AI in Manufacturing: Start With the Data, Then Size the Compute

Manufacturers have moved fast on the first step of AI. Parsec Automation surveyed 1,200 manufacturing leaders in February 2026 and found that 72% had adopted AI in some form, up from 53% in 2024. Scaling is harder: 10% run AI at scale across their operations. McKinsey's COO100 survey of 101 operating executives at companies above US$1 billion in revenue tells the same story from the top. About two-thirds say AI is still at the exploration or targeted-implementation stage, and 2% call it fully embedded.

The money is moving anyway. One-third of those COOs put under 1% of cost of goods sold into digital and AI over the past five years. For the next five years, 93% plan to spend more, and almost one-third intend to reach at least 5% of cost of goods sold. The question for a plant leader is where that money goes first. This guide makes the case for data first, then compute sized to each workload, and it uses GMI Cloud's services to show what that looks like in practice.

Figure 1. Adoption outpaces scale. Source: Parsec Automation 2026 survey of 1,200 manufacturing leaders.

What is AI in manufacturing today?

In practice it is four jobs that happen to share a factory. Cameras check parts for defects. Models read machine sensor data and predict failures before the line stops. Simulation and digital twins test a layout or a robot program before anyone touches equipment. And operators ask a copilot questions that used to mean digging through manuals and maintenance logs. Quality control leads the list: it has been the number one AI use case for two years running in Rockwell Automation's State of Smart Manufacturing report, with 50% of 1,560 respondents planning to apply AI to product quality. Parsec's survey puts quality control at 50% as well, followed by IT operations (46%) and supply chain management (45%).

Each of those four jobs needs data from the floor and compute somewhere to train and run the models. The two problems manufacturers hit first are on the data side.

Why manufacturing data slows AI adoption

The first problem is governance. Finance and healthcare collect, classify and retain data under rules that established basic data standards years ago. Manufacturing varies by sector and by plant. A pharmaceutical site following good manufacturing practice (GMP) or an automotive supplier following the IATF 16949 quality standard has formal data requirements; many discrete and process plants have practices that differ line by line.

An automated production line with robotic assembly equipment. AI-generated illustration.

The second problem is connecting to operational technology (OT), the equipment and systems that run the factory. Machine controllers, drives and sensors use different communication standards. Equipment records and production systems may already hold years of useful data, but getting it out safely is the first engineering task. A practical approach is to move data through a protected network buffer between factory equipment and business systems, while keeping outside systems from connecting directly to the machines. Security leads the list of barriers in the survey below, so involve the plant’s OT team early.

The surveys match the shop-floor experience. RSM's 2026 middle-market study identifies security, data quality and integration with older systems as the main barriers, as the chart below shows. Among companies whose pilots underperformed, data quality was the most frequently cited reason. These findings make data readiness the first priority. As a pilot expands to more lines, the compute requirements also need to be reassessed.

Figure 2. Barriers to AI deployment. Source: RSM US Middle Market AI Survey 2026, manufacturing respondents.

A useful first exercise is to inventory the four kinds of data a factory produces and note what each one asks of the infrastructure. Table 1 is a starting template.

Table 1. Four factory data types and their infrastructure needs. Source: GMI Cloud.

Where AI is used in manufacturing: four workloads and their GPU needs

Each workload has its own compute pattern. Table 2 lays them out side by side, with the GMI Cloud service that fits each pattern.

Table 2. Manufacturing AI workloads mapped to compute patterns. Source: GMI Cloud.

Machine vision checks parts at the production line. AI-generated illustration.

For visual inspection, one design point matters. The device at the production line makes the reject decision, while training and retraining can run on cloud GPUs. A camera model learns from a labeled image set, often starting with a pretrained model with a few hundred defect examples, gets evaluated, then ships to the line as a deployable model in ONNX format or optimized with TensorRT. When the product changes or false rejects creep up, the loop runs again. That loop is where cloud GPU time gets used. The result also has to return to the line in a form the plant understands: a signal to the equipment, a reject code in the production system, or a maintenance work order. That is what an OT lead means by integration.

The simulation row is where the term physical AI shows up. Robot policy training, synthetic data generation and digital twins are the same stack that robotics companies use, and Deloitte's 2026 outlook cites a Manufacturing Leadership Council survey in which close to a quarter of manufacturers plan to deploy physical AI within two years, more than double current adoption.

How does AI predictive maintenance work, and what does downtime cost?

