Article

Nvidia GTC 2025: AI Matures into Enterprise Infrastructure

Companies are changing the architecture for how they compete.

By Eric Sheng, Brock Simon, Sarah Elk, Daniel Reznikov, Srikanth Munjuluri, Sandeep Singh, and Apoorv Sharma

Published on March 24, 2025
min read


We came away from Nvidia’s GTC event last week with a reinforced belief that AI has moved well beyond pilots and proofs of concept to scaled deployment and operational impact. Consistent with what we see in our work with clients, organizations are no longer testing AI at the edges. They are re-architecting how they compete. Across industries, AI is moving from innovation teams into the business core. Below are nine key themes we observed at GTC that are shaping the enterprise AI landscape. 1. “No data, no AI”—and now, AI is generating the next layer of data. Data remains the biggest challenge and the biggest opportunity. Every successful AI deployment presented at GTC rested on clean, connected, and accessible data. But the frontier is no longer just data consumption; it’s data creation. AI is surfacing previously invisible insights: operational patterns, best practices, and in-the-moment decisions propelling activities like frontline sales that can be reused and optimized. Generative AI is also helping address classic data challenges, from search and retrieval to confidence scoring and synthetic augmentation. Organizations that invest in feedback loops and change management are building true data flywheels. 2. Smaller, specialized models are rewriting the economics of AI. The dominance of large, general-purpose models is giving way to smaller, fine-tuned systems built for specific domains. Techniques like quantization, pruning, and retrieval-augmented generation (RAG) are unlocking cost savings without sacrificing performance. Enterprises are increasingly embracing model fine-tuning and self-hosting for better control, lower latency, and improved privacy—though doing so brings operational complexity that few organizations are fully prepared for at present. 3. Agentic AI is starting to gain traction—and trust depends on structure. The progression has moved from RAG to AI assistants and now agentic AI. Fully autonomous AI agents, however, are rarely deployed at scale. The biggest challenge remains evaluating the accuracy of an agent’s output, something both business leaders and engineers are grappling with. But a clear theme has emerged: Structure matters. Enterprises are prioritizing transparency, escalation paths, redundancy guardrails, traceability and auditability in production, and predictable behavior. While fully autonomous agents remain rare, semiautonomous systems—with human oversight—are the pragmatic near-term standard. Frameworks like Nvidia AgentIQ are helping bring order to this new frontier. There is also growing momentum around “agent orchestration platforms,” with some companies developing in-house solutions. This trend reflects a broader push to simplify the creation and integration of AI agents into enterprise systems. 4. Digital twins and simulation are now everyday enterprise tools. Simulation has shifted from innovation showcase to standard operating practice. Teams are using digital twins to model factories, stores, and supply chains, testing changes virtually before implementing them physically. The result is faster rollout cycles, lower risk, and more confident decision making. Executives increasingly prefer virtual walkthroughs over site visits, especially as twins become more integrated with real-time spatial and operational data. 5. Video is becoming the next major dataset. Computer vision and video language models are transforming video from passive monitoring to active intelligence. Organizations are using video to analyze customer behavior, product interaction, compliance, and safety in real time. These insights are now directly influencing merchandising, labor planning, and layout optimization, turning physical spaces into intelligent, data-rich environments. 6. Enterprises are shifting from build to buy for large language model operations (LLMOps) and infrastructure—and adoption is accelerating. Off-the-shelf tools like Nvidia DGX Cloud and Inference Microservices (NIM) are lowering the barriers to enterprise AI adoption. Companies are launching copilots, knowledge assistants, and other custom gen AI applications without having to build deep machine learning operations or LLMOps stacks. We’re also seeing increased accessibility to advanced capabilities like decision optimization and simulation, now running on out-of-the-box, GPU-accelerated infrastructure. 7. Simulation is emerging as the new collaboration layer. Beyond modeling, simulation is becoming a unifying platform for cross-functional teams. With Nvidia Omniverse integrating seamlessly into design and operations tools, teams can cocreate in virtual environments before making real-world changes. This is reducing iteration cycles, improving coordination, and unlocking faster, lower-risk decision making across functions. 8. Custom model deployment is the new organizational bottleneck. Fine-tuning foundation models is becoming easier, but deploying them into production is still hard. Teams face challenges around performance optimization, latency, hardware compatibility, and security. Enterprises are realizing that production-readiness requires strong internal capabilities and deliberate investment in infrastructure. 9. Multimodal AI is transforming creative workflows and brand expression. Tools like Nvidia Picasso, Adobe Firefly, and open-source diffusion models are enabling teams to generate product visuals, videos, 3D assets, and social content—all from natural language prompts. Companies are scaling content generation with creative pipelines from platforms like RunwayML, Canva, and Synthesia. This shift is speeding up campaign cycles, enabling hyper-personalization, and making high-quality creative accessible to teams of any size.

