AI Factories (AIF)

• AI Factories (AIF) are integrated AI-enabled production systems that combine supercomputing, cloud-native services, and advanced manufacturing processes.
• They support Sovereign AI by co-locating HPC and cloud infrastructures, ensuring secure data management and efficient multi-tenant resource utilization.
• AIF extends to dynamic factory models and embodied intelligence, reshaping production topology through lifecycle management and real-time decision-making.

AI Factories (AIF) is a polysemous term in contemporary research. It denotes, in one major usage, a national or regional AI infrastructure platform that co-locates and tightly integrates supercomputing resources with cloud-native services for “Sovereign AI” (Lopez et al., 16 Sep 2025). In a second usage, it denotes a manufacturing environment or reference architecture in which AI assets are systematically developed, validated, deployed, monitored, and retired under CPPS-specific governance and MLOps-derived lifecycle control (Rauh et al., 15 Sep 2025). In a third, more recent usage, it denotes distributed, demand-proximal “AI Factories” that emerge when embodied intelligence crosses critical thresholds in dexterity, generalization, reliability, and tactile-vision fusion, thereby restructuring manufacturing economic geography itself (Fang et al., 3 Mar 2026). Related literature extends the concept to smart-factory decision support, self-driving laboratories, in-orbit production, and even automated influence-content production, making AIF a cross-domain label for end-to-end AI-enabled production systems rather than a single standardized institutional form.

1. Semantic scope and conceptual lineage

The manufacturing-centered literature roots AIF in “Industrial AI,” defined as “a systematic discipline, which focuses on developing, validating and deploying various machine-learning algorithms for industrial applications with sustainable performance” (Lee et al., 2019). In that framing, an AI Factory is not merely “a collection of sensors plus a cloud service,” but a self-organizing, closed-loop production environment in which AI modules sit alongside PLCs, robotics and MES systems, with AI algorithms treated as first-class production assets subject to version control, performance SLAs, and automated re-training pipelines.

The infrastructure-centered literature uses the same label for public or sovereign digital infrastructure. There, an AIF is “a national or regional AI infrastructure platform that co-locates and tightly integrates supercomputing (HPC) resources with cloud-native services,” with goals that include training large deep models, serving inference and agentic AI applications, securing data sovereignty, and bridging HPC performance with cloud usability (Lopez et al., 16 Sep 2025). This usage emphasizes technological autonomy, multi-tenancy, open access, and public-facing AI services.

A third research strand places AIF within economic geography. That work argues that the topology of manufacturing has remained structurally Fordist—centralized mega-factories located near labor pools—despite later innovations, and that embodied intelligence can trigger a paradigm-level transformation once capabilities cross critical thresholds (Fang et al., 3 Mar 2026). In this sense, “AI Factories” are not simply better automated plants; they are a new manufacturing topology.

Taken together, these usages show that AIF refers less to one artifact than to an architectural principle: AI is embedded as an organizing substrate of production, orchestration, and service delivery. A plausible implication is that the term’s analytical value depends on specifying whether the object under discussion is a compute platform, a factory-operating model, a spatial manufacturing form, or a broader autonomous production system.

2. Sovereign AI infrastructure and the Cloud–HPC convergence

In the infrastructure literature, the defining architectural proposal is an “Integrated Dual-Stack” that places an HPC stack and a cloud-native stack inside the same AIF (Lopez et al., 16 Sep 2025). The HPC stack comprises SLURM or an equivalent batch scheduler, a parallel file system such as GPFS or Lustre, and direct-to-InfiniBand/GPU programming through MPI and CUDA. The cloud-native stack comprises Kubernetes clusters, S3-compatible object storage, container runtimes such as Docker and Singularity, and managed PaaS/FaaS services. Between them sits a shared resource plane: a single pool of GPUs and CPUs, a unified data layer combining PFS and object storage behind a common data catalog, and performance-portable containers with acceleration-as-a-service middleware. Above that lies an AI service layer with JupyterHub, LLM inference endpoints using OpenAI-compatible APIs and vLLM, Model Context Protocol servers, agent orchestrators, and ETL, graph, streaming, and vector-database pipelines. An external federation layer provides secure endpoints and peering with sovereign public clouds, edge clusters, and other AIF “antennas.”

