AI in Manufacturing: Visual Inspection, Predictive Maintenance, Digital Twins, and Vendors

Three manufacturing AI use cases have earned production status: visual inspection on the line, predictive maintenance on rotating equipment, and digital twins of assets and processes. The most important 2026 signal for buyers is negative: the hyperscalers have retired most of their turnkey industrial AI services, which moves the build-versus-buy line and makes vendor viability a first-order selection criterion.

By the numbers

The 2026 reality: the hyperscalers retreated from turnkey industrial AI

Five years ago the obvious advice for a plant starting its AI program was to reach for a managed cloud service: upload defect images to a vision service, stick wireless vibration sensors on motors, and let the cloud provider own the models. That advice is now obsolete, because the services themselves are being switched off. Amazon Lookout for Vision — AWS's service for building custom quality-inspection vision models — stopped accepting new customers on October 10, 2024 and was discontinued on October 31, 2025, with AWS pointing existing users toward Amazon SageMaker AI, Amazon Bedrock, or partner solutions.^[4] Amazon Lookout for Equipment, which trained anomaly-detection models on up to 300 equipment sensors, loses all console and resource access on October 7, 2026.^[3] Amazon Monitron, the end-to-end sensor-plus-service condition-monitoring product, is closed to new customers.^[5]

The retreat is not an AWS quirk. Microsoft stopped new resource creation for Azure AI Anomaly Detector — whose multivariate mode was explicitly marketed for predictive maintenance on factory machines — in September 2023, and retires the service on October 1, 2026, recommending migration to anomaly detection in Microsoft Fabric or to its open-source anomaly-detector project.^[6] Google Cloud's Vertex AI Vision platform is deprecated as of June 15, 2026 and reaches end of life on September 30, 2026^[7].

Read the pattern, not the individual obituaries. The hyperscalers are not abandoning manufacturing; they are abandoning the thin, prescriptive application layer and consolidating on general-purpose platforms — SageMaker, Bedrock, Fabric, Vertex AI — where foundation models and custom training live. The application layer they vacated is being claimed by industrial specialists that own sensors, domain models, and workflow integration. For a buyer, this redraws the map: a hyperscaler logo on an industrial AI product is no longer a durability signal by itself, and every shortlist now needs an explicit answer to the question of what happens to your models, data, and sensor fleet if the product is sunset.

Use case 1: Visual inspection and defect detection

Automated visual inspection is the most mature of the three use cases: cameras image every unit on the line, a trained model classifies or localizes defects, and results feed the quality system in real time. The system has four layers that should be evaluated separately — imaging (cameras, optics, lighting), inference (edge devices at the line versus cloud endpoints), the model itself, and integration into the manufacturing execution system so a detection becomes a reject, an alert, or a root-cause record. Latency budgets are set by line speed, which is why inference usually runs at the edge; the cloud's role is training, fleet management, and analytics.

The modeling substrate is shifting under this use case. The classical approach — a convolutional network trained on a few thousand labeled defect images — still delivers the best unit economics for a stable, high-volume line. What has changed is the arrival of large multimodal models that can reason about images with little or no task-specific training, which AWS now explicitly positions as a defect-detection path via Bedrock's foundation models for customers leaving Lookout for Vision.^[4] Multimodal models are slower and costlier per image, but they handle open-ended questions ("describe the anomaly"), rare defect classes with no training examples, and natural-language triage of inspection results. A pragmatic architecture uses a small specialized model for the 99 percent of frames that are routine and escalates ambiguous frames to a multimodal model for richer analysis; the architectural tradeoffs of that pattern are covered in depth at /guides/enterprise-multimodal-ai-guide.

Fusion beyond pixels matters more in manufacturing than in most domains. Inspection accuracy improves when image data is combined with synchronized process signals — vibration, temperature, torque — and with textual context such as operator annotations and maintenance logs, because many visual anomalies are only interpretable against process state. That multimodal fusion requirement is also what makes data engineering the dominant cost: aligning camera frames with sensor timestamps and labels is harder than training the model. Plan for labeling as an ongoing operation, not a project phase — product changeovers cause model drift, and the teams that sustain accuracy are the ones that route operator-verified detections back into retraining.

