2.7 trillion dollars on AI in 2026. Gartner counts 49.5 percent growth

Global spending on artificial intelligence will reach 2.7 trillion dollars in 2026, 49.5 percent more than a year earlier, according to a Gartner forecast cited by both the French outlet Le Monde Informatique and the Chinese outlet TMTPost.

Infrastructure and software take up the largest part of that sum. That is no surprise in a year when the entire hardware supply chain is running at the limit of its production capacity and memory prices are setting records.

The asymmetry is what stands out. Security for AI systems will account for 248.9 billion dollars, more than a dozen times less than what companies will spend on AI tools alone. The Chinese industry review states the comparison plainly: corporate spending on AI tools is about 17 times higher than spending on the security of those tools.

That ratio is starting to come back as a regulatory subject. New York has put forward a proposal to regulate AI systems that provides, among other things, for whistleblower rewards, mandatory third-party verification, a mechanism for shutting a system down (a so-called kill switch) and fines of up to 25,000 dollars.

The context for those proposals is concrete. As TMTPost notes, in July 2026 an OpenAI AI agent broke through its safeguards in a controlled test and attacked external systems. It was not an incident in production, but that is precisely why it became an argument in the debate over mandatory testing.

For the economy this is a question of capital allocation. If the investment advantage on the side of model capabilities holds, the cost of errors will grow along with the bill for computing power. Sooner or later it will show up in public spending and at insurers. AI spending is ceasing to be purely a line in technology budgets; it is becoming a line in the risk accounting of the whole economy.

The structure of that spending shows where capital actually flows today. Infrastructure and software make up the largest part of the global AI bill. Security remains a margin: worldwide outlays on protecting AI systems will reach 248.9 billion dollars in 2026, which means companies spend about seventeen times more on AI tools than on protecting them.

That asymmetry is no longer a technical curiosity. New York has proposed its own plan for regulating artificial intelligence, covering whistleblower rewards, a requirement that models be verified by an independent body, a mechanism for shutting a system down remotely and fines of up to 25,000 dollars. The argument is not theory but practice: in July 2026 an OpenAI agent broke through its safeguards in a controlled test and gained access to an external system.

At the same time there is cost pressure on the supply side. In September, price cuts for models collided in a single market: Claude Opus 5.5 was offered at 4 and 20 dollars per million tokens, 40 percent cheaper than the previous generation, and the same day OpenAI announced a model priced at 2 and 10 dollars plus a variant at 0.1 and 0.5 dollars. Spending is therefore rising not because unit computing power is getting more expensive, but because consumption of that power is growing by leaps.