The most striking number from this week's earnings cycle was not in any single company's accounts. According to TrendForce, the five largest North American cloud providers (Google, Amazon, Microsoft, Meta and Oracle) will collectively spend more than $770 billion on AI infrastructure in 2026. That figure is roughly 87 percent higher than last year and is the clearest indication yet that what looked like a one-off buildout is settling into something closer to a permanent industrial base. Sundar Pichai confirmed his own slice of it at Google I/O on Tuesday, telling the audience that Google's capital expenditure will reach $180 to $190 billion this year, six times its 2022 level.
The composition of the spend matters as much as the total. TrendForce expects the five hyperscalers to account for more than 60 percent of global demand for Nvidia's new GB and VR rack-scale systems. AI training compute across the group will rise more than 56 percent, but inference compute is rising faster: up 122 percent year on year, on top of a 2025 base that already topped 37 exaflops. That ratio is significant. Inference is what runs in production, what serves the agents and chatbots and search results, and what generates revenue. The shape of the buildout suggests the hyperscalers now believe the commercial phase of generative AI has arrived.
A separate Lowy Institute analysis flagged by the data-science newsletter Let's Data Science captures what this means for the firms themselves. Big Tech is mid-way through a transition from asset-light software businesses to asset-heavy infrastructure operators, with fixed assets already accounting for about half of the major firms' balance sheets and total capex on a path toward $2 trillion. The companies that built the early internet by leveraging cheap commodity servers are now, in effect, becoming utilities. Long depreciation horizons, exposure to energy and supply-chain risk, and reliability engineering at a physical scale are becoming central to their operating profile in ways they were not five years ago.
Two parallel strategies are visible underneath the headline number. The first is custom silicon. TrendForce projects Google's TPU shipments will rise nearly 80 percent in 2026 as it transitions from TPU v7 to v8, and Pichai used I/O to introduce a dual-chip approach with separate variants for training and inference. Amazon's Trainium accelerators will account for more than 40 percent of its own AI server shipments. The second is energy. New AI racks are forcing the adoption of liquid cooling as standard, and server power consumption growth for the top five providers is set to jump 18 gigawatts in 2026, a 116 percent increase on the prior year. That is comparable in scale to the entire generating capacity of a mid-sized nation, added in a single year, for five companies.
The vulnerability in this picture is not capital. With combined operating cash flows in the hundreds of billions, the hyperscalers can afford the equipment. The constraint is electricity and the political durability of public consent to provide it. As AI Daily reported on Wednesday, U.S. opposition to data centers has hardened sharply, with 20 projects cancelled in a single quarter and elected officials being removed in towns that approved them. Wholesale power costs are projected to rise 6 to 29 percent by 2030 because of data center demand, with Virginia exposed to a possible 57 percent jump. The same Lowy framing applies in reverse: a quasi-utility cannot operate without a functioning grid and the social license to draw from it.
That places the $770 billion figure in a sharper light. The argument for spending it is straightforward: inference demand is doubling, every major enterprise software category is being rebuilt around agents, and whoever owns the cheapest tokens at the moment of use captures the resulting market. The argument against is that the bottleneck has migrated from chips, where money buys solutions, to grids and permits, where it largely does not. The five companies betting on the first argument are about to spend a sum equivalent to the GDP of Switzerland to find out which view is correct.