From the chip to the chiller
Behind every AI computation lies a chain of physical infrastructure, starting with a single watt of electricity. It begins at a processor, the component that attracts the most attention, but to reach that chip, it must pass through cooling systems, power distribution units, transformers, substations and, ultimately, a generating plant. Each layer presents a business opportunity. In 2026, however, the layers furthest from the chip are demonstrating some of the strongest economics.
There is one number buried in Microsoft’s disclosures that tells the story of the AI trade in 2026. The company is sitting on a US$80 billion ($103 billion) backlog of Azure orders that it cannot fulfil, not for want of demand, capital or chips, but because it cannot secure enough power.
That single constraint has fundamentally reshaped the investment case. For the past three years, the market viewed AI primarily as a semiconductor story, with the debate hinged on whether hundreds of billions in promised spending would ever convert into revenue. That debate is largely settled. The world’s fourteen largest data centre operators will spend close to US$750 billion this year, up from less than US$450 billion in 2025. This capital has already been committed and financed and is being deployed. The more pressing question now is whether the physical infrastructure can keep pace, and who stands to benefit if it does.
