AI and Manufacturing: Insights from NVIDIA GTC 2025
In this blog
NVIDIA GTC has evolved and grown substantially over the past few years. The event that I attended in 2025 was definitely different from a couple of years ago, in good ways. The event brought together a great intersection of industries, experience and use cases, not the least of which was manufacturing, to my delight!
Booth demonstrator
I had the opportunity to interact with many people during the two days that I was in the WWT booth and enjoyed the discussions with people who shared their journeys with the technologies NVIDIA and our common partners are enabling for their business. The WWT booth was focused on providing demonstrations and discussions on our journey with AI and how our teams were utilizing the tools to gain efficiency and insights. I will touch on a few of the tools here, but the team brought many more to the conference.
Atom Ai
Atom Ai is WWT's generative AI assistant that uses natural language processing to make it intuitive to engage with. The demonstration started with the interface and how natural language prompting enabled our teams to efficiently access multiple data sources from our internal platforms and repositories. I expected many to relate to this but be a bit desensitized to the impact since almost all have likely engaged with ChatGPT, Grok, Copilot and others. Still, people were interested in how they could enable this enterprise-specific functionality for themselves, especially when shown how WWT has integrated Atom into the wwt.com platform to enable seamless work enablement, such as summarizing a blog or white paper.
As our guests continued to talk and think through this, they ultimately landed on two questions: How long did it take? And what is the ROI? I will answer the ROI question first by saying that ROI in AI, to me, is both a mindset (more on that later) and quantitative value. From a qualitative aspect, the teams have seen 25- to -35 percent of what I will refer to as "task energy," meaning that they have reduced the human energy to complete research, document creation and communication using Atom Ai. The "how long" question was addressed by describing the iterative journey WWT went on over 2+ years. Atom Ai began with a basic understanding of WWT processes and knowledge base. Through data integration with ServiceNow, document repositories, including the WWT platform, Atom expanded in experience, which was enhanced when the WWT development team implemented technical advancements with the use of natural language processing conversion to insightful queries that produced more accurate and timely responses.
Many of these discussions resulted in the outcome of how WWT can help our clients achieve this enterprise-relevant AI tool for themselves, taking their specific data sources and security into account while also considering the human aspect of training and adoption.
AI Agentic Network Assistant
This demonstration truly resonated with both the attendees and me, as it explored possibilities aligned with my career journey from a network engineer with programming experience to adapting to the evolution of DevOps, NetOps, SecOps and SecDevOps. The reality is that there is many data in the infrastructure that can and has been used for the benefit of the enterprise for a while now through multiple platforms. Still, the elusive SPOC (single pane of glass) has never truly revealed itself, at least to me. I am not saying that this is it, but perhaps progress towards a simplified interface into the infrastructure for the varying personas and skill levels.
The Agentic Network Assistant allowed our visitors to leverage natural commands, not specific syntax based on platform or vendor, to query the environment and receive a response. As examples, "What interfaces are currently up?" or "What is the current CPU usage?" would be "translated" into commands based on the equipment vendor and platform, Cisco in this case, and issued against the device(s) to give a response back to the requestor.
The tool also offers the ability to run compliance, think diff of "golden config" against a running config, and execute run books (YAML) that execute commands and return the output/result of the commands back to the assistant. Again, this is not a "new" capability. Still, the ability to create multiple workstreams, utilize varying sources of data and utilize NLP prompting to query individual devices or entire locations is helpful to engineers and architects.
I started to think about how we could continue to evolve aspects of this foundation to create the ability to gain insights into networking and even processes running within a manufacturing environment based on the health of the infrastructure that supported them. This would allow control engineers and operators to ask, "What is the status of line 1?" and provide insights into the overall status based on the returned values from the translated commands back to the operator. I saw examples of this idea becoming a reality as industrial chatbots were shown throughout the expo halls, which were integrated into industrial equipment and machine makers for their own process and equipment training, maintenance, and operation. The iteration of agentic assistants means that the varying machine and platform could potentially be utilized as sources of information to an overall plant assistant by asking the equipment agent, "How are things going?" I'm excited to see how this interaction with our day-to-day environments and processes continues to develop.