This architecture is motivated by two reciprocal problem classes. “Serverless HPC” asks how to bring cloud elasticity to supercomputing: S3-style object storage alongside GPFS, multi-tenant GPU-backed Kubernetes on InfiniBand, and function-level inference or ETL with sub-second startup rather than SLURM-scale minutes. “High-performance Cloud” asks how to raise cloud infrastructure to HPC levels: object-store protocols with reduced metadata overhead and direct PCIe/NVMe paths, FaaS on bare metal or disaggregated servers, and Kubernetes networking adapted to RDMA/InfiniBand so that distributed training can use MPI-style collectives inside containers.

The performance discussion is correspondingly hybrid. The cited model for point-to-point communication is the classical Hockney relation,

$T(m) \simeq \alpha + \beta \cdot m,$

with $\alpha$ as software/hardware latency and $\beta$ as per-byte transfer time. The paper also references serverless message passing benchmarks with “up to 162× communication speedup and 397× cost reduction over storage-backed FaaS,” burst-compute startup times contrasting “HPC batch ($T_{\text{batch}} \approx \text{minutes}$) vs. FaaS ($T_{\text{FaaS}} \approx \text{milliseconds}$),” and standard efficiency measures

$$\eta = \frac{\text{Useful Compute\_Time}}{\text{Allocated\_Cores} \times \text{Wallclock\_Time}}, \qquad U = \frac{\sum (\text{Active\_GPU\_cycles})}{\sum (\text{Provisioned\_GPU\_cycles})}.$$

The operational target is $\eta \to 1$ under mixed workloads by dynamically sharing GPUs across stacks.

The EuroHPC case studies frame AIF as an institutional program as well as a technical design. The European Commission is described as having committed “€ 1.5 billion across 13 AI Factories” with plans for “€ 20 billion in 5 AI GigaFactories.” The comparison among “Conservative Front-end,” “Isolated Dual-Stack,” and “Integrated Dual-Stack” approaches is instructive: OpenOnDemand on top of SLURM showed low uptake by non-HPC users; isolated “AI Factory islands” avoided contention but underutilized GPUs and raised security complexity; and an integrated dual-stack pilot in Spain hosted at BSC, by unifying SLURM and Kubernetes GPU scheduling, yielded “20 % higher average GPU utilization” and “cut job queuing times in half.” Security, usability, and accessibility are treated as first-order concerns through unified IAM, Kubernetes Network Policies and SLURM cgroups, hardened public API gateways, visual resource sliders, pre-baked experiment templates, low-code pipeline builders, hosted data catalogs, and notebook-to-FaaS export.

3. Manufacturing AIF as lifecycle, architecture, and governance

In manufacturing research, AIF is formulated as both a layered technical architecture and a governed process model. One architecture organizes the factory in four layers: a Perception Layer that collects raw status and environment data through heterogeneous sensors, RFID, and machine-vision cameras; a Network/Communication Layer using Software-Defined Industrial Network components, 5G/Wi-Fi/LTE gateways, and TSN/EtherCAT for required QoS, latency and reliability; a Computation Layer partitioning time-sensitive inference to edge servers and large-scale analytics to cloud clusters running Hadoop/Spark and Kubernetes; and a Service/Application Layer hosting predictive maintenance, order management, digital twin services, KPI dashboards, recommender systems, RESTful APIs, MES/ERP interfaces, and mobile applications (Wan et al., 2021). The same work identifies four characteristic capabilities of the AI-driven customized factory: self-perception, operations optimization, dynamic reconfiguration, and intelligent decision-making.

