JPMorgan lifts AI infrastructure estimate to $5.5T

The bank’s new forecast is a meaningful step up from its prior $5.1 trillion estimate. It now sees as much as $4.1 trillion of that being funded through debt, reflecting higher loan-to-value structures and a deeper role for credit markets in the data-center race. For investors, that is crucial: it means the AI infrastructure trade is not confined to a handful of megacap balance sheets. It is spreading across investment-grade bonds, project finance, construction lenders, equipment suppliers and the semiconductor complex.

The concentration of spending is what makes the opportunity attractive and the risk asymmetric. JPMorgan expects the biggest US hyperscalers to spend about $697 billion in 2026, up $173 billion from the start of the year, and sees that figure climbing further as demand for AI compute expands. The bank also projects their combined operating cash flow could top $900 billion by 2027, giving the sector more internal financing power than skeptics assume. That is the market’s blind spot: investors often look at the headline capex and assume strain, when the better lens is the pace at which AI revenue, cloud usage and enterprise adoption are swelling the cash engine behind the buildout.

The financing data back up the thesis. JPMorgan expects investment-grade bond issuance tied to data centers to contribute more than $2.1 trillion over five years, including roughly $150 billion from US hyperscalers in 2026 alone and another $100 billion from outside the US. It also sees another $170 billion coming from data-center and chip issuers outside the core investment-grade group. That is a massive capital-markets pipeline, and it creates clear winners: lenders, underwriters, utilities, grid operators, networking vendors and the most efficient AI hardware suppliers.

The geography matters too. JPMorgan says the US still captures about 85% of AI and machine-learning venture capital, while China, South Korea and Taiwan benefit as second-order winners through the semiconductor supply chain. That reinforces the investable map for the next leg of the cycle. In the US, the best positioning is in hyperscaler enablers and the infrastructure stack. In Asia, it is in foundry, memory and advanced packaging exposure. The market underestimates how much of the AI spend will leak into the broader industrial economy through power, cooling and construction.

The most important signal is that this cycle is being financed, not just funded out of equity optimism. JPMorgan pointed to deals it helped structure in 2026, including $9.6 billion of construction loans for the Stargate data-center project in Abilene, Texas, a $4.25 billion bond for Hut 8’s Beacon Point facility that was 95% loan-to-value financed, and a $5.25 billion debt package for CoreWeave. Those transactions show that lenders are willing to stretch for high-quality AI-linked assets, which is exactly what you would expect near the front end of a secular infrastructure boom.

For investors, the message is not to chase every AI name. It is to own the companies that charge rent on the buildout and to avoid the businesses that need perfect demand to justify their valuations. Nvidia, Microsoft and AMD remain key bellwethers for the compute cycle, but the broader trade is now extending to networking, data-center infrastructure, power equipment and specialty financiers. Even after recent volatility in Nvidia, Microsoft and AMD, their price action still reflects a market that is wrestling with durability, not collapse.

My view is simple: JPMorgan’s upgrade confirms that AI capex is entering its monetization phase, and that makes the next winners less about who invented the model and more about who supplies the physical backbone. If the bank is right, the opportunity is not just in AI stocks — it is in the picks-and-shovels ecosystem that benefits every time another trillion dollars is pushed into compute.