Are companies still treating AI spending as an experiment, or has it become a regular line item they budget for every year?

This looks at where enterprise AI money is going now: whether companies build their own AI tools or buy them, whether AI is still paid for as an experiment or as an ongoing operating cost, and what the money ends up paying for. The corpus shows two clear changes, both from a16z's 2025 survey of enterprise CIOs. AI has moved out of experimental budgets and into regular IT spending. Innovation funds made up about a quarter of AI budgets a year earlier and now make up only 7%, while leaders expect overall AI spending to grow about 75% next year Are enterprises actually spending more on AI that works?. Companies are also building less and buying more. Many in-house AI tools turned out to need constant model tuning and upkeep, and that maintenance cost more than the advantage the tools gave Are enterprises abandoning custom AI apps for third-party solutions?.

The move to buying makes more sense once you see what the vendors are paying for. Epoch AI's analysis of three AI labs found that compute for research and for running models takes 54–62% of operating costs, while staff costs stay under 25% Do compute costs really dominate AI company budgets?. Inside the data centers, the biggest cost is the servers themselves, not electricity. A one-gigawatt facility costs about $8.5B a year, and roughly $5B of that is hardware epoch-ai-estimates-a-one-gigawatt-ai-data-center-annualized-total-cost-of-owner. Prices also change quickly. A new state-of-the-art capability is expensive when it first appears, and then its cost falls by about two-thirds per quarter as competitors catch up Do AI performance prices fall faster when capabilities debut?. A company that builds its own tool on today's frontier model is building on something that will soon be cheaper and outdated. That is the maintenance burden the CIOs described.

Buying the software doesn't solve the harder problems, though. Evans argues that making tools easier to build ignores two barriers: most workers don't see their own tasks as things AI could do, and adopting AI means organizational decisions across departments, not just technical setup Does easier tool-building actually solve enterprise adoption problems?. OpenAI's ChatGPT Enterprise usage data shows the same unevenness. Adoption is concentrated in larger, research-heavy companies, and inside those companies marketing staff and early-career employees use the tool far more than executives do Who adopts enterprise AI first and how do they use it?. Even among companies that buy, the payoff differs from firm to firm. Firms with more tasks that AI can do replace contract labor with AI faster and more cheaply than other firms, which suggests that internal know-how keeps growing in value as a company uses AI more Do firms substitute labor for AI at different rates?.

That raises a question the budget figures don't answer: is the spending paying off? Brynjolfsson, Rock and Syverson's "productivity J-curve" gives one answer. Much of what a company actually spends to make AI work goes to training, redesigning processes and integration, and official productivity figures don't count that spending as investment. Productivity therefore looks weak at first and only later looks strong Why does AI productivity appear weak when investment is high?. A survey of 750 executives fits this pattern. They report bigger gains than the data shows so far, probably because revenue lags behind improvements in how work gets done Do AI productivity gains feel larger than they actually measure?.

Put together, AI's line item has become routine and its share of the budget is growing, while the costly and uneven work of getting value from it is spread across training, process changes and staff time that don't appear as AI spending. A company's AI budget may be the smaller part of what it actually spends on AI.