News · 2026-06-26
Big Tech is set to spend up to three-quarters of a trillion dollars on AI in 2026
Behind every chatbot reply is a staggering amount of physical infrastructure, and the bill for building more of it in 2026 is coming into focus. Projected capital spending by the big cloud and AI companies runs into the hundreds of billions of dollars for the year, with combined estimates for the largest players landing in the range of roughly two-thirds to three-quarters of a trillion dollars, and some broader industry tallies running higher. A funding and capex roundup collects the figures, and a notable wrinkle is how it is being paid for: increasingly with borrowed money, as Big Tech issues large amounts of bonds to fund the buildout.
The background a newcomer needs. Capital expenditure, or capex, is money spent on long-lived physical things: in this case data centers, the buildings full of specialized computers that train and run AI models. Training a frontier model and then serving it to millions of users requires vast clusters of expensive chips, plus the power and cooling to keep them running. For years these companies funded such spending out of their enormous profits. The shift now is that the appetite has grown so large that they are turning to debt markets, the same way utilities and telecom companies historically borrowed to build power grids and networks.
What is actually happening. The largest cloud providers are each planning to spend somewhere between many tens of billions and around two hundred billion dollars in a single year, much of it AI-related. Separately, OpenAI has moved into designing its own chip, announced with Broadcom and nicknamed Jalapeno, with a claim that it can make running models, the inference step, substantially cheaper. Designing your own silicon is the strategy Google, Amazon, and others have pursued to escape paying a premium to outside chip vendors and to tune the hardware to their exact workloads.
An analogy. The financing shift is the AI industry growing up from a cash business into a capital-intensive one, like the difference between a software startup that runs on a credit card and a railroad that floats bonds to lay track across a continent. When a sector starts borrowing at this scale to build physical assets, it is betting that the demand will be there for decades, the way railroads, electric utilities, and telecoms once did. That is a vote of confidence, and also a source of risk, because debt has to be repaid whether or not the demand materializes on schedule.
The custom-chip move has its own logic. Running a popular model is like running a toll road: every user query costs a little electricity and compute, and at billions of queries those pennies become the dominant expense, often larger over time than the one-time cost of training. If OpenAI can design a chip that serves its own models more cheaply, it lowers the toll on every trip, which matters enormously in a week when frontier model prices are rising and the economics of inference are under scrutiny. For the difference between the two phases, see our explainer on training versus inference.
Why it matters: this spending is the physical foundation under everything else in AI, the government-gated model launches, the talent wars, the open-versus-closed pricing fight. The scale, and the move to debt financing, signals that the major players are treating AI infrastructure as core, decades-long industrial capacity rather than a speculative bet. It also concentrates power, because only a handful of companies can marshal this kind of capital, which is part of why the frontier keeps consolidating into a few hands.
The honest caveat: most of these numbers are projections and estimates, not audited results, and they vary widely between sources, so treat the totals as a range rather than a precise figure. The custom-chip claims are shakier still: OpenAI's cost-saving figure for Jalapeno is a vendor claim with no independent benchmarks or real-world deployment data yet, and a full technical report is still awaited. History is also full of infrastructure booms that overbuilt, the late-1990s fiber glut being the classic example, where the demand eventually came but arrived years after the debt did. Enormous, debt-financed buildouts are a bet on the future, and bets can be early, or wrong, even when the underlying technology is real.