.
• Value accrues to the model layer. This depends on models remaining differentiated. If capability converges, models become a commodity input and pricing power erodes.
• Value accrues to the application layer. This depends on applications building genuine defensibility — workflow integration, proprietary data, distribution — rather than being thin wrappers over a commoditising input.
• Value accrues to incumbents who embed the technology into existing distribution and customer relationships, capturing the benefit without any of the platform economics.

The uncomfortable observation is that capital was deployed across all four theses simultaneously, at valuations implying each would capture substantial value. They cannot all be right, because they are partly claims on the same pool of economics.

That does not make the aggregate investment irrational — a portfolio spanning the chain is a reasonable response to genuine uncertainty. But it does mean that the aggregate capital deployed is unlikely to earn the return implied by summing the individual cases.

Exits reopened, narrowly

Listing activity improved from the trough. This was genuine, but its shape matters more than its existence.

The window that opened was narrow. It admitted large companies with profitability or a clear path to it, in favoured sectors, with scale sufficient to support institutional trading. It did not open for the broader cohort of companies that raised in 2021 at valuations that would not survive public scrutiny.

For those, three paths remained: continue extending runway while growing into the valuation, accept a sale at a price below the last private mark, or restructure. Each was occurring, mostly without announcement.

The consequence for institutions was that distributions improved somewhat but remained below the level needed to normalise commitment pacing. A partial reopening relieves the pressure without resolving it — which, for planning purposes, is a materially different situation from either a closed market or a functioning one.

Secondaries and continuation vehicles remained structural. Having become permanent infrastructure in 2023, they did not recede when conditions partially improved. This appears to be a durable change in how private capital manages duration rather than a crisis response.

The early-stage cohort gap

The most consequential structural development of 2025 was also the least visible, because it concerns something that did not happen.

Seed and pre-seed formation remained constrained for a third consecutive year outside the AI segment. Capital that would historically have funded a broad base of early-stage companies concentrated instead into fewer, thematically-favoured ones.

The effect is deferred, which is why it attracts little attention. A shortfall in seed formation does not show up in Series A data for two to three years, and in Series B data for four to five. The pipeline thins from the bottom, and the consequence appears only when the affected cohort should have been raising.

This produces a predictable dynamic: a future period in which capital is available at growth stage but the supply of quality companies at that stage is unusually thin — because they were never funded. That is a setup for exactly the sort of valuation inflation that occurs when too much capital chases too few qualifying assets.

The pattern is well established in venture history and reliably underappreciated at the time, because the cost is invisible in the year it is incurred.

How to read a bimodal market

If averages describe nobody, the practical question is what to use instead. The answer is a small set of substitutions, each of which is available in published data.

Substitute count for value where the question is about access. Deal value measures how much money moved, which in a concentrated market is a statement about a handful of transactions. Deal count measures how many companies were funded. For a founder assessing whether capital is available, or an early-stage investor assessing pipeline health, count is the relevant series and value is actively misleading.

Substitute medians for means. A mean in a distribution with a long tail is pulled by the tail. The median round size in 2025 and the mean round size describe different markets, and the gap between them is itself a useful measure of how concentrated the market has become.

Substitute segment-level data for market-level data. This is the largest and most demanding substitution. It requires defining the segments, which is where the analytical difficulty sits — "AI company" has no agreed boundary and the definition materially affects every resulting statistic. But an approximate segmentation is more informative than an exact aggregate, because the aggregate is measuring a composite that does not exist.

Substitute distribution shape for point estimates. The most informative single question about a market statistic in 2025 was not its level but its dispersion. A market where the interquartile range widened substantially is a market where the median has stopped being a summary.

Check whether the two sides moved in opposite directions. The diagnostic for bimodality is straightforward: if the top quartile and bottom quartile moved in opposite directions over the period, the aggregate is averaging two conditions. This is computable from any dataset that reports quartiles and is rarely done.

An aggregate statistic is a summary when the underlying distribution clusters and a fiction when it does not. The test is cheap and almost nobody runs it.

The same substitutions apply to manager assessment. A fund's 2025 return said as much about its thematic exposure as about its selection skill. Assessing a manager against a market benchmark in a bimodal market compares them to a number that describes neither segment. The more informative comparison is against managers with similar exposure — which requires knowing what the exposure was.

What allocators actually did

Several behavioural patterns were visible through the year.

Thematic concentration was accepted deliberately. Rather than treating AI exposure as a risk to be diversified away, many allocators concluded that under-exposure was the greater risk, and sized accordingly.

Manager selection weighted access over strategy. In a market where the best opportunities were heavily competed, the practical question became which managers could get into rounds — a return to the access-constrained dynamic last seen in 2021.

Liquidity planning became explicit. After several years of constrained distributions, institutions modelled liquidity as a primary variable rather than assuming it.

Vintage diversification received renewed attention. The experience of 2021 and 2022 vintages made the case for consistent deployment across years, rather than attempting to time entry, considerably easier to argue internally.

Why bifurcation is the normal consequence of a shock

A pattern worth naming, because it runs against the intuitive expectation and appears repeatedly across this archive.

Shocks are widely assumed to level. A common disruption affects everyone, so the reasoning goes, and differences between participants should narrow.

They usually concentrate, because the capacity to respond to a shock is itself unevenly distributed. Three examples from the archive make the point:

• The 2020 Asia-Pacific report describes digital adoption accelerating fastest where infrastructure already existed. Markets without the infrastructure experienced far less acceleration, because the constraint was the base to accelerate from. The gap widened.
• The same report describes fiscal capacity determining the policy response available. Economies with capacity limited their damage; those without absorbed more permanent damage. The gap widened.
• The 2024 US venture report describes the emerging manager pipeline thinning while established firms raised more easily. The constraint that reduced total commitments advantaged incumbents by the same mechanism. The gap widened.

2025's bifurcation is the same pattern applied to a capital cycle. A contraction in available capital did not reduce funding proportionally across the market. It concentrated what remained into the segment with the strongest thematic case, which meant the rest experienced a contraction larger than the aggregate figure implied.

The mechanism is general: when a resource becomes scarce, it flows toward whatever has the strongest claim on it. That claim may be thematic momentum, existing relationships, demonstrated performance, or institutional capacity. In every case the participants who already had the most of it end up with proportionally more.

The aggregate figure in a concentrating market describes the winners. Everyone else is in the residual, and the residual is where most of the participants are.

The practical implication for 2026 and beyond is that any statement about a market recovering should be checked for whether the recovery is broad. A rising aggregate with a falling count is not a recovery — it is further concentration, reported as improvement.