Vendor-wise, this segment is where the hyperscaler retreat bites hardest, because Lookout for Vision and Visual Inspection AI were the turnkey options. The choices that remain: machine-vision incumbents (Cognex, Keyence, Hexagon) that bundle cameras, lighting, and deep-learning tooling; software specialists such as Landing AI and Instrumental that focus on the model lifecycle across heterogeneous lines; and building in-house on general ML platforms — the path AWS itself now recommends, using SageMaker for custom models or Bedrock for foundation-model inference.^[4] For regulated products — medical devices, aerospace — explainability and audit trails weigh heavily in that choice; the cross-industry regulatory picture is compared at /compare/ai-regulated-industries-comparison.

Use case 2: Predictive maintenance with vibration analysis and anomaly detection

Predictive maintenance is the largest digital-industrial AI application by spend — 39.9 percent of digital-twin implementation expenditures, per the market data compiled by NIST.^[2] The physics is well understood: rotating equipment (motors, pumps, compressors, gearboxes) telegraphs developing faults through its vibration signature. Accelerometers capture the oscillation, signal processing (filtering, Fourier transforms, features such as RMS energy and kurtosis) turns raw waveforms into features, and models watch for deviation from the machine's healthy baseline. Amazon's now-closed Monitron product is a useful reference architecture even in retirement: purpose-built vibration and temperature sensors, a gateway, and analytics that combined ML models with ISO 20816 vibration-severity thresholds — with up to 20 sensors per asset.^[5]

The modeling reality is that labeled failure data is scarce — most plants cannot produce a training set of documented bearing failures — so the field leans on anomaly detection: learn normal behavior, flag deviation. The research base here is solid and worth citing when a vendor claims proprietary magic. Graph attention networks that learn cross-sensor dependencies in multivariate time series underpinned Azure's multivariate detector, which correlated up to 300 signals per model;^[8][6] on the supervised end, published CNN approaches to bearing fault diagnosis report accuracy gains of up to 15.8 percent over prior methods under time-varying speed conditions while running in real time.^[9] The methods are public; the differentiation is in sensor quality, deployment engineering, and failure-mode libraries.

With Lookout for Equipment, Monitron, and Anomaly Detector all retired or closed,^[3][5][6] the predictive-maintenance market splits into two buyable shapes. Integrated sensor-to-insight products from specialists — Augury, SparkCognition, Hitachi's industrial units, and the condition-monitoring arms of the automation incumbents — own the full chain from accelerometer to work order, and are the fastest route to value for a maintenance organization without data scientists. The platform path builds on general time-series tooling — Microsoft's stated migration target is anomaly detection in Fabric, or its open-source anomaly-detector library^[6] — and suits enterprises that already run a data platform and want maintenance analytics unified with the rest of operations data. The tradeoff is classic: the integrated product is faster and stickier; the platform build is slower, cheaper at scale, and yours.

Whichever shape you buy, treat deployed models as production ML, not appliances. Baselines drift as machines age, duty cycles change, and sensors degrade; alert thresholds need periodic recalibration against confirmed outcomes, and a model that pages technicians for phantom faults loses the floor's trust in weeks. The operating discipline — drift detection, alert quality review, retraining cadence — is the same one that governs any production model, detailed at /guides/model-monitoring-production-guide. Measure the program on maintenance outcomes (unplanned downtime, mean time between failures, planned-versus-reactive work ratio), and instrument those KPIs before the pilot so the baseline is defensible.

Use case 3: Digital twins for simulation and optimization

"Digital twin" is the most abused term in industrial software, so anchor on NIST's working definition: a computer model of a physical system, such as a machine or building, that has the potential for high accuracy, precision, and flexibility in modeling aspects of that system — with the more demanding definitions (NIST cites the AIAA's) requiring that the model be dynamically updated with data from its physical twin and carry predictive capability.^[2] NIST itself notes the threshold between "a model" and "a twin" is not easily defined and that the term is often applied to what is really just data tracking.^[2] That definitional slack is a sales weapon; make vendors state precisely what is synchronized, from which sensors, at what latency.

