Artificial Intelligence: Citi's Inference Ahead - Powering the Next Generation of Agents

Citi's report notes that longer agent runs enabled by top models such as Muse Wonder Full will require further developments like QS PowerBlock and Nvidia's more open Vera Rubin architecture to support networking, power, optics, energy storage and accelerated computing even as model efficiency improves. After the release of 7 proprietary and 0 open models this week, the average cost of running AA tasks on proprietary models rose 4% WoW to $3.51 per task, widening the cost-per-task premium over open models to 4.4x from 4.2x, illustrating the trade-off between intelligence, speed and cost. On financing, AI infrastructure funding is growing more expensive and shifting toward riskier capital, with AI-related issuance reaching $75B in 2026, 7% above the 2025 total;