Gartner: AI Inference Costs Per Agentic Workflow Will More Than 5x Through 2028

Gartner is warning enterprise leaders not to bank cost relief on falling per-token prices: inference costs per agentic workflow will rise more than fivefold through 2028, even as unit token economics keep improving. The firm calls this the “Inference Paradox” — three forces compound faster than price-per-token drops. Foundational model costs keep falling, which lets teams justify deploying more powerful, more expensive models. And agentic workflows themselves burn far more tokens than a chatbot exchange, since an agent must continuously reason, negotiate, and self-check rather than interpret one query and return one answer. “Product leaders cannot rely on more efficient token economics to rationalize AI costs,” said Will Sommer, Senior Director Analyst at Gartner, “because each successive generation of AI capability will necessitate more, and often more expensive, tokens.” Gartner’s fix is a multimodel ecosystem with inference-tiering, routing, and orchestration calibrated to task complexity — warning that defaulting to generic autonomous intelligence risks unbounded costs.

The forecast lines up with a structural shift Gartner flagged in its own AI-infrastructure spending numbers a week earlier: an August 10 forecast projected worldwide AI-optimized IaaS spending to grow 96% in 2026, to $42 billion. Buried in that forecast is the same trend restated in dollars — 2026 is the first year global inference spend ($23.3 billion, 55% of the total) surpasses training spend ($19 billion, 45%), as enterprises shift from building models to running them continuously in production. Gartner analyst Hardeep Singh tied that shift directly to agentic AI, since multistep autonomous execution drives sustained, real-time inference rather than the periodic bursts tied to model training.

For consulting and sales organizations setting client-facing AI ROI expectations, the two forecasts read as one story: the cost center of enterprise AI has already moved from training to inference, and Gartner’s own numbers say that shift is compounding, not stabilizing, through 2028.