Session attendee
Admittedly, I did not have the opportunity to attend as many sessions as I would like or had registered for. There was a topic and depth of that topic for all participants; remember, this conference, at its core, is a developer's conference. I cannot say enough about the quality of all the speakers I attended sessions with. They were relevant and engaging, with many storytellers who walked the attendees through their personal and professional journeys, developing use cases, strategies and solutions.
Throughout the customer-led sessions, a few common themes resonated with me, and I believe they helped develop a mindset toward AI in our industry and individual companies.
AI is a journey
As I was asked in my booth demonstrator persona about the timeline to accomplish outcomes, many presenters revealed that the solutions and benefits they were presenting were not short pursuits. Rather, they were years (often three to eight years) of effort to achieve what they were speaking to us about that day. There were missteps and disappointment along the way, but a typical development method of learning from the prior attempt and iterating won out in many cases. I want to amplify that the missteps mentioned were not just in the technology decisions or configurations but also included the other aspects of a holistic solution: people and process. The elements of understanding how people will use the solution and be trained and how the solution will change or integrate into existing processes are essential to successful AI solution outcomes. The reality is that there WILL be disruption, not all of it positive, but ultimately, believing that there is an opportunity for us to progress the industry like never before is not one we can walk away from.
ROI status is "complicated"
I stated there would be more on this topic. I've received questions on ROI internally and externally, which I understand entirely. I wanted to hear from others who were well on their AI journey about this, and what I took away was that even after years, the ROI was not clear and was one of the consistent barriers to progress as they tried to justify investments within their organizations.
There are immediate areas of consideration for ROI, and they follow the same metrics used for years to justify technology investment within the manufacturing industry.
This is not an exhaustive list and will vary by organization, but some examples are reduction of unplanned downtime, reduced scrap, quality improvement and OEE. Continuing to justify the investment in AI in those same areas limits the opportunity for AI to transform the business because many manufacturing clients have solutions for these issues and are satisfied with the results. Areas of impact to consider are workforce (retention, attraction, enablement), customer experience and risk. The risk topic can branch into multiple teams and outcomes, such as human safety, unplanned action, security, and intellectual property risks.
The areas our manufacturing team at WWT speaks with clients about are looking at solutions with a broader view across the network, computing, security and data architectures. This means that AI investment will cross multiple budgets and impact various workflows between enterprise (IT) and line of business (OT). This concept accentuates the need for IT and OT to partner; the convergence has been happening for a while now and needs to be either enhanced or accelerated through a more developed partnership between the teams.
I am not arguing that AI should unthinkingly be funded without any payback to the business, but rather consideration for the complexity and duration of AI varying from traditional technology adoption. Some of the above topics have been said to be six to 18 months with a 150 percent -200 percent ROI. However the consideration of how large the scale of these deployments across the company is and how broad the impact is will determine if those observations hold true or if the grander scale needed to support varying use cases, locations and business models change these for each client.
Start small
Starting small sounds contradictory to my call to think more broadly in the ROI section, but planning broadly and at scale can be separate from the action of actually starting somewhere, and this is what all of the sessions I attended emphasized. This mentality has served other areas of technology well, moving from waterfall to agile development as an example, and helps with iteration and innovation. There has been history for many of us throughout our careers and trends in the industry. (I)IoT comes to mind specifically, where we have learned that the frameworks and equipment can be made to work for a focus area or piece of equipment, but once we work to scale out the areas of performance, management and cost quickly put a strain on the adoption of what the solution is capable of.
