The AI Capex Gap Every Board Should Be Studying
AIMG’s August 2026 Special Report, The Economics of AI, frames this cycle as a paradox, not a bubble: industrial-scale capital deployment alongside returns that have not yet arrived. Global AI spending is on course for USD 2.52 trillion in 2026, up ~44 percent, split across USD 1.37 trillion of infrastructure, USD 452 billion of software and USD 588 billion of services. Aggregate CY2026 hyperscaler capex has been revised up to exceed USD 900 billion, a 101 percent rise, heading beyond USD 1.2 trillion in CY2027. Yet 95 percent of enterprise generative AI pilots deliver no measurable ROI, and only 5 percent reach scaled production.
The spend is now macro-material
Global data-centre spend passes USD 650 billion in 2026, up 31.7 percent, with capacity nearly doubling from 114GW to 226GW by 2030 and cumulative capex reaching up to USD 5 trillion. AIMG puts top-five hyperscaler spend at roughly 2 percent of annual US GDP, which is why the BIS, the IMF and central banks now treat this as a financial-stability question. FactSet’s rising capex guidance, Epoch AI’s 0.8 percent-of-GDP estimate and Bridgewater’s ~140bp 2026 growth contribution point the same way.
The constraint has moved off silicon
AIMG’s key supply-side finding: capacity planning is now gated by long-lead power, grid, cooling and networking components, not chips. Microsoft has noted two-thirds of its capital spend goes to short-lived GPU and CPU assets. That short-duration mix is exactly why the IMF warns on “rapid capital obsolescence risks.”
Concentration is the sharper risk
AIMG records USD 147 billion of AI venture funding in Q2 2026 and USD 387.5 billion in H1, already above 2025’s full-year USD 264.1 billion. OpenAI and Anthropic alone took roughly 47 percent of H1 AI funding; mega-rounds above USD 100 million took 89 percent of capital raised; sovereign wealth funds contributed ~USD 46 billion. CB Insights (81 percent of Q2 funding in mega-rounds, deal count at a decade low), Global SWF and Invesco (52 percent of SWFs cite concentration as their top portfolio risk) corroborate it. The IMF’s structural read is the same: scale economies in frontier models raise entry barriers and concentrate power in a handful of firms.
Why capex rises as prices fall
AIMG’s answer is a Jevons paradox in cognition. Token pricing has fallen 5x to 10x a year, but cheaper cognition expands viable workloads. Multi-agent systems consume up to 15x more tokens per task than a single query, and total cost is dominated by retrieval, vector search, embeddings, caching and governance, not the model invoice. Falling unit prices drive volume rather than deflation.
Returns lag by design
AIMG frames flat productivity as a diffusion lag: innovation is outrunning the reorganisation firms must complete. The eventual dividend is 0.5 to 0.75 points a year on trend productivity, but not on a quarterly cadence. And 60 to 70 percent of AI economic value accrues in the workflow, retrieval, governance, evaluation and distribution layers – what AIMG calls the productivity harness – not the model. BCG’s 10-20-70 rule lands in the same place.
Put the two facts together. Capital is committed on a two-year cycle against a value curve measured in three to five years, funded increasingly by debt and private credit. That is the gap behind the BIS warning that “disappointment in returns could trigger a sudden pullback in financing and turn the capex boom into a protracted investment bust,” and its flag on poorly disclosed circular financing between chipmakers, hyperscalers, labs and neoclouds. McKinsey’s USD 6.7 trillion to 2030 and Morgan Stanley’s USD 1.5 trillion financing gap size how much must come from outside operating cash flow.
AIMG View. Not a bubble call. A concentration, leverage and diffusion-lag call. The value is accruing in a harness layer that almost no capex line names, and the same few firms sit on every side of the trade.
AIMG Bottom Line. Four questions for boards with hyperscaler or AI-supplier exposure. What do our numbers look like under a two-year capex pause? What is our indirect exposure via private credit, ABS and sovereign co-investors? Does our valuation assume no repricing of AI infrastructure debt? And first: is our own AI spend going into the harness or the model? That determines whether we benefit from this cycle or merely finance it.
Sources: The Economics of AI, AIMG Research