AIM4M recasts AIF as a lifecycle-management blueprint for AI assets in CPPS (Rauh et al., 15 Sep 2025). It partitions the AI-asset lifecycle into four sequential phases and eleven stages, with a governance gate after each phase and “AI Cards” for Use Case, Data Set, Model, and Deployment. Phase 1, Ideation, covers problem analysis, conceptual design or PoC, and approval for development. Phase 2, Development, covers data and model development, application development, operation onboarding, and approval for release. Phase 3, Operation, covers operation and monitoring, improvement and optimization, and approval for update. Phase 4, Retirement, covers end-of-life decommissioning, archival or erasure, final performance comparison, and lessons learned. The formal stage set is

$$S=\{\mathrm{Ideation},\,\mathrm{Development},\,\mathrm{Operation},\,\mathrm{Retirement}\},$$

with a directed transition function $\delta:S\times S\to\{0,1\}$ and total process cost expressed as

$C_{\mathrm{total}}=\sum_{s\in S} c(s).$

Three refinements distinguish this process model from generic MLOps. First, an extended OT/IT role model introduces Hardware Engineer, Infrastructure Engineer, Service Engineer, and Quality Inspector. Second, CPPS-specific activities are embedded explicitly: lab and shop-floor hybrid testing, OT/IT interface definition, sensor and IIoT device provisioning, low-latency inference design, and controlled rollouts such as canary and shadow modes with rollback procedures tailored to physical processes. Third, compliance is operationalized through structured artefacts and three quality gates that enforce segregation of duties, sign-off, and audit trails aligned with the EU AI Act and Industry 4.0 standards.

The broader Industrial AI literature provides the deployment logic around this process view (Lee et al., 2019). It maps AI adoption onto the 5C-CPPS architecture—Connection, Conversion, Cyber, Cognition, Configuration—and emphasizes data capture, model building, deployment pipelines, and continuous feedback loops. This includes instrumenting assets, pre-processing signals, aggregating data into machine or line digital twins, containerized micro-services on Docker/Kubernetes, CI/CD-style workflows with A/B testing, and audit trails of inputs, outputs, and decisions.

Empirical case studies show that these architectural and governance structures are tied to measurable production changes. In a customized packaging prototype with “5 robots, 2 AGVs, conveyors, RFID tags, OPC UA/DDS network, edge+cloud compute,” AI scheduling plus dynamic reconfiguration increased throughput “from 100 units/hr to 130 units/hr,” improved the Production Flexibility Index “from 0.72 to 0.88,” and reduced unplanned downtime by “45%, saving ~\$25 K/month” (Wan et al., 2021). In the Lighthouse-factory discussion, further examples include Foxconn’s Shenzhen plant, BMW’s Regensburg line, Tata Steel, and Bosch Automotive, with reported improvements in labor, yield, OEE, downtime, scrap, and maintenance cost (Lee et al., 2019). A common misconception is that AI factories are defined by model deployment alone; the literature instead treats traceability, OT/IT integration, staged approval, and continuous feedback as constitutive elements.

4. Embodied intelligence and phase transitions in economic geography

The embodied-intelligence literature formalizes an AIF transition in terms of a four-dimensional “Capability Space”

$\alpha$0

where $\alpha$1 is Dexterity, $\alpha$2 is Generalization, $\alpha$3 is Reliability, and $\alpha$4 is Tactile-Vision Fusion (Fang et al., 3 Mar 2026). Dexterity is measured by a normalized manipulation score on benchmark object sets; generalization is the out-of-distribution success rate on unseen tasks; reliability is mean time between failures normalized to $\alpha$5; and tactile-vision fusion is a scalar index of multimodal sensor fidelity and latency. The claim is that these four axes suffice because fine motor manufacturing requires dexterous control and robust perception, micro-factory viability depends on generalization across product variants, and economic geography hinges on machine uptime in place of human reliability.

