Agentic AI

Research evidence excerpt

Agentic AI

Macquarie Equity Research

25 May 2026

Innovation Watch

Software & Services

GlobalAgentic AI

"What if" AI models embrace LTA?

Ellie Jason

Key Points Jiang Zhang

• AI is shifting from hardware- to software/model lock-in, as capacity

constraints enable multi-year commitments and improve revenue See our latest reports below:

visibility.

• Model providers are moving up the stack, increasing switching costs, as Agentic AI:

enterprises embed agent-driven workflows into core operations. Large models, tokens, and new traffic order

• Token unpredictability limits true LTAs today, but stronger upstream Entering a dramatic stage of agentic reform

lock-in should drive gradual adoption across the AI value chain. China Innovation Watch series:

DeepSeek-V4: breaking the million-token context

barrier

• From hardware lock-in to software lock-in. Hardware vendors have Seedance 2.0: taking video generation to the next

increasingly benefited from long-term agreements (LTAs), which help frontier

narrow supply–demand mismatches and improve contract visibility. Qwen: Putting AI discovery into real-world actions

For large AI models, however, a key concern remains the persistence

of this supply-demand imbalance. This is driven by surging and often Figure 1 - The higher the LTA

unpredictable usage alongside an unprecedented CAPEX buildout, fuelling maturity, the stronger the binding...

periodic “bubble” concerns. Against this backdrop, we view OpenAI’s push

into Guaranteed Capacity as a step towards a more structured “lock-

in” model. In a capacity-constrained environment, secured and reserved

access carries tangible value and can be formalised contractually. If

scaled into multi-year commitments, this approach could help align