The economics are attractive but bounded. NIST estimates the potential impact of digital twins at $37.9 billion annually if fully adopted across US manufacturing, with a deliberately down-biased Monte Carlo putting the 90 percent confidence interval at $16.1 to $38.6 billion — "likely in the low tens of billions of dollars."^[2] More useful than the headline number is NIST's decision logic: a twin is cost-effective when the system being modeled is complex, sensitive to its settings, and carries a high cost consequence for running at non-optimal configuration.^[2] A simple machine with forgiving tolerances does not earn a twin; a bottleneck line where a 1 percent yield change moves millions does. That test — complexity times cost-of-being-wrong — is the portable takeaway, and it is also why the same NIST data shows predictive maintenance, business optimization, and performance monitoring dominating actual twin spending.^[2]

The platform layer has two live poles among the allowlisted infrastructure providers, plus the industrial incumbents. Azure Digital Twins is a platform-as-a-service for building twin graphs of entire environments — factories, buildings, energy networks — with entities modeled in the Digital Twins Definition Language (DTDL), live data flowing in from IoT Hub devices, and a query API over the running graph.^[10] NVIDIA Omniverse approaches from the simulation and visualization side: a collection of accelerated libraries and microservices for physical AI simulation, built on the OpenUSD framework for 3D interoperability, with reference workflows for industrial digital twins and factory-scale simulation.^[11] Around them sit the engineering incumbents — Siemens, Dassault Systèmes, PTC, Rockwell, GE's industrial software lines — whose twins are anchored in CAD, PLM, and automation estates. The practical pattern is layered: an IoT/graph layer for live state, a physics or simulation layer for what-if, and AI models that learn from both; the durable choice is less which logo than which interoperability standards (OpenUSD, DTDL, OPC UA) keep your asset models portable.

The vendor landscape, segmented for a buyer

Discard alphabetical vendor lists; the market sorts into four shapes with different risk profiles. Buy by shape first, then shortlist within it.

Two selection rules follow from the 2026 retirements. First, weight roadmap durability as heavily as capability: ask every vendor — hyperscaler or startup — for its end-of-support policy in writing, because the past two years prove that even category-defining products from trillion-dollar vendors get switched off.^[4][6] Second, prefer stacks whose artifacts survive the vendor: models exportable in standard formats, sensor data in your own historian or lakehouse, asset models in DTDL or OpenUSD rather than proprietary schemas.^[10][11] Lock-in you choose knowingly for speed is a legitimate trade; lock-in you discover during a forced migration is not.

Honest objections

The adoption story is thinner than the vendor decks imply. Census Bureau Business Trends and Outlook Survey data show overall AI use among US businesses ran between 17 and 20 percent from mid-December 2025 through May 3, 2026 — 19.8 percent nationally at the endpoint — and the sectors leading adoption are Information (39.7 percent) and Finance and Insurance (33.9 percent), not manufacturing; even among large firms with 250-plus employees, the rate is 37 percent.^[12] Manufacturing's Q2 2026 productivity gain of 1.9 percent is real, but it is a whole-sector number that no one should attribute to AI adoption specifically.^[1] If you are early, you are not behind a stampede; you have time to sequence deployments properly.

The ROI evidence base is also softer than it looks. NIST is explicit that its digital-twin impact estimates carry "a wide range of error" and rest on assumptions that are "not certainties,"^[2] and most vendor case-study percentages — downtime cut by X, scrap reduced by Y — are unaudited marketing claims that this guide deliberately omits. The honest planning posture is to treat published benefit figures as hypotheses, instrument your own baseline before the pilot, and fund the unglamorous prerequisites — sensor installation, data labeling, timestamp alignment, MES integration — that typically consume most of the budget. And the counterargument to urgency deserves a fair hearing: a plant with weak data infrastructure will get a better return from historian consolidation and basic condition monitoring than from a premature digital-twin program. The NIST cost-effectiveness test cuts both ways — many assets genuinely do not earn a twin.^[2]

The read

For a manufacturing CIO or platform lead, the 2026 decision stack looks like this. Visual inspection: buy integrated vision from the machine-vision incumbents for standardized high-volume lines; build on general ML platforms — with a multimodal escalation tier — where product mix changes fast, and treat labeling operations as the core competency. Predictive maintenance: buy sensor-to-insight from a specialist to get moving, but land the data in your own platform from day one so the Fabric-style build path stays open; anchor severity logic in open standards, not vendor thresholds.^[5] Digital twins: apply NIST's complexity-times-cost-consequence test asset by asset, start where the twin's spend already concentrates — predictive maintenance and optimization^[2] — and standardize on DTDL, OpenUSD, and OPC UA so the twin outlives any single vendor.^[10][11] Across all three: after watching five hyperscaler industrial services retire in two years, make exit engineering a signing condition, not an afterthought.