Site selection is then written as a per-unit delivered-cost objective,

$\alpha$6

where $\alpha$7 is local wage or its residual, $\alpha$8 is transport cost to major demand centres, $\alpha$9 is a convex machine-climate penalty, and $\beta$0 capture fixed setup and scale economies. As $\beta$1 crosses critical surfaces such as $\beta$2, $\beta$3, and $\beta$4, the topology of the level sets $\beta$5 changes. Below $\beta$6, low-$\beta$7 robots cluster in a few low-wage, low-transport-cost factory nodes; once $\beta$8 passes $\beta$9, those level sets break up into many small, disconnected islands around local demand hubs.

Three specific pathways define the phase transition. In weight inversion, when dexterity and generalization satisfy

$T_{\text{batch}} \approx \text{minutes}$0

the robot’s per-unit handling cost falls below its per-unit transport cost, and factories move from large central plants to many small proximal sites. In batch collapse, once

$T_{\text{batch}} \approx \text{minutes}$1

the break-even batch size $T_{\text{batch}} \approx \text{minutes}$2, eliminating scale-economy lock-in and enabling “single-piece flow” near demand. In human-infrastructure decoupling, when

$T_{\text{batch}} \approx \text{minutes}$3

the system no longer requires high-density human-support infrastructure, and siting is freed from legacy labor markets.

The climatic term is elevated to a first-class determinant through “Machine Climate Advantage,” defined as

$T_{\text{batch}} \approx \text{minutes}$4

Here $T_{\text{batch}} \approx \text{minutes}$5, $T_{\text{batch}} \approx \text{minutes}$6, and $T_{\text{batch}} \approx \text{minutes}$7 are local humidity, irradiance, and thermal stability; $T_{\text{batch}} \approx \text{minutes}$8 is the machine-optimal triple; and $T_{\text{batch}} \approx \text{minutes}$9 penalizes temporal variance. In the paper’s formulation, this replaces the standard labor–land–climate trade-off of Weber–Fujita models with geography tilted toward stable, robot-friendly climates.

The reported empirical illustrations are explicitly topological rather than merely cost-based. Before $T_{\text{FaaS}} \approx \text{milliseconds}$0, “25% of U.S. counties had zero local production capacity (‘deserts’),” and after surpassing $T_{\text{FaaS}} \approx \text{milliseconds}$1 the predicted share falls to “2%.” A concentration measure in “Country X” rises “from 0.65 to 0.82 pre-threshold, then collapses to 0.30 once $T_{\text{FaaS}} \approx \text{milliseconds}$2.” In electronics, median plant size falls “from 150,000 m² to 1,200 m²,” the number of plants increases by “8×,” and average distance to final consumer shrinks “from 1,200 km to 180 km.” These results define AIF as a new manufacturing topology produced by embodied capability thresholds, not simply an efficiency upgrade to existing factories.

5. Planning, maintenance, and retrieval systems inside smart factories

Several works instantiate AIF as an operational software stack embedded within factory workflows. Yantra AI is an intelligent production system in which sensors and PLCs stream temperature, vibration, pressure, power-load, and AGV-load data into an ingestion layer based on Pandas and NumPy, after which three ML modules run in parallel: a Random Forest Classifier for predictive maintenance, an Isolation Forest for anomaly detection, and a Random Forest Regressor for energy management (Krishnamurthy, 12 Dec 2025). Streamlit and Plotly provide a dashboard with machine-health risk scores, anomaly notifications, energy views, and “what-if” sliders, while a GPT-4 virtual assistant routes natural-language queries to the ML modules, time-series store, and process-optimization engine. The report states “machine-failure recall > 90% in synthetic runs,” “false positives kept ≲ 5% of normal windows,” “Energy forecast RMSE on hold-out ≲ 8%,” “Streamlit callbacks” redrawing figures in “< 1 s latency,” and end-to-end dashboard plus chat latency “< 1.5 s.” Because the system was tested with synthetic data, these figures characterize a designed AIF prototype rather than a validated production deployment.

