FundaAI released a research report on enterprise AI applications, indicating that enterprise AI budgets are still expanding, but there is a divergence in trajectories in the second half of 2026 and 2027. The AI spending guidance from large U.S. telecom operator A shows an increase from a baseline of 100 in January to about 190 in December, with an expected year-on-year increase of 40%–50% in 2027; large European automaker A has only increased by 10%–15% so far this year, with guidance for next year remaining roughly flat.

Incremental spending is shifting from paid seats to API/Token consumption and production workflows. The aforementioned telecom operator’s subscription and API ratio has changed from about 50%/50% to 40%/60%, and it may trend towards 35%/65%; mid-to-large biopharmaceutical company A has adjusted from 80%/20% to about 70%/30%. Open-source adoption is uneven, with active scenario usage accounting for 30%–40%, as the unit price is lower, leading to a smaller spending proportion; experts estimate that open-source inference can be about 40%–70% cheaper than closed-source cutting-edge models, with the gap narrowing to 20%–40% under full cost metrics, and model routing, caching, and context compression could further reduce API spending by about 20%–30%.

On the production side, AI budgets are increasingly built from the bottom up based on workflow ROI. The typical production ROI for this telecom operator is about 1.5–2 times, with a payback period of 6–18 months, and mature use cases can reach 3–5 times. The next wave of spending is related to agents, software modernization, network operations, commoditized workflows, and longer-cycle business processes, but engineering capacity, process reengineering, governance, and data readiness are becoming tighter constraints than funding.