AI’s biggest problem is no longer just access to computing power. In VentureBeat’s October survey of 156 qualified AI professionals, 38% said their organization delayed, scaled back, or cancelled an AI project because it cost more than expected. Only 12% cited insufficient computing capacity—making cost overruns more than three times as common.
The “AI honeymoon” is ending as budgets meet business reality. Even though 79% of respondents expect their AI budgets to rise over the next year, only 26% say their organization measures AI spending through a financial return that finance can defend. In other words, spending is growing faster than proof of value.
Costs extend far beyond the price of a model. Agents can repeatedly query company data, retain context, generate monitoring records, and require human review; each step adds expense. Uber’s feature-flag cleanup example shows the value of measuring cost per completed task: it reported a 90% reduction in cost per cleanup after benchmarking and choosing an appropriate model.
The winners will be those who treat AI as an economic system, not a technology demo. Companies should measure completed work, quality, rework, human intervention, infrastructure use, and total operating cost — not just token prices. Open models, gateways, and flexible platforms can help, but they save money only when they improve the cost and quality of real outcomes.
THE HONEYMOON IS OVER—AI MUST NOW EARN EVERY RUPEE, EVERY GPU, AND EVERY BUSINESS CASE.
Sanjay Sahay
Have a nice evening.