FactoryLLM realizes AIF as a safe and open-source AI playground for evaluating LLM-based RAG models on cross-machine maintenance reasoning (Pulse et al., 12 Jun 2026). Its six-layer pipeline comprises configuration, knowledge ingestion, interaction, retrieval, reasoning, and storage. It supports vector RAG through ChromaDB, graph RAG through NebulaGraph, prompting strategies including IO, CoT, ToT, and GoT, and dual evaluation through RAGAS and NVIDIA’s LLM-as-a-Judge metrics. In the case study, “approximately 600 pages of cross-machine documentation” are segmented into “≈3 000 chunks,” and “30 cross-machine maintenance queries” are evaluated across Qwen3-235B-A22B-Instruct-2507, Llama 4 Maverick, and Gemma 3 27B. Reported averages include Context Precision of 0.48, 0.46, and 0.51; Context Recall of 0.82, 0.76, and 0.89; Faithfulness of 0.72, 0.62, and 0.70; NVIDIA Context Relevance of 0.90, 0.86, and 0.89; and NVIDIA Groundedness of 0.95, 0.88, and 0.92. The paper’s own interpretation is that “retrieval is the bottleneck” while “generation is robust.”

SMarTplan represents a different operational layer: online task planning and rescheduling for the smart factory (Bit-Monnot et al., 2018). It combines a special-purpose component, RCLLPlan, which generates domain-specific SMT/OMT encodings in flat and macro-action variants, with a general-purpose Lifted Constraint Planner that compiles actions into chronicles and solves a CSP via SMT. The formal planning relation is

$T_{\text{FaaS}} \approx \text{milliseconds}$3

with optimization over

$T_{\text{FaaS}} \approx \text{milliseconds}$4

On 120 RoboCup Logistics League instances, the macro-action version “solves all 120 problems,” often in “sub-second times for C0” and “under 1 s even in the more complex C1 feasibility runs.” The configured general-purpose version solves most C1 cases within “~10 s,” while OPTIC and SMTPlan fail beyond the simplest settings. These results position AIF planning as a hybrid of domain specialization for real-time responsiveness and domain independence for adaptability.

6. Science factories and in-orbit factories

Outside terrestrial manufacturing proper, the AIF concept has been generalized to modular self-driving laboratories. In the science-factory architecture, the atomic unit is the module, defined by a set of actions, a uniform six-function interface $T_{\text{FaaS}} \approx \text{milliseconds}$5, a physical device, and a network address (Vescovi et al., 2023). Modules are composed into workcells represented as tuples $T_{\text{FaaS}} \approx \text{milliseconds}$6 or equivalently as bipartite graphs linking modules and stations. Multiple workcells, together with mobile-robot links and online compute/data services, form a science factory $T_{\text{FaaS}} \approx \text{milliseconds}$7. Workflows are finite step sequences over modules; applications orchestrate workflows, AI inference, model training, simulation, and data publication. The architecture integrates laboratory hardware with distributed compute through Globus Compute and Flows, while digital twins in Omniverse or Isaac Sim validate motions and generate synthetic data.

