Contemporary artificial intelligence (AI) systems predominantly operate under an execution-oriented paradigm in which computational requests are processed as soon as possible after submission. Although immediate computation is essential for latency-sensitive applications, it is not necessarily the most economically efficient strategy for every computational task. Electricity prices fluctuate, GPU rental costs vary, cloud computing resources experience changing supply and demand, and different AI requests exhibit different degrees of tolerance for delayed execution. These conditions suggest that computational efficiency should not be understood exclusively in terms of execution speed. This paper proposes Postponed Computation (PC) as a central temporal computing paradigm in which an AI system deliberately postpones eligible computational operations until a more favorable future execution opportunity becomes available. Unlike conventional post-hoc computation, which analyzes or processes information after an event has occurred, Postponed Computation concerns the intentional scheduling of computation itself for a later time. The paper develops an Economic-Aware Postponed Computation framework for generative AI services, in which the decision to compute is separated from the decision of when to compute. Four principal mechanisms are introduced: User-Adaptive Postponement, Predictive Economic Timing, Deadline-Constrained Postponement, and Cross-Task Aggregation. The framework considers electricity prices, GPU and cloud resource costs, computational availability, waiting costs, service-level commitments, and user-selected execution preferences. A particularly relevant application is an economy-oriented AI service in which eligible user requests, including voluntarily deferred requests from free-tier users, can be retained and executed during periods of lower computational expenditure. The framework further investigates whether compatible requests from multiple users can be aggregated into larger execution batches to improve resource utilization. Existing research has established important foundations in deferred workload scheduling, electricity-aware computing, cloud resource optimization, and inference batching. Accordingly, the contribution of this paper is the formalization and proposed integration of Postponed Computation as a broader temporal decision principle for generative AI services, rather than a claim that workload deferral itself has never been studied. A conceptual architecture, mathematical formulation, operational mechanisms, research hypotheses, and evaluation methodology are presented. No experimental results or empirical performance improvements are claimed.
Year: 2026
Type: preprint
Abstract: The rapid expansion of generative artificial intelligence (GenAI) has intensified the demand for computational resources, particularly high-performance graphics processing units (GPUs), electricity, memory, and large-scale data-center infrastructure. Contemporary AI services are predominantly designed around immediate or near-immediate inference, whereby computational requests are processed shortly after submission regardless of whether the current economic and infrastructural conditions represent an optimal opportunity for execution. Although existing research has investigated deferred cloud computing, deadline-aware scheduling, dynamic electricity pricing, carbon-aware workload shifting, and opportunistic utilization of low-cost computing resources, the broader possibility of treating AI computation as a long-horizon economic decision that may be postponed for months, reassessed, reorganized, and subsequently executed remains an important research direction. This paper introduces Ultra-Deferred Computation (UDC) as a proposed computational paradigm that extends conventional immediate inference and short-horizon deferred scheduling toward long-horizon, economically adaptive AI computation. The proposed framework distinguishes request submission, computational commitment, resource acquisition, execution, and result delivery as separate events that need not occur within the same operational period. It integrates future GPU availability, electricity prices, computing-resource rental costs, renewable-energy conditions, workload characteristics, computational opportunity, user service tiers, task utility, and acceptable waiting periods into a unified decision-making architecture. A particularly important application concerns generative AI services in which users of free or lower-priority service tiers may voluntarily authorize computationally intensive requests to remain in a persistent deferred queue for several weeks or months in exchange for reduced service costs or free execution. Rather than merely delaying an existing queue, UDC introduces the possibility of periodically reassessing task value, predicting future computational opportunities, reallocating execution priorities, and reorganizing compatible workloads before activation. A formal optimization model, task-lifecycle architecture, scheduling procedure, and research propositions are developed to establish the theoretical foundations of this paradigm. This study is conceptual and methodological; it does not claim experimental validation or measured cost reductions. To the best of our knowledge, this paper proposes one of the first unified conceptual frameworks specifically positioning long-horizon, user-tier-adaptive, economically conditioned deferral of generative AI computation as a distinct computational-service paradigm. The framework opens a research direction beyond immediate AI inference in which the timing of computation becomes an explicit dimension of intelligent resource allocation.
Author: Zhang, Jincheng
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7226158303