A predictive maintenance model reads vibration, temperature, current and cycle data from the equipment records, learns what normal looks like for each asset, and flags changes early enough to schedule the repair. Most production models are gradient-boosted trees or small neural networks that train in minutes and run on a CPU. GPUs earn their place when a plant trains deep sequence models across hundreds of assets at once, or retrains the whole fleet on a schedule.

The money involved is large. Siemens' True Cost of Downtime 2024 report, based on 181 interviews at large industrial organizations, puts unplanned downtime in automotive at about US$2.3 million per hour, roughly double the 2019 figure, and estimates the annual cost across Fortune Global 500 industrials at US$1.4 trillion, or 11% of revenue. Facilities are having fewer incidents, 25 a month versus 42 in 2019, but each one now takes 81 minutes to recover from instead of 49. Almost half of the surveyed firms now run a predictive maintenance team, double the 2019 share.

Figure 3. Downtime incidents and recovery time per facility. Source: Siemens, The True Cost of Downtime 2024.

GPU infrastructure supports training and simulation workloads. AI-generated illustration.

On-premises or cloud GPUs: a rough cost picture

The workloads above sit in very different budget lines. Retraining a visual inspection model on tens of thousands of images takes a few hours on a single H100-class GPU, which at GMI Cloud on-demand rates is tens of dollars per run. A fleet-wide predictive maintenance retrain is similar. A digital twin or synthetic data cluster runs for days on eight or more GPUs and needs a separate budget assessment.

A simple rule covers most plants. Steady 24/7 inference at the line belongs on the edge device, which is cheap and keeps the decision local. Bursty work, which means retraining, simulation and experiments, belongs in the cloud, where the GPUs are paid for by the hour and released when the job ends. Buying GPUs for bursty work means paying for idle hardware most of the month, and the hardware generation turns over before the purchase is amortized.

Matching the stage of adoption to the GMI Cloud layer

GMI Cloud is a Reference Platform NVIDIA Cloud Partner, providing infrastructure built on NVIDIA's reference design. On October 1, 2026, the company announced US$668 million in new financing: US$223 million in Series B equity, led by ARCHIV with participation from NVIDIA, and a US$445 million credit facility led by CTBC. The financing will support capacity expansion in the United States, Taiwan and the rest of Asia-Pacific. For manufacturers, location matters alongside capacity. GMI previously announced a US$500 million AI factory in Taiwan in November 2025, with WiAdvance among the partners announcing smart-manufacturing and computer-vision use cases. Plants that need production data to remain in-country should confirm the deployment location and data-handling arrangements with GMI as part of their compliance review.

Table 3 shows which layer to reach for at each step.

Table 3. Matching the stage of adoption to the infrastructure layer. Source: GMI Cloud.

A 90-day plan for AI in manufacturing

Weeks one to four go to data that already exists. Pick one line and one problem, usually visual inspection or a recurring failure on a critical machine. Pull the history from the equipment records and the camera archive instead of installing new sensors, and have the quality engineers label it, because labeling is often the most time-consuming step. Name the owners: a controls engineer for the machine connections, a quality engineer for labels, IT for the secure data connection. Then set the target. McKinsey, citing a Manufacturing Leadership Council survey, reports that nearly 60% of manufacturers lack clear targets for measuring AI value, and where targets exist nearly two-thirds of companies meet or exceed them. False-reject rate, minutes of unplanned downtime, scrap percentage: pick one.

Weeks five to eight go to the model. Train on cloud GPUs against the data you collected, starting from a pretrained vision or time-series model, and evaluate against the target. For inspection, export the compact model to the edge device at the line and run it alongside the current inspection process without letting it reject parts. This shadow-mode trial lets the team validate its decisions before putting it in control. For maintenance, deploy it as a service that processes each batch of machine data.

Weeks nine to twelve go to measurement and the scale decision. Compare the target metric against the baseline, document how often you needed to retrain, and price the next step with numbers: how many GPU hours the training loop consumed, how many inferences per day the line generated, and what in-country placement would change. That is the business case for the second line, and it is built on production data.

Where to start

The surveys point to a gap between adopting AI and scaling it. Expanding beyond a pilot means preparing data from more production lines and reassessing compute needs. The sequence matters. Get telemetry and images out of the OT layer first, pick one target metric, and let the workload decide the compute: burst training for inspection, small frequent jobs for maintenance, multi-node clusters for simulation, steady inference for copilots. If you want to price that first 90-day loop with GPU hours, the GMI Cloud team can walk through Table 2 against your line's data and return an estimate within a week. Customer stories show how other teams made the same move.