⚡ electrical engineering

Modular Integrated AI Infrastructure (MiAI): from deployment modularity to lifecycle modularity

This paper introduces Modular Integrated AI Infrastructure (MiAI), a paradigm that decouples the renewal cycles of computing and support assets through co-designed physical and functional modularity, enabling selective renewal and multi-cycle reuse while demonstrating that favorable environmental outcomes are achieved when retained service credits exceed the burdens of transition.

Original authors: Yuan-Hao WEI

Published 2026-09-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yuan-Hao WEI

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Artificial intelligence is reshaping the world, but the physical machines that power it are often built with a hidden flaw in their design. These massive data centers rely on two very different types of equipment: the computing hardware that runs the software, and the supporting infrastructure like power systems, cooling units, and structural frames. The computing gear evolves at a breakneck pace, becoming obsolete every few years, while the heavy-duty support systems are built to last for decades. In traditional construction, these two timelines clash. When the computers need an upgrade, the entire building is often treated as a single unit, forcing engineers to tear out perfectly good, long-lasting support systems just to swap out the short-lived processors. This creates a cycle of waste where durable assets are discarded prematurely, consuming vast amounts of energy and materials to rebuild what could have been kept.

A new approach called Modular Integrated AI Infrastructure, or MiAI, proposes a different way to build these facilities. Instead of treating the data center as a monolithic block that must be replaced all at once, MiAI treats it as a collection of distinct, interchangeable parts. By designing the physical connections and digital interfaces from the start, engineers can isolate specific functions. This allows the fast-changing computing elements to be swapped out while leaving the slower-changing power and cooling systems in place, provided they still meet the new requirements. The goal is to extend the useful life of the infrastructure, preserving the value of the heavy assets across multiple generations of technology.

The researcher behind this concept, Yuan-Hao Wei at The Hong Kong Polytechnic University, developed a framework to test whether this strategy actually saves resources. They compared the MiAI approach against a standard modular baseline, which already keeps some support assets but does not fully optimize for selective renewal. The team created a detailed model to track the environmental cost, specifically the global warming potential, of upgrading a facility over time. They calculated the "burden" of the new work required—such as the materials needed for new connection points and the labor to disconnect and reconnect systems—and weighed it against the "credit" gained by keeping old, functional equipment in service. The study focused on a specific scenario where a facility undergoes two major technology transitions over a fifteen-year period.

The findings reveal a clear tipping point where the strategy becomes beneficial. The researchers found that for MiAI to be environmentally superior, the amount of qualified service preserved must exceed the effort required to enable the renewal. In their specific simulation, they determined that if the additional physical retention of support assets reaches a certain threshold, the system pays off. Specifically, within the group of assets exposed to change, the design must preserve an extra twenty percent of the physical components beyond what a standard modular design would keep. When this condition is met, the environmental credit from keeping the old equipment outweighs the cost of the new interfaces and the work to swap the parts. In their example, meeting this requirement resulted in a net reduction of global warming potential equivalent to seven and a half percent of the manufacturing impact of the entire support infrastructure.

However, the study also highlights that this benefit is not automatic. If the design fails to preserve enough of the existing equipment, or if the work required to disconnect and reconnect the systems is too high, the strategy can actually increase the environmental burden. The researcher identified three distinct zones of outcome: configurations that are favorable because they preserve enough service, configurations that are burden-dominated because the renewal work is too costly, and configurations that are physically impossible because they demand more retention than the existing hardware allows. The key to success lies in the design of the interfaces themselves. By creating standardized, reversible connections that allow for easy inspection and adaptation, engineers can lower the cost of renewal and make it easier to keep assets in service.

This work moves the conversation beyond simply how fast a data center can be built or how efficiently it runs while new. It introduces the concept of a "renewal boundary," a specific design line that determines which parts of a facility can be changed and which must stay. In a MiAI system, this boundary is not fixed to the original delivery of the building. It can shift depending on the technology transition. For instance, if a facility switches from air cooling to liquid cooling, the boundary might expand to include the new piping and pumps, but it can still exclude the structural frame or the main power transformers if they remain compatible. This flexibility allows the infrastructure to evolve without being scrapped. The study suggests that by treating the ability to renew as a primary design objective, rather than an afterthought, we can build AI infrastructure that is not only powerful but also durable and resource-efficient.

The researcher validated their model by running it through various scenarios, including different schedules for when technology updates occur and different ways of accounting for the work involved. In every case, a clear boundary emerged where the benefits of retention outweighed the costs of renewal. They also checked the consistency of their findings by simulating discrete events over time, confirming that the logic holds even when the renewal process is broken down into individual steps rather than treated as a single block. The results show that the transition from deployment modularity to lifecycle modularity is not just a theoretical idea but a calculable engineering path. It offers a way to align the rapid pace of artificial intelligence development with the slower, more sustainable rhythm of physical infrastructure, ensuring that the machines of the future do not require the constant destruction of the past.

Technical Summary: Modular Integrated AI Infrastructure (MiAI)

Problem Statement
Current artificial intelligence (AI) infrastructure increasingly relies on prefabricated modular assemblies to streamline deployment. However, a fundamental mismatch exists between the "delivery modularity" used to install these systems and the heterogeneous "renewal clocks" of their constituent components. Computing hardware typically has a short service life (e.g., 6 years), while supporting infrastructure—such as power systems, cooling, structural elements, and energy storage—often possesses significantly longer lifespans (e.g., 25–30 years).