The prototype reported in that work spans “15 different robotic apparatus, five applications (one in education, two in biology, two in materials), and a variety of workflows, across four laboratories” (Vescovi et al., 2023). In the Color Picker application, loop time “varies from 10–60 min depending on batch size and compute location.” In Electrochromic Polymers, the metric reported is “Number of new candidate polymers screened per day (~100).” The concurrency discussion goes further and states that no fundamental bottleneck prevents deployment of “$T_{\text{FaaS}} \approx \text{milliseconds}$8 workcells,” each submitting “$T_{\text{FaaS}} \approx \text{milliseconds}$9 experiments per day,” for aggregate “$$\eta = \frac{\text{Useful Compute\_Time}}{\text{Allocated\_Cores} \times \text{Wallclock\_Time}}, \qquad U = \frac{\sum (\text{Active\_GPU\_cycles})}{\sum (\text{Provisioned\_GPU\_cycles})}.$$0 experiments per day,” sufficient to support “$$\eta = \frac{\text{Useful Compute\_Time}}{\text{Allocated\_Cores} \times \text{Wallclock\_Time}}, \qquad U = \frac{\sum (\text{Active\_GPU\_cycles})}{\sum (\text{Provisioned\_GPU\_cycles})}.$$1 active scientists.” This suggests that, in this literature, AIF denotes not only automated production but also scalable experimental discovery infrastructure.

The in-orbit factory literature adapts the same logic to robotic assembly of small satellites (Leutert et al., 2024). The reported system centers on a dual-arm YUMI robot with “7 DoF per arm,” integrated six-axis force/torque sensors, an optical inspection station with “overview camera + two microscopes under polarized illumination,” an electrical test bench, a real-time compositor bus, and a central Digital Process Twin. The physical process is modeled as $$\eta = \frac{\text{Useful Compute\_Time}}{\text{Allocated\_Cores} \times \text{Wallclock\_Time}}, \qquad U = \frac{\sum (\text{Active\_GPU\_cycles})}{\sum (\text{Provisioned\_GPU\_cycles})}.$$2 with observation model $$\eta = \frac{\text{Useful Compute\_Time}}{\text{Allocated\_Cores} \times \text{Wallclock\_Time}}, \qquad U = \frac{\sum (\text{Active\_GPU\_cycles})}{\sum (\text{Provisioned\_GPU\_cycles})}.$$3 and a digital-twin update $$\eta = \frac{\text{Useful Compute\_Time}}{\text{Allocated\_Cores} \times \text{Wallclock\_Time}}, \qquad U = \frac{\sum (\text{Active\_GPU\_cycles})}{\sum (\text{Provisioned\_GPU\_cycles})}.$$4. Optical fault detection uses a modified UPerNet semantic-segmentation architecture with ConvNeXt blocks over five classes; electrical-fault detection uses Local Outlier Factor; insertion uncertainty is handled through force sensing, Q-learning over offset grids, and an impedance-like insertion law; and teleoperation uses shared control

$$\eta = \frac{\text{Useful Compute\_Time}}{\text{Allocated\_Cores} \times \text{Wallclock\_Time}}, \qquad U = \frac{\sum (\text{Active\_GPU\_cycles})}{\sum (\text{Provisioned\_GPU\_cycles})}.$$5

The reported metrics are unusually detailed. Optical inspection achieves “Mean Precision > 0.9 and Recall > 0.9” for “components” and “tombstone,” while “solderball” reaches approximately $$\eta = \frac{\text{Useful Compute\_Time}}{\text{Allocated\_Cores} \times \text{Wallclock\_Time}}, \qquad U = \frac{\sum (\text{Active\_GPU\_cycles})}{\sum (\text{Provisioned\_GPU\_cycles})}.$$6. Electrical outlier detection reports “true positive rate > 95%, false positive rate < 5%.” For connector insertion, nominal insertion succeeds in “≈85% of trials” without fallback, whereas “with Q-learning fallback the system attained 100% insertion success over 200 consecutive assemblies,” with learned offsets within $$\eta = \frac{\text{Useful Compute\_Time}}{\text{Allocated\_Cores} \times \text{Wallclock\_Time}}, \qquad U = \frac{\sum (\text{Active\_GPU\_cycles})}{\sum (\text{Provisioned\_GPU\_cycles})}.$$7 mm and peak contact forces reduced by “30%.” Shared control reduces mean completion time “from 60 s (pure teleop) to 35 s,” raises success rate “from 88% to 99% under 500 ms round-trip latency,” and eliminates “>90% of operator-induced off-axis collisions.” In this setting, the factory is cyber-physical, AI-enhanced, and fault-tolerant, but also explicitly teleoperated and twin-supervised.

