At a Glance
- Enterprise AI is shifting from model training toward inference, where operational value depends on keeping data moving at the pace of the models and GPUs consuming it.
- M2M Tech’s Smart Cities solution combines DDN Infinia with NVIDIA accelerated computing to turn traffic detection into an evidence-based workflow.
- NVIDIA AI infrastructure supports the inference path, while DDN Infinia provides the governed data layer for video, detections, audit history, model context, and persistent KV cache available as the workload scales.
- DDN Infinia integrates with the NVIDIA Inference Transfer Library (NIXL) and the prebuilt NVIDIA Dynamo container, enabling KV-cache acceleration through configuration rather than code changes.
- Human approval remains central: the system recommends actions, while municipal operators retain control over infrastructure decisions.
A camera at a busy Seattle intersection picks up traffic backing up across two lanes. For most of the last decade, that’s roughly where the technology’s job has ended. An alert fires, a dashboard turns red, and an operator starts the real work: pulling up the feed, checking whether this is normal for a Tuesday afternoon, digging through permit records, and calling to find out whether there’s a crew on site. The system saw the problem but it didn’t help anyone understand it.
That gap between seeing something and knowing what to do about it is where much of enterprise AI is stuck today. It’s also the gap M2M Tech set out to close with a solution built for municipalities.
The Market Is Moving to Inference. Most Organizations Haven’t Caught Up.
Over the past few years, the AI conversation has been dominated by training: bigger models, bigger clusters, bigger budgets. Training and fine tuning is how a model gets built. Inference is how it earns its keep. Every answer, detection, summary, and recommendation a model produces in the real world is inference, and that’s where the return on all that training investment shows up or falls flat.
Inference doesn’t look the same everywhere. It can be a vision model counting vehicles, a language model answering questions, or a retrieval system pulling the right document at the right moment. In the enterprise, it increasingly takes the form of agents, systems that string those capabilities together to perceive something, gather context, reason about it, and propose an action.
The promise is enormous, and so is the frustration. Many enterprises are still working out where inference fits in their operations, how to apply it to a workflow people depend on, how to scale it beyond a pilot, and how to show a return that justifies the spend. Pilots are multiplying, but production deployments aren’t progressing as quickly. The organizations making progress are tying inference to specific business or operational decisions someone already has to make, then building the infrastructure to support it at scale.
Where AI-Native Companies Step In
This is where the market is shifting. AI-native companies are capitalizing on the move to inference by doing the hard part on behalf of their customers: identifying a specific, valuable decision, building the models and agents around it, and packaging the result as a solution an organization can deploy. The enterprise gets the outcome without first having to become an AI lab.
M2M Tech’s Smart Cities solution brings together DDN Infinia and NVIDIA accelerated computing in a deployment designed for municipal traffic operations. Rather than stopping at detection, the architecture supports the inference workflow that turns live conditions and operational data into evidence-backed recommendations for city operators.
M2M’s system carries an operator through the full path: detect, understand, explain, recommend, approve. It brings together the sources a traffic operator would otherwise check one by one, including live video, vehicle counts, permits, and incident records, into a single operational workflow.
Return to that backed-up intersection. The system notices the buildup and compares live counts against the historical baseline for that location and time, distinguishing ordinary rush-hour traffic from something genuinely unusual. It looks for a likely cause, perhaps a permitted lane closure or a stopped vehicle, and puts the supporting video and records in front of the operator. Then it recommends a next step. The operator can ask follow-up questions in plain language, review the evidence, and approve or reject the recommendation. That decision, along with the evidence behind it, is retained for later review.
Built for How Cities Actually Buy and Operate
What makes this a product rather than a one-off project is how closely it fits the realities of a municipal buyer:
- The AI advises. People decide. Nothing touches city infrastructure without a human operator’s approval, which is the right posture for public-sector operations where accountability matters.
- Every recommendation shows its work. Tying outputs and decisions to the underlying video and records makes them easier to trust; and to defend to a council or the public.
- The data stays home. The solution supports sovereign, on-prem, hybrid, and air-gapped deployments, so city video and records don’t have to leave the city’s environment.
- It builds on what cities already own. It sits on top of existing cameras, traffic systems, and data sources. There’s no new sensor network to fund before a city can see value, which makes a pilot far easier to approve.
- The architecture is the product. The same perception, reasoning, evidence, and approval workflow can be configured with different cities, departments, cameras, data sources, and branding. That’s what lets M2M deploy it city after city, and lets a city extend it from traffic to other departments.
For a municipality, that adds up to a realistic path through a challenge many enterprises face: start small, measure against work operators already do, and expand once the approach proves itself.
How DDN and NVIDIA Support Inference at Scale
As agentic workflows scale, the question is not only how much compute is available, but whether data and context can reach it fast enough. In M2M’s solution, NVIDIA accelerated computing supports the inference environment, while DDN Infinia provides the data layer designed to keep pace with it.
Infinia serves as the system of record for the video, detections, evidence, audit history, and model context the workflow depends on. Built on a scalable key-value store, it brings structured and unstructured data together, including existing data lakes, and is designed for the low-latency, high-concurrency access inference workloads demand. For a city, that means the evidence behind every recommendation and approval lives in one governed place, inside its own environment.
The less obvious piece is what happens as deployments grow. When a language model works through a request, it builds an intermediate context known as KV cache. Discard that context, and the GPU has to recompute it the next time similar information is needed. Agentic workflows revisit the same context constantly: the same intersection baselines, permit records, historical data, and operating procedures, queried by many operators across many shifts.
With Infinia serving as a persistent KV cache layer, that context can be stored and reused rather than rebuilt from scratch. That capability connects directly to the NVIDIA inference stack: DDN Infinia is integrated with the NVIDIA Inference Transfer Library and the prebuilt NVIDIA Dynamo container, allowing KV-cache acceleration to be enabled through configuration rather than code changes. NIXL enables fast and efficient data transfers between servers and storage while Dynamo optimizes inference at scale and KV cache context reuse across multiple GPU servers. In a pilot covering a handful of intersections, that is a useful efficiency. Across hundreds of cameras, multiple departments, and many concurrent users, reuse matters to both performance and economics: less context has to be recomputed, and the infrastructure can keep pace as demand grows. For an AI-native company selling into many cities, it also helps make each new deployment economically repeatable.
The Broader Lesson
It’s tempting to read this as a traffic story that became an interesting technology project. It’s really a story about how inference creates value, and who is turning it into something enterprises can use. The return doesn’t come from a model that can see or talk. It comes from putting that capability inside decisions people already make, grounding it in evidence, keeping humans in control, and running it on infrastructure that can scale without starting over.
Some organizations will get there by adopting solutions like M2M Tech’s, built by AI-native companies that have already done the hard work for a specific domain. Others will build their own agents around processes only they understand. Many will do both. Either way, the same foundation sits underneath it all: the data those systems depend on, and whether it can be reached quickly, governed properly, and shared across every application that comes next.
A city that starts with traffic will want to move on to permitting, emergency response, and maintenance. An enterprise that starts with one agent will want ten, then one hundred or more. The organizations that plan for that from the beginning, with a data layer built to sustain inference performance as they grow, are the ones that will turn a promising pilot into a lasting return.