It is advisable to commence with the creation and prioritization of use cases well before considering technology solutions. Next, active stakeholders and subject matter experts should be involved to ensure that what is being considered and developed is accurate and that the assumptions are limited. The team will be small but will assist in preventing loss of focus and scope creep while working to accomplish the solution outcome. Understanding existing tools and data sources is an important step in determining the what and how of building an AI solution to impact the use cases defined. Enter into a lifecycle of development and iteration based on decided platforms, tools, data and partnerships.
Take a builder mentality
Interacting with people at the conference and listening in sessions made it clear that AI takes a village, like many things in our lives. For me, a builder mentality clicked as I thought back to when I built my house and had to interact with many partners and functions of building the house. I acted as a general contractor and engaged multiple partners who I felt could best help me provide a home for my family.
The decision I made on who to work with was based on multiple qualifiers; one of the most heavily weighted ones was who I could rely on to communicate, collaborate and deliver on what we agreed to offer together. I have discussions with internal WWT teams and clients where we discuss the experience layer of their considerations in this, and my input to them is to recognize that many companies and partners are developing their portfolios and capabilities at the same time that the client is, so the consideration of who do you feel brings the same passion, insights and capabilities to help them be successful is more important to me than how many "times" they have done it.
When I was done with my other personas of booth demonstrator and session attendee, I was excited to walk through the multiple expo halls set up to experience and interact with the vendors and partners that bring solutions to life with NVIDIAand their own platforms and intellectual capital. A few of the sights I took in and stopped to discuss further were digital twins, robotics and industrial agentic assistants.
Digital twins were everywhere and were shown to impact so many facets of a company's departments and processes. There were demonstrations of how research and design are accelerated due to the ability to make non-destructive changes to other team members' work within a digital twin and any physical models that may be built. This use case helps in the reasoning of ROI in my mind, as discussed above, where capital is saved by not having to physically remodel based on miscalculation and consolidation of design platforms between the team members. Digital twins are also shown to enable other facets of client business, with marketing as an example. The ability to generate realistic models of new products within varying environments without the need to physically go to "location" to create the idea of the product working within varying client scenarios.
Robotics was, of course a huge draw. Admittedly, I did pat a head of a quadrupedal (I could not resist) as it was lying on the floor. The humanoids were talking on the floor, vacuuming carpet, shaking hands, and drawing large crowds. There are advances in the intelligent capabilities of this physical AI, though many are still questioning where this technology can be most effectively enabled for the business. WWT has clients that are bringing bipedal and quadrupedal humanoids into their environments to understand this, and we see many possibilities in repetitive tasks, human co-working, and monitoring (paper route) functions that the technology can impact with the monitoring use case helping in one area I mentioned before with risk in human safety. There are challenges to overcome for specific movements that some of the robotics cannot accommodate and need physical modification. Still, without the builder mentality and lifecycle approach, the outcomes may not come to reality for clients in their work streams and environments.
The final topic I put forward was industrial agentic assistant. The industrial aspect of the agent has nothing to do with the form factor but rather the intelligence, platforms and processes that it is interacting with. I had put forward my vision of this earlier in the Atom Ai section. As mentioned, this is enabled by the capabilities of agents amongst various data sources, OEMs and processes. An example of a starting point was shown by an industrial automation company that had integrated an industrial assistant into its HMI, allowing an operator to utilize natural language to engage with the machine for training, maintenance, and overall process health. An additional aspect of this solution was that the machine had integrated NVIDIA hardware and software to allow the internal process that was using X-ray to be exposed to the operator in the form of a digital twin and AR glasses. I was very interested in this as the machine enabled two AI topics, a digital twin and an industrial assistant, in one, which could potentially be used as a source for an agentic community within the overall facility, which was a single part of the production process.
Closing
In closing, I walked away from the conference more informed and energized to continue the journey of AI within manufacturing. I have experienced many success stories, heard about missteps, and been exposed to some of the capabilities yet to come. I encourage everyone to keep thinking of ways to positively impact your company's AI journey and who you can go on it with from internal and external partnerships.