We came away from Nvidia’s GTC event last week with a reinforced belief that AI has moved well beyond pilots and proofs of concept to scaled deployment and operational impact. Consistent with what we see in our work with clients, organizations are no longer testing AI at the edges. They are re-architecting how they compete. Across industries, AI is moving from innovation teams into the business core. Below are nine key themes we observed at GTC that are shaping the enterprise AI landscape.

1. “No data, no AI”—and now, AI is generating the next layer of data.

Data remains the biggest challenge and the biggest opportunity. Every successful AI deployment presented at GTC rested on clean, connected, and accessible data. But the frontier is no longer just data consumption; it’s data creation. AI is surfacing previously invisible insights: operational patterns, best practices, and in-the-moment decisions propelling activities like frontline sales that can be reused and optimized. Generative AI is also helping address classic data challenges, from search and retrieval to confidence scoring and synthetic augmentation. Organizations that invest in feedback loops and change management are building true data flywheels.

2. Smaller, specialized models are rewriting the economics of AI.

The dominance of large, general-purpose models is giving way to smaller, fine-tuned systems built for specific domains. Techniques like quantization, pruning, and retrieval-augmented generation (RAG) are unlocking cost savings without sacrificing performance. Enterprises are increasingly embracing model fine-tuning and self-hosting for better control, lower latency, and improved privacy—though doing so brings operational complexity that few organizations are fully prepared for at present.

3. Agentic AI is starting to gain traction—and trust depends on structure.

The progression has moved from RAG to AI assistants and now agentic AI. Fully autonomous AI agents, however, are rarely deployed at scale. The biggest challenge remains evaluating the accuracy of an agent’s output, something both business leaders and engineers are grappling with. But a clear theme has emerged: Structure matters. Enterprises are prioritizing transparency, escalation paths, redundancy guardrails, traceability and auditability in production, and predictable behavior. While fully autonomous agents remain rare, semiautonomous systems—with human oversight—are the pragmatic near-term standard. Frameworks like Nvidia AgentIQ are helping bring order to this new frontier. There is also growing momentum around “agent orchestration platforms,” with some companies developing in-house solutions. This trend reflects a broader push to simplify the creation and integration of AI agents into enterprise systems.

4. Digital twins and simulation are now everyday enterprise tools.

Simulation has shifted from innovation showcase to standard operating practice. Teams are using digital twins to model factories, stores, and supply chains, testing changes virtually before implementing them physically. The result is faster rollout cycles, lower risk, and more confident decision making. Executives increasingly prefer virtual walkthroughs over site visits, especially as twins become more integrated with real-time spatial and operational data.

5. Video is becoming the next major dataset.

Computer vision and video language models are transforming video from passive monitoring to active intelligence. Organizations are using video to analyze customer behavior, product interaction, compliance, and safety in real time. These insights are now directly influencing merchandising, labor planning, and layout optimization, turning physical spaces into intelligent, data-rich environments.

6. Enterprises are shifting from build to buy for large language model operations (LLMOps) and infrastructure—and adoption is accelerating.

Off-the-shelf tools like Nvidia DGX Cloud and Inference Microservices (NIM) are lowering the barriers to enterprise AI adoption. Companies are launching copilots, knowledge assistants, and other custom gen AI applications without having to build deep machine learning operations or LLMOps stacks. We’re also seeing increased accessibility to advanced capabilities like decision optimization and simulation, now running on out-of-the-box, GPU-accelerated infrastructure.

7. Simulation is emerging as the new collaboration layer.

Beyond modeling, simulation is becoming a unifying platform for cross-functional teams. With Nvidia Omniverse integrating seamlessly into design and operations tools, teams can cocreate in virtual environments before making real-world changes. This is reducing iteration cycles, improving coordination, and unlocking faster, lower-risk decision making across functions.

8. Custom model deployment is the new organizational bottleneck.

Fine-tuning foundation models is becoming easier, but deploying them into production is still hard. Teams face challenges around performance optimization, latency, hardware compatibility, and security. Enterprises are realizing that production-readiness requires strong internal capabilities and deliberate investment in infrastructure.

9. Multimodal AI is transforming creative workflows and brand expression.

Tools like Nvidia Picasso, Adobe Firefly, and open-source diffusion models are enabling teams to generate product visuals, videos, 3D assets, and social content—all from natural language prompts. Companies are scaling content generation with creative pipelines from platforms like RunwayML, Canva, and Synthesia. This shift is speeding up campaign cycles, enabling hyper-personalization, and making high-quality creative accessible to teams of any size.