Artificial Intelligence: Citi’s Inference Ahead - Constraints v. Capabilities

Research evidence excerpt

Artificial Intelligence: Citi’s Inference Ahead - Constraints v. Capabilities

Viewpoint |

24 Jul 2026 15:51:01 ET │ 19 pages

Artificial Intelligence

Citi’s Inference Ahead - Constraints v. Capabilities

CITI'S TAKE

Heath Terry AC

The realized risk from autonomous agents escaping sandboxes has moved +1-212-723-4624

from interrupting lunch in April to infiltrating Hugging Face infrastructure in heath.terry@citi.com

July. This creates new fault lines around security and regulation while open-

weight models proliferate and grow more performant (Poolside, 7/21; Figure Shelby Spencer

2). Even as a swift resolution to the debate (Nvidia, 7/24) on AI regulation +1-212-816-0416

(Scott Bessent, 7/22) remains unlikely, the landscape raises the compliance shelby.spencer@citi.com

bar for enterprises that maintain their own abstraction layers for running

models. Meanwhile, the underlying interpretability gap persists, with the Ashley Kim

providers training these models still unable to fully inspect why they act. As +1-212-816-6689

systems grow more autonomous and the economic tradeoff between some ashley.kim@citi.com

agents and people shows potential to narrow (METR, 7/21), token spend will

continue to accelerate, driving the increasing demand for infrastructure Janna Withrow

reflected in our updated AI Industry model. +1-212-723-0439

janna.withrow@citi.com

Power. Kimi's release of the largest open model to date underscores how

dramatically models are scaling in both size and user base. While these systems

scale, a growing share of wall-clock time goes to moving weights and KV-cache data

across HBM and the inter-GPU fabric rather than to the matrix math itself. As a

result, trillion-parameter systems are increasingly gated by memory, interconnect,