- AT&T and Cisco are running AI Grid with Nvidia at AT&T's Dallas Discovery District
- The carrier has a second commercial pilot underway at an industrial services firm
- Comcast, Akamai, Spectrum and T-Mobile are also building Nvidia AI grids, signaling a wave of operator momentum behind distributed AI inference
NVIDIA GTC, SAN JOSE, CALIF. — AT&T and Cisco are putting Cisco AI Grid with Nvidia to work in production, moving from reference design to real deployments an architecture that promises to transform telecom networks from data pipes into distributed AI platforms.
Cisco AI Grid with Nvidia is a full-stack reference architecture that pushes AI inference deeper into network infrastructure and closer to customers. The architecture connects AT&T's dedicated IoT core and radio access network directly to Nvidia accelerated compute housed in Cisco-managed facilities, co-located with the mobile packet core and orchestration stack. The underlying GPU hardware is Nvidia's RTX PRO 6000 Blackwell Server Edition.
A live use case is running at AT&T's Discovery District, a four-block entertainment and dining destination in Dallas. A commercial pilot is underway with an industrial services firm in Louisiana.
In addition to AT&T, Comcast is also using AI Grid in three trial use cases, and other carriers are implementing the technology as well.
AT&T's structural advantage
AT&T's converged 5G and fiber footprint gives it a structural advantage for delivering deterministic, zero-trust AI inference at scale — positioning the network not as a transport layer, but as the key differentiator for a new class of enterprise AI applications, AT&T said. That convergence differentiates AT&T's approached from overlay-based networks that introduce additional network hops and expose control planes to third parties, according to the carrier.
"You can't manage or secure a network you don't control," said Shawn Hakl, AT&T Business SVP of product, who spoke with Fierce at the conference alongside his Cisco and Nvidia counterparts. "When you've got more endpoints than anyone else and you've got direct access to the packets, I can guarantee complete end-to-end control over policy, whether it's routing or security, from the very point of inception of the device all the way to where we drop it off."
AT&T's use cases: from corporate campus to industrial yards
At the Discovery District, AT&T is running Linker Vision cameras through its network and into Nvidia compute to monitor common areas in near real time, flagging spills, obstructions and perimeter anomalies. A commercial pilot is now underway with TanMar Companies, a Louisiana-based oilfield services firm deploying the platform for license plate recognition, perimeter intrusion detection and restricted-area monitoring across geographically dispersed industrial sites.
A key enabling technology is vision language models, which allow cameras to function as intelligent sensors rather than simple video feeds. "You can talk to your video," said Ronnie Vasishta, Nvidia SVP of telecom. "You can say, 'Tell me when there's a spill in the garage.' The camera will immediately respond." Over time, he added, autonomous systems could respond without human intervention. Nvidia supports this capability with its Video Search and Summarization blueprint, now in version three, built on the Metropolis video pipeline model.
AT&T's deliberate shift in go-to-market
The three-way partnership also represents a deliberate shift in how AT&T goes to market. Masum Mir, SVP and GM of Cisco's Provider Mobility business, drew the contrast with the industry's traditional approach of building bespoke solutions for individual enterprise verticals. "What we are doing now is actually turning that into a platform — you build it, you integrate these packages of software and scale it without compromising the trust and the security that needs to be delivered by telecom," Mir said.
Cisco's role in the stack spans from bottom to top: mobile packet core, orchestration, intelligent routing and the security layer.
Hakl described Cisco's function as the connective tissue — someone has to make the capability secure and consumable for developers, and that is what Cisco software provides. Nvidia supplies the accelerated compute, ISV blueprints and vision language models. AT&T brings the network, the zero-trust policy enforcement and the end-to-end control plane.
The Cisco Mobility Services Platform underpinning the deployment already supports 293 million mobile IoT subscribers — including 130 million connected vehicles — across more than 31,000 enterprises worldwide, according to Cisco.
The partnership is also explicitly designed to open the aperture to a developer ecosystem rather than bundle everