Having recently attended NVIDIA’s GTC 2025, widely dubbed the “Super Bowl of AI,” I was struck by how it underscored the extent to which artificial intelligence has redefined computing.
As often emphasized by Salute, the convergence of advanced compute capabilities and the increased demands prompted by modern infrastructure design has radically redefined data center operations. With that has come a significant impact on power and cooling demands, requiring precise planning and execution courtesy of novel innovations such as digital twin environments, both crucial to optimizing performance in line with sustainability concerns.
Hearing one of the industry’s most forward-thinking leaders explore these themes firsthand
brought fresh clarity to the opportunities and challenges facing other leaders. Though there were
numerous takeaways from this talk, I feel the following best reflect the future of data center
innovation:
Agentic AI – A New Chapter and Era
NVIDIA CEO Jensen Huang began the keynote by tracing the trajectory of AI from perception to generative and now to agentic AI. Each chapter has brought with it increased capabilities, with agentic AI marking a newfound ability to reason, plan, and act. By understanding context, using tools, and executing complex problems, agentic AI introduces a new level of autonomy. However, this leap in functionality comes at a cost. Specifically, it requires 100 times more computational power than prior models. Naturally, this rapid increase in processing makes higher-density compute, low-latency networks, and energy-efficient design foundational.
Redefining Performance Through Architecture Advancements
Another key highlight of the keynote was NVIDIA’s Blackwell platform, which represents a major shift in data center architecture. Each Blackwell GPU contains two GPU dies, enabling previously unimaginable compute performance. With liquid-cooled racks housing 600,000 components, organizations can now achieve one ExaFlops of compute in a single rack—a feat that previously required more than 1,400 racks. This, in turn, has enabled a jump in AI token generation capacity from supporting 300 million tokens in a 100-megawatt facility to 12 billion. Similarly, Jensen highlighted the industry shift toward rack power densities exceeding 130kW per rack with Grace Blackwell, projections of 200kW per rack with Rubin, and 600kW with Feynman in the years ahead. Certainly, it seems no understatement to suggest these advancements will fundamentally reshape the architecture and operation of AI data centers worldwide.
AI Factories and Digital Twins: Designing for What’s Next
As Jensen emphasized, modern AI infrastructure is no longer a generic compute environment but an “AI factory,” a term that reflects the ability of data centers to process billions of tokens, coordinate massive model training, and execute reasoning-based inference. To manage this, NVIDIA has introduced tools like the Omniverse and Cadence Reality digital twin platform, both of which allow for precise modeling of a data center’s performance, thereby identifying potential design flaws and optimizing thermal flows in real time.
Scale-Up Before Scale-Out: Reinventing the Rack
NVIDIA’s keynote also highlighted a strategic shift in infrastructure philosophy, namely the importance of scaling up before scaling out. Rather than distributing compute horizontally, maximizing performance within each rack is the new priority. Technologies like NVLink and SpectrumX Ethernet enable extremely low-latency communication between GPUs, making scale-up feasible and efficient. NVIDIA also introduced Dynamo, a new open-source operating system that orchestrates multi-GPU workloads with seamless coordination, allowing faster inference, better utilization, and lower operational overhead.
From Newton to Robotics: The Broader Impact of AI
Beyond data centers, NVIDIA also showcased how AI is transforming the physical world. Their partnership with GM to deliver autonomous vehicles, along with the introduction of Newton—a humanoid robot powered by the ISAAC GROOT N1 foundation model—demonstrated the real-world applications of AI reasoning. Of course, the fact that these systems rely on continuous learning, large-scale model training, and instant inference requires robust, scalable infrastructure behind the scenes, something Salute is working to provide.
Salute’s Approach to Powering the Next Generation of AI Infrastructure
This evolution of AI workloads from generative to agentic, driven by the emergence of innovations such as Grace Blackwell, Rubin, and Feynman, signals a significant shift in what data centers must deliver. These platforms are driving rack power densities far beyond the limits of traditional air-cooled environments, making the delivery of direct-to-chip liquid cooling at scale a vital next step for sustainably supporting high-performance compute. As expected, this transition introduces considerable operational complexity, requiring precise commissioning, specialized training, and resilient facilities management capable of supporting increasingly dense and sensitive infrastructure.
Across the industry, data center operators are actively exploring how best to respond, with questions surrounding reliability, performance, and sustainability playing a central role in protecting GPU investments and driving return on investment. In this respect, operational readiness depends not only on a focus on the technical side of things but also on planning and deployment strategies that prioritize tighter coordination between design and delivery teams, more refined internal processes, and resilient long-term service models. Put simply, as technological demands grow more complex, so too must the strategies that guide how infrastructure is designed, delivered, and sustained.
It is for this reason that Salute is so dedicated to building the capabilities required to support this next chapter of infrastructure transformation. Through a focus on advisory, commissioning, and facilities management services tailored to direct-to-chip liquid cooling environments, we are actively aligning our approach with the emerging demands of NVIDIA’s AI and other HPC workloads. Ultimately, our goal is to help operators navigate this complexity with confidence, developing strategies that support scalability, protect infrastructure investments, and ensure operational excellence across the lifecycle. In so doing, we ultimately aim to lead the AI Data Center Evolution, the next step in our already extraordinary journey.
Discover a Smarter Way to Scale Your AI Infrastructure with Salute
As the AI era accelerates, Salute stands at the forefront of delivering the physical infrastructure that will power it. With a global footprint spanning 12 offices, 1,800+ employees, and operations in over 102 markets—adding up to support for 80% of the world’s data center operators—we bring trusted expertise across the full lifecycle. Whether scaling an AI factory, upgrading for dense rack requirements, or deploying next-gen cooling systems, Salute delivers the insight, execution, and operational excellence to scale safely and efficiently.