Artificial Intelligence for Investors
Understanding artificial intelligence before valuing it. Independent macro research on what the data-centre build-out means for growth, inflation, rates and asset classes. Last updated: September 2026.
If you never bought an AI position, you still hold one
The build-out of AI data centres has become large enough to move growth, inflation, monetary policy, and bond and equity markets. That is why artificial intelligence is one of our six structural drivers shaping the macroeconomic environment and financial markets in the years ahead — alongside geoeconomics, China Shock 2.0, productivity, demographics, and fiscal policy and debt.
This is not a sector story. It is a macro event that reaches investors who hold no technology exposure at all.
The essentials in brief
- A macroeconomic event, not a sector story. The five large US technology groups are investing over USD 800bn in data centres in 2026, after roughly USD 420bn in 2025. In the first quarter of 2026, three quarters of US growth came from this spending — around 40% after adjusting for imports.
- The build-out is already in the major indices. Anyone holding a global equity index holds a good fifth of their equity risk in the four largest operators and the six largest suppliers — both hinge on the same investment decision. The same operators are on track to overtake banks as the largest group of borrowers in the US investment-grade index.
- These are two businesses, not one. Building a model is a one-off investment; running it is an ongoing unit-cost business. In 2026, operation costs more than training for the first time — which makes utilisation the decisive variable for returns.
- Physics, not money, sets the pace today. Current-generation compute chips are allocated through 2027, memory is scarce, and over 90% of leading-edge fabrication sits in Taiwan. A new grid connection for a data centre takes three to seven years in the US, depending on the region.
- Whether the maths works can be checked. The installed capacity needs to earn roughly USD 260bn a year to cover depreciation, operation and cost of capital. Per million tokens processed, that is about USD 2.20 — while USD 1.02 to 1.48 comes in. The gap has so far been closed by rapidly growing volumes.
- The cycle would turn on financing rather than demand. Capital is plentiful today, but it is the first thing to give way when expectations shift.
Our AI research series
Core thesis. The investment boom is real, physically constrained and initially inflationary. The bottleneck sits not in demand but in manufacturing capacity — and increasingly in financing.
The eight themes
- The AI capex cycle — hyperscaler spending and the export channel into Taiwan and South Korea
- Is AI inflationary or disinflationary? — investment phase versus maturity
- Chinese models and US–China systems competition
- AI usage and monetisation — price per token versus volume
- Capital requirements — bond and equity issuance, real yields
- Compute as an asset class — rental rates, residual values, financing vehicles
- Strategic and tactical market view
- The three indicators used to test the base case
The three early indicators. We review the base case continuously against: rental rates for GPU hours, the depreciation schedules of the hyperscalers, and the prices of older memory chips.
Note: Our research is intended for professional and sophisticated investors. It contains general macroeconomic assessments and constitutes neither investment advice nor recommendations regarding individual securities.
Frequently asked questions
Does the AI build-out affect investors who hold no AI stocks?
Yes. It works through growth, inflation and interest rates on every asset class. It is also embedded in any global equity index: a good fifth of equity risk sits with the largest operators and suppliers. Those same operators are on track to overtake banks as the largest group of borrowers in the US investment-grade bond index.
Does the compute capacity already built actually pay for itself?
Not yet in full. Around USD 2.20 per million tokens processed is required; between USD 1.02 and 1.48 actually comes in, depending on how it is measured. In aggregate the installed base would need to earn roughly USD 260bn a year; a little over half of that is demonstrable. The gap has so far been closed by strongly growing volumes.
What separates training from inference — and why does it matter economically?
Training is the one-off creation of a model; inference is running it. In 2026, operation costs more than training for the first time. Utilisation therefore becomes the key variable — much as in a factory with high fixed costs.
How would you recognise a turn in the AI investment cycle?
First in the rental price of the current chip generation: if it falls markedly while volumes keep rising, more has been built than is needed. The second early indicator is financing, which is the first thing to give way when expectations shift.
Who is this research written for?
For anyone analysing the economy or basing investment decisions on it: researchers, strategists, traders and investors across money, bond, currency, commodity and equity markets — and anyone allocating a portfolio.
About the author
Daniel Pfändler is managing director of Research Ahead GmbH and a macro strategist with decades of experience as an economist and market strategist. He has repeatedly been ranked for his macro and rates research and writes a regular column for „MeinungsMacher" at Manager Magazin. The AI research series is part of his macro research for institutional investors.
For questions and disagreement: daniel.pfaendler@researchahead.com