7. Giga-scale networks and contested extensions of the term

At the largest scale, AIF research includes network fabrics designed for distributed model training across “hundreds of thousands of GPUs” (Khashab et al., 20 May 2026). The cited example is the Spectrum-X multiplane architecture, in which each NIC breaks an 800 Gbps port into lower-rate planes such as “4×200 Gbps,” each plane connecting into an independent two-level leaf–spine fat-tree. The design principle is “topological parallelism vs. hierarchical depth”: rather than adding spine tiers and increasing hop count and tail jitter, the network adds parallel planes while keeping per-plane diameter at two hops. Load balancing is decomposed into in-switch adaptive routing, NIC-based per-plane load balancing, and sender-based congestion control reacting to ECN marks. The paper states reaction times of $$\eta = \frac{\text{Useful Compute\_Time}}{\text{Allocated\_Cores} \times \text{Wallclock\_Time}}, \qquad U = \frac{\sum (\text{Active\_GPU\_cycles})}{\sum (\text{Provisioned\_GPU\_cycles})}.$$8 for switch-side exclusion and $$\eta = \frac{\text{Useful Compute\_Time}}{\text{Allocated\_Cores} \times \text{Wallclock\_Time}}, \qquad U = \frac{\sum (\text{Active\_GPU\_cycles})}{\sum (\text{Provisioned\_GPU\_cycles})}.$$9 for NIC failover, with bisection bandwidth

$\eta \to 1$0

The empirical results position networking as a constitutive layer of giga-scale AIF. On a “1 K-GPU H100 testbed,” SPX first-percentile bandwidth is “377 Gbps (98 % of 384 Gbps),” with “P99 tail latency at 75 % load” of “8–9 µs (±0.2 µs).” Under cross-tenant contention, a victim All2All workload recovers “>98 % of isolated performance.” With “10 % of leaf-to-spine links disabled uniformly,” RDMA bisection bandwidth drops “11 % (377 → 335 Gbps)” and p99 latency increases by “7 %.” Under a single host-link flap, “SPX PLB failover” reaches “2.68 ms to 75 % line rate,” whereas a software load-balancer takes “≃1.08 s (>400× slower).” These figures place AIF within the design space of hardware-accelerated Ethernet fabrics, not only cluster schedulers or model-serving stacks.

The term AIF is also used outside industrial or sovereign-infrastructure contexts. In “AI Propaganda factories with LLMs,” AIF denotes a modular end-to-end system for automated political messaging, comprising persona engineering, prompt construction, message generation on small LLMs, automated evaluation by a judge model, and optional memory or policy modules (Olejnik, 27 Aug 2025). The study evaluates “180 ChangeMyView (CMV) threads × 8 personas × 2 modes × 4 models = 11,520 generated replies,” reporting “persona-over-model” and “engagement as a stressor” as its two main findings. Quantitatively, mean IAS rises “from $\eta \to 1$1 (response) to $\eta \to 1$2 (engagement),” while EIC jumps “from 42–64% to 69–85%” for far-extreme personas. The paper argues that defense should shift toward “conversation-centric detection and disruption of campaigns and coordination infrastructure.”

This non-industrial usage is not a mere terminological anomaly. It shows that AIF can denote any modular, end-to-end AI production system capable of high-throughput generation, evaluation, and deployment of outputs. At the same time, it creates a genuine ambiguity in the term. A common misconception is that “AI Factory” refers only to GPU infrastructure or only to smart manufacturing. The cited literature shows both usages, and also shows further extensions to scientific discovery and influence operations. The most precise reading therefore depends on the surrounding research program: sovereign compute, industrial CPPS operationalization, embodied manufacturing topology, autonomous laboratories, or automated content production.