Las Vegas (DTW 2026) - We didn't come to Dell TechWorld to talk about the future of AI. We came to show what's already working and to listen.

For the past year, Khasm Labs, together with AT&T, Microsoft, Dell Technologies, and Archetype AI, has been running a live AI deployment at intersections in Bellevue, Washington. Our Traffic Management Lens application, built on Archetype AI's Newton model, runs at the network edge with Microsoft Azure handling enterprise orchestration above it. It monitors pedestrian behavior in real time and physically extends walk signals when someone needs more time to cross safely. Not a pilot. Not a simulation. A production system making safety under three seconds, continuously, at live intersections with AT&T providing the connectivity fabric and Microsoft infrastructure backbone being built on Dell Technologies hardware in cloud governance layer that ties it together.

That year of building taught us something more valuable than the technology itself: the hardest part of edge AI isn't the model. It's the system. We went through a full iteration cycle foundation models that were built for the edge but is now, video path latency we didn't anticipate, cellular network topology that added hundreds of milliseconds we couldn't afford. We failed to trigger signal extensions 100% of the time in early testing. We fixed it layer by layer, shaving milliseconds at each step, until we drove efficacy above 90% — meaning the system successfully extends pedestrian walk phases in over nine out of ten scenarios where intervention is physically possible within the signal timing window. That's a meaningful bar in a safety critical, real time environment. And we're not done.

The question most enterprises aren't asking yet

Here's what gets overlooked in most AI conversations: when we first began working with the City of Bellevue, edge AI inference as a cloud service simply didn't exist commercially. Like most enterprises, Bellevue was evaluating what was available which meant any serious edge AI deployment required physical infrastructure at every intersection. Servers in the cabinet. Hardware to procure, install, secure, and maintain. A proprietary platform that would be obsolete before it was fully deployed.

Based on Bellevue's own infrastructure cost benchmarks from comparable deployments, that model carried a potential price tag of $30,000-$50,000 per intersection in edge server CapEx, plus $10,000 per intersection annually in operational overhead. Across 60 intersections, that's $1.8M-$3M before a single inference runs — and a five-year total cost of ownership approaching $5.7M.

The Telco AI Cloud architecture changes that equation entirely. No hardware. No CapEx. A cloud based inference model with a five-year TCO roughly $3.8M lower than the legacy alternative and no lock-in to a hardware platform that ages while the AI around it accelerates. Bellevue gets the benefit of continuous technology improvement at cloud refresh rates, not hardware replacement cycles.

Most enterprises don't know this model exists yet. They're still designing AI strategies around the infrastructure constraints of three years ago and more importantly limiting their ability to pull critical data from the edge necessary for improving the performance, and accuracy, of their AI investments . That's the conversation we're trying to change.

What would your AI strategy look like if you could inference your data at the edge securely, at cloud scale, without deploying a single server? What operational decisions that feel out of reach today would suddenly become possible? What agentic AI actions would you put in motion?

The pattern that emerged at Dell TechWorld

What struck us during the session was how quickly enterprises in the room recognized the Bellevue pattern in their own operations. Different industries. The same underlying problem.

A client from the energy and utilities sector is exploring how routine fleet vehicles, already driving established patrol routes, can become continuous sources of corridor intelligence. Cameras stream visual data as vehicles move. The edge inference layer processes that footage in near real time, flags vegetation encroachment and corridor exceptions, and generates prioritized work packages for operations supervisors before the vehicle returns to the yard. No new infrastructure. No change to existing routes. The fleet they already operate becomes an always on sensing network.

A client from the industrial sector has thousands of legacy analog gauges on equipment that was never designed to be connected. AT&T SIM connected cameras mount directly on existing gauges and stream continuously into the edge inference layer, where a computer vision model reads the gauge face, digitizes the value, and monitors for anomalies. When something looks wrong, it surfaces into Azure for agentic action within the client's existing operational environment. Not a single gauge replaced. Not a single legacy system ripped out.

We're also extending TML itself into a new context: detecting traffic accidents at intersections in real time, assessing incident criticality, and delivering a structured incident summary to an emergency dispatch operator before the first 911 call arrives. The same inference architecture. A completely different operational outcome.

In every case, the pattern is identical: data being generated at the edge faster than it can be acted on, with decisions that lose value every second they're delayed. The enterprises in that room didn't need convincing that the problem was real. They needed to know the infrastructure to solve it finally exists.

What’ s next

The Telco AI Cloud architecture is moving from proof point to platform. Bellevue validated the model. The deployments we're developing across utilities, industrial operations, and public safety are expanding it. And the conversations we had at Dell TechWorld told us the demand is real, cross industry, and further ahead than most enterprises realize.

If you're working through what edge AI could mean for your operations – or if you’re still trying to figure out where to start — we'd welcome the conversation. The infrastructure that makes this possible exists today. The question is what you'll build with it

Khasm Labs Contact: [email protected]