Jags Kandasamy of Latent AI winning the Networking & IT category at the 2026 Data Center World Innovation Challenge, Powered by ABB.
The startup LatentAI began with a focus on defense industry use cases, optimizing AI models so they require significantly less computing power. That efficiency is essential for edge environments where power availability is constrained, such as drones, autonomous systems, and military deployments running on batteries.
What began as a defense challenge is now becoming increasingly relevant to data centers.
“When we looked at the power budgets and costs spiralling within data centers, we realized there was a parallel,” says Jags Kandasamy, Co-Founder and CEO of Latent AI. “Could we take the same technology and reduce the power budget there as well?”
At the center of Latent AI’s philosophy is the idea that not all data deserves equal treatment.
Instead of constantly expanding infrastructure to accommodate growing AI workloads, Kandasamy believes organizations should first rethink how those workloads are designed, processed, and distributed across networks.
Latent AI’s software helps organizations run AI models efficiently on edge devices such as cameras, sensors, drones, and industrial equipment. By optimizing and compressing AI workloads, the company allows data to be analyzed closer to where it is generated, reducing the amount of information that must be transmitted, stored, and processed in centralized data centers.
“AI workload is causing that explosion in data center buildouts and power consumption,” he says. “The question is whether we can fundamentally reduce that footprint.”
Latent AI won the Networking & IT category at the 2026 Data Center World Innovation Challenge powered by ABB. The Innovation Challenge invited 24 startups to pitch their data center innovations Shark Tank-style to a panel of judges. Judges selected winners in seven categories.
Consider Data Value Before AI Processing
Today, many systems are designed to capture, transmit, and process enormous amounts of information without considering its actual value. AI then sits at the end of the chain, consuming significant compute resources to make sense of it all. Kandasamy argues that future architectures should work differently.
He points to motion-triggered doorbell cameras, which capture and transmit footage whenever movement is detected. In practice, that means insects, shadows, passing vehicles, and countless other irrelevant events can generate alerts and network traffic.
Latent AI’s software lets companies build AI-first architectures where intelligence is distributed across multiple layers. A lightweight model at the edge could determine whether a human is present, and only then would relevant information move further into the system. Then more sophisticated AI models could do facial recognition, behavioural analysis, or threat assessment.
“Now you’ve cut down the whole process to only sending information when certain events are triggered,” he explains.
AI and Internet of Things Will Also Strain Infrastructure
Generative AI may dominate headlines, but Kandasamy believes many of the biggest infrastructure challenges will emerge from vision-based applications and real-world sensors generating continuous streams of data.
For example, a municipality with hundreds of intersections and thousands of cameras could spend tens of millions of dollars annually simply moving data to cloud environments before any processing takes place. In many scenarios, reducing the amount of data that needs to be transmitted can create major savings in networking, compute, and energy.
In enterprise computing environments, AI workloads still represent a relatively small percentage of overall computing today. Most advanced generative AI systems are concentrated within hyperscalers and other cloud providers. But over the next decade, organizations across manufacturing, logistics, transportation, healthcare, and industrial operations are expected to deploy AI systems closer to where data is generated. As that shift occurs, efficiency becomes critical.
Latent AI recently worked with a large manufacturer using vision-based quality control systems. Initial pilots required dedicated NVIDIA GPU servers for individual production lines. Scaling that deployment across hundreds of production lines would have required substantial IT infrastructure investment. By optimizing inference workloads, Latent AI helped it reduce hardware requirements by about 93%.
“This is where we help them scale without breaking the bank,” Kandasamy says.
The AI Pain Comes After the Pilot Project
Kandasamy says Latent AI typically enters the picture as companies look to move from successful AI pilot projects to large-scale deployments.
“The pilot is about proving the concept works,” he says. “When you try to scale that same architecture, all the costs come with it.”
When those costs become clear, organizations then revisit and redesign systems to focus on operational efficiency and not just functionality.
Networking’s an Underappreciated Data Center Constraint
Kandasamy believes networking will become even more critical as AI workloads become distributed across multiple layers of centralized and edge computing environments.
While GPUs have become the defining hardware story of the AI era, he expects the chip architecture to become increasingly heterogeneous and complex, with specialized accelerators handling specific workloads alongside CPUs and GPUs.
The challenge will then be ensuring data moves efficiently between them. “Networking has always been the key factor for compute performance,” he says.
Rather than operating as isolated facilities, data centers will increasingly function as part of distributed ecosystems where workloads shuttle across locations depending on available resources, latency requirements, and energy constraints.
“The way workloads are distributed is going to be the biggest factor that changes,” Kandasamy says.
Latent AI is helping organizations move beyond the constraints of heavy, cloud-dependent AI by enabling efficient, adaptive models that run closer to the data, reducing latency, cost, and infrastructure complexity at the edge. As AI demand continues to grow, efficiency could become every bit as important as scale.