Investment flooded in faster than grids could absorb it, models kept doubling on schedule, and the industry quietly split into a frontier tier and everyone else. A field report from the twelve months that ended the debate about whether AI was a sector at all.
Global spending on artificial intelligence crossed the $1.2 trillion mark in the twelve months ending June 2026, according to a tally assembled from 247 disclosure filings, earnings transcripts, and government budget documents reviewed by this desk.
The figure — a 38% jump over the prior year — exceeded the most aggressive Wall Street forecast by nearly $180 billion and confirmed what every boardroom had already begun to suspect: that AI is no longer a line item but the line item.
Roughly 61 cents of every dollar flowed into three buckets — frontier model training, data-center construction, and the silicon that powers both — leaving software, services, and the long tail of applications to fight over the remainder.
"The capital is no longer the constraint," said one infrastructure investor granted anonymity to discuss live deals. "Power, permits, and people are."
Compute remained the dominant story. Nvidia alone booked more revenue in the period than the entire global semiconductor industry did in 2014, and its order book at quarter's end stretched into late 2027.
The build-out of new data-center capacity, meantime, set records of its own. Construction starts for dedicated AI facilities exceeded 22 gigawatts of contracted power — roughly the nameplate capacity of Sweden's entire grid.
Yet for all the headline numbers, the year's most consequential shifts happened quietly. Three bottlenecks matured into structural problems the industry can no longer treat as transitional.
The first is power. Grid interconnection queues in three of the largest U.S. markets now average 4.7 years. The second is talent. Senior research engineers command compensation packages north of $8 million. The third — and least visible — is data.
The data wall moved closer. A study published in May estimated that the share of high-quality public text remaining for pretraining, on current consumption trajectories, would be exhausted by mid-2027.
Labs responded by accelerating investments in synthetic data, multimodal corpora scraped from video, and proprietary licensing deals with publishers — including, notably, this newspaper's parent company.
Regulators, meanwhile, moved at their own pace. The EU's AI Act entered full enforcement in August 2025; the U.S. issued no comprehensive federal statute, leaving a patchwork of state rules and executive orders that one senator called "a regulatory Moiré pattern."
Below, the desk's full accounting of the year — by the numbers, by the deals, and by the moments that mattered.