Spend any time with manufacturing executives and you’ll start to hear the same message: The robotics push of the 2010s delivered real gains, at least for a while. Then, enterprise resource planning upgrades brought cleaner data and standardized processes, while IoT rollouts generated enormous amounts of sensor data.
And yet, the core metrics tell a different story — one of yield, unplanned downtime, defect escapes and inventory turns. Across the board, progress over the past decade has been marginal, at best, for most manufacturers.
In the US, for example, labor productivity in the manufacturing sector increased at an annualized rate of only 0.5% between the fourth quarter of 2019 and the first quarter of 2026, as output rose by 0.1% and hours worked declined by 0.4%.
These figures point to a strategy problem that needs to be solved, and urgently, as the global market for AI in manufacturing is projected to grow from $34.2 billion in 2025 to $155 billion by 2030, at a compound annual growth rate of 35.3%.
Where manufacturing falls short on intelligence
Traditional automation makes processes faster and more repeatable, but it doesn’t make them smarter.
A robot can weld the same joint consistently 10,000 times a day, but it won’t notice that a new batch of steel behaves differently, that a bearing is likely to fail in two shifts or that a production schedule should be adjusted because a downstream bottleneck is about to clear. These judgment calls have long depended on human experience, which is costly, scarce and difficult to scale.
There are four broad factors making this an urgent problem to address:
- Supply chains remain fragile: Problems that started in 2020 have only become more complex. Geopolitical tensions aren’t helping.
- Skilled labor is thinning out: Retirees are outpacing new entrants in most manufacturing economies, and much of their knowledge leaves with them.
- Sustainability requirements are tightening: Regulators and major customers now expect detailed, auditable proof, grounded in operational data.
- Customization is the norm: Shorter production runs, more variants, greater personalization and faster changeovers — the bar keeps rising.
Conventional automation doesn’t ease these pressures, but physical AI can. It enables systems to perceive what’s happening in your operations in real time, then both understand those conditions and act on them.
What’s changed: The underlying AI and environment around it
Perception and emerging world models are getting better at understanding physical environments and anticipating what happens next. More capable compute can support training at scale, while digital twins can test those models and generate synthetic data from virtual factories, equipment and workflows rather than relying only on real-world trials.
But having these capabilities is only part of the equation. For manufacturers, the bigger challenge is integrating them with the operational technology, enterprise systems, data and workflows that run the plant.
At NTT DATA, we see physical AI as an integration challenge as much as an AI challenge.
NVIDIA’s physical AI stack
NVIDIA provides an important part of the technology foundation for this work. Its physical AI stack spans models, libraries, frameworks, and accelerated computing for training, simulation and real-world deployment.
The stack can be thought of in three layers:
- Models: The NVIDIA Cosmos world foundation models help physical AI systems understand and reason about the physical world.
- Software, libraries and frameworks: NVIDIA Isaac Sim provides a simulation environment for developing and testing robotics applications. NVIDIA AI Enterprise provides software for deploying and managing AI in production and includes NVIDIA Omniverse, Metropolis and Triton.
- Compute: NVIDIA DGX provides centralized infrastructure for training AI models, RTX PRO Servers provide the compute needed to run simulation and digital-twin workloads, and Jetson brings AI inference to robots, machines and other edge systems.
How manufacturers are already benefiting
The improvements seen in manufacturing operations are striking:
- A global automaker deployed digital twins built on NVIDIA Omniverse libraries across more than 30 plants. The time needed to perform virtual automated collision checks — to make sure new vehicles heading to the production line don’t collide with their surroundings — dropped from four weeks to three days, and there was a projected 30% reduction in production planning costs.
- Electronics manufacturers are finding ways of improving quality control and productivity. Foxconn Industrial Internet is working with NVIDIA Metropolis partners to automate circuit-board quality inspections and NVIDIA RTX PRO Servers servers to run large language models developed on DGX SuperPOD at the edge. Innodisk is using NVIDIA Metropolis to automate optical inspection on production lines, while Wistron is deploying it to partially automate the optical inspection of circuit boards.
- In industrial defect detection and classification, NVIDIA’s end-to-end deep-learning platform achieved a 96.4% recall rate, a 99.3% precision rate and a 0.1% false-alarm rate in a demonstration while delivering major inference gains through TensorRT optimization. On edge deployments, the same approach processed 228.1 frames per second (FPS) on Jetson AGX Xavier compared with 18 fps on Jetson Nano, illustrating how AI-based inspection can combine detection quality with production-ready throughput.
NTT DATA’s services complement the NVIDIA stack
Many manufacturers’ AI initiatives make the mistake of treating each AI use case as a standalone project. They source vision AI from one vendor, predictive maintenance from another and digital twins from a third. Each comes with its own data pipelines, deployment stack and maintenance needs.
Then they discover that every additional use case introduces another integration layer, and with it, more complexity and a weaker ROI case.
There’s no shortage of vendors in this space. But in manufacturing, AI delivers only when the organization implementing it understands the realities of the factory floor as well as the technology itself.
NTT DATA’s global Manufacturing practice has spent more than two decades working with industrial clients in the automotive, electronics, aerospace, process, and food and beverage industries. We know what manufacturing execution system (MES) integration looks like when the MES is 15 years old, how a predictive maintenance model behaves when on-the-ground practices drift from training conditions, and which data governance issues can derail a program before it gets off the ground.
Our AI for Manufacturing platform builds on that foundation. It’s an agentic AI orchestration layer designed specifically for this industry, sitting on top of NVIDIA’s stack to turn AI outputs into operational decisions in real time. It includes a library of more than 50 prebuilt agents and several specialized industrial AI models, with low-code orchestration that integrates directly into NVIDIA’s computing and inference infrastructure.
NVIDIA provides the AI infrastructure and open models, libraries and frameworks. We connect that capability to the systems and workflows that run the plant.
What’s next in this series
There’s a wide gap between what manufacturing AI can do today and what most plants are running. In many cases, the missing piece is a structured way to move from isolated pilot projects to an enterprise operating model.
In our next blog post, we’ll look at how NTT DATA’s AI for Manufacturing platform brings these capabilities together, combining NVIDIA’s physical
AI technology with the agents, orchestration and integration needed to connect AI to real manufacturing operations.
In the final installment, we’ll lay out a 90-day implementation framework to provide a practical starting point for operations leaders who are ready to move. The intelligent factory isn’t something you arrive at all at once; it’s built step by step, through the decisions you make today.
WHAT TO DO NEXT
Read more about NTT DATA’s Manufacturing services and get in touch with our team to see how we can help you produce faster, move smarter and cut costs.