A $2.5trn question

There is a $2.5 trillion question hanging over venture and the broader technology market, and it is the one every serious allocator is now being forced to confront: how much of the capital pouring into artificial intelligence will actually earn a return, and how much is being spent on infrastructure and models that may never pay for themselves. The figure is large enough to reshape portfolios, and the honest answer is that no one yet knows.

The scale of AI investment has reached a point where it is no longer a thematic bet but a macro force. Vast sums are flowing into data centres, chips, model training and the companies building on top of them. Some of that spending will underpin the defining businesses of the next decade. Some of it will turn out to be capacity built ahead of demand that never fully arrives, or models commoditised faster than their builders can monetise them. Telling the two apart in advance is the central challenge.

The capital-intensity problem

What makes the question acute is how capital-intensive this wave has become. Earlier technology cycles were often defined by asset-light software businesses that scaled with minimal incremental cost. The current AI build-out is the opposite: it demands enormous up-front spending on compute and infrastructure before any return is visible. That changes the risk profile entirely. When success requires billions in capital expenditure, the margin for error narrows, and the cost of building for demand that does not materialise rises sharply.

For investors, the capital intensity raises the stakes on every judgement. Backing the right company in an asset-light cycle is forgiving; the losers cost little. Backing the wrong infrastructure thesis in a capital-heavy cycle is expensive, because the money is committed to physical assets and long contracts that cannot easily be unwound. The $2.5 trillion question is really a question about where, in a hugely expensive build-out, the durable economics will actually sit.

Picking the layer that captures value

Part of the answer lies in which layer of the AI stack ends up capturing value. Infrastructure providers, model builders and application companies all have different economics, and history suggests value does not distribute evenly across a stack. In past cycles, some layers commoditised quickly while others captured durable margins. Whether the foundation models become commoditised utilities or remain defensible franchises, and whether the lasting value accrues to chips, clouds, models or applications, will determine which slice of the $2.5 trillion earns its keep.

This is where discipline separates investors. The temptation in a boom is to fund every layer indiscriminately, on the logic that AI is transformative and therefore everything connected to it will win. The more rigorous approach is to ask, for each investment, where the defensible economics are, who can be displaced, and what happens to margins as capacity and competition increase. Those questions are harder to answer in a hype-driven market, which is exactly why they matter most.

What it means for capital

The signals for allocators are sobering and useful in equal measure. First, the sheer scale of AI investment means it now carries macro-level risk, not just thematic risk, and a misjudgement at this size can move whole portfolios. Second, the capital intensity of the cycle raises the cost of being wrong, demanding more discipline than the asset-light booms that preceded it. Third, returns will concentrate in specific layers of the stack rather than spreading evenly, so the question is not whether to invest in AI but where.

For founders, the implication is that capital remains available but scrutiny is rising, and businesses with a credible path to durable economics will stand out from those riding the theme. For investors, the $2.5 trillion question is a discipline test disguised as an opportunity. The capital is being deployed at a pace that guarantees both extraordinary winners and expensive failures. The allocators who do the unglamorous work of figuring out where the returns actually sit, rather than betting that all of AI will pay off, are the ones most likely to be on the right side of the answer.