Existing modular practices, such as Modular Integrated Construction (MiC) and Multi-trade Integrated MEP (MiMEP), optimize for coordinated delivery and installation but do not inherently define boundaries for selective renewal. Consequently, when a technology transition occurs (e.g., shifting from air to liquid cooling), the entire integrated package is often treated as a single unit for assessment, potentially leading to the premature replacement of compatible, long-lived support assets. The core problem is the lack of an architectural framework that aligns physical assembly with functional integration while explicitly decoupling these from the timing of technological renewal.

Methodology
The paper introduces Modular Integrated AI Infrastructure (MiAI) as a system architecture that co-designs physical assemblies, functional interfaces, and renewal boundaries. The methodology proceeds in three stages:

Architectural Definition: MiAI is defined as a paradigm where interoperable physical-digital modules are factory-integrated, transported, and assembled via standardized, reversible connections. Crucially, it distinguishes between a delivery unit (what is manufactured and installed together) and a renewal unit (what can be independently assessed, retained, adapted, or replaced during a transition). This allows renewal boundaries to cross delivery units, isolating only the functions affected by a specific technology shift.
Analytical Framework: The author develops a differential lifecycle Global Warming Potential (GWP) model to evaluate MiAI against a "deployment-oriented modular baseline" (B1). The baseline (B1) already retains a non-zero fraction of support assets ( $r_B$ ) during transitions. MiAI is modeled as adding an additional physical retention ( $\Delta r$ ) within the transition-exposed subset.
Metric Formulation: The study defines Service-Weighted Effective Retention ( $R_{eff}$ $R_{e f f}$ ) as the core metric for lifecycle value:
$R_{eff} = q \Delta r v$
Where:
- $q$ : The fraction of total support stock exposed to the transition.
- $\Delta r$ : The additional physical retention achieved by MiAI relative to the baseline.
- $v$ : A qualification factor (0 to 1) representing the condition, compatibility, and remaining service life of the retained assets.

The model balances the "retained-service credit" ( $N R_{eff}$ ) against the "enabling burdens" (interface premium $\alpha$ and transition work $b$ ) to determine the differential GWP ( $g$ ).

Key Contributions

Conceptual Shift: The paper shifts the design focus from deployment modularity (how infrastructure is delivered) to lifecycle modularity (what infrastructure preserves across technology generations). It treats the "renewal boundary" as an explicit design object rather than a fixed consequence of delivery packaging.
The Break-Even Condition: The author derives a condition for favorable lifecycle outcomes: $R_{eff} > R^*_{eff}$ . This establishes that MiAI generates a net reduction in GWP only if the qualified service preserved exceeds the interface and transition burdens allocated per transition.
Design Space Mapping: The study maps the relationship between design variables (interface premium, transition burden, exposure fraction, and retention) to define a bounded applicability space. This space categorizes configurations into three regions:
1. Favourable (Retention-dominated): Retained service exceeds burdens.
2. Burden-dominated: Burdens exceed retained service.
3. Physically Infeasible: The required retention exceeds the available non-retained stock ( $\Delta r > 1 - r_B$ ).
Digital-Physical Integration: The paper emphasizes that effective retention requires not just physical reversibility but also digital traceability (module identity, maintenance history, condition evidence) to qualify the remaining service ( $v$ ).

Results
Using a normalized same-site configuration with a 15-year horizon and two compute transitions ( $N=2$ ), the study provides illustrative results:

Baseline Parameters: Interface premium $\alpha = 0.10$ , transition burden $b = 0.05$ , exposure fraction $q = 0.50$ , baseline retention $r_B = 0.40$ .
MiAI Performance: With an additional retention $\Delta r = 0.30$ and qualification factor $v = 0.75$ , MiAI achieves an effective retention of $R_{eff} = 0.1125$ .
Break-Even Analysis: The required effective retention threshold is calculated as $R^*_{eff} = 0.075$ . Since $0.1125 > 0.075$ , the configuration is favorable.
GWP Impact: The resulting differential GWP is $g = -0.075$ , indicating a 7.5% reduction in lifecycle emissions relative to the support-infrastructure manufacturing reference.
Thresholds: The additional physical retention threshold within the exposed subset is $\Delta r^* = 0.20$ . Configurations retaining more than this increment (relative to the baseline) are favorable.
Consistency Check: A multi-cycle consistency check across 24 schedule-and-accounting combinations (varying refresh intervals and burden allocation rules) confirmed that a feasible internal crossing exists in every case, validating the robustness of the boundary condition.

Significance and Claims
The paper claims that MiAI extends the utility of modular construction beyond the initial delivery phase. By making the renewal boundary an explicit design variable, MiAI allows infrastructure owners to anticipate how technology transitions propagate through integrated facilities. The significance lies in the ability to:

Preserve Service: Enable the retention of compatible, long-lived support assets (power, cooling, structure) across multiple generations of short-lived compute hardware.
Quantify Trade-offs: Provide a rigorous analytical mechanism to evaluate whether the investment in standardized interfaces and reversible connections is repaid by the preservation of qualified service.
Guide Design: Offer a bounded design space where engineers can adjust interface premiums, transition work, and retention strategies to ensure lifecycle outcomes remain favorable.

The author concludes that the next generation of AI infrastructure should be designed not only for the day of deployment but for every technology generation it must outlive, transforming preservation into an explicit objective of infrastructure design.

Technical Summary: Modular Integrated AI Infrastructure (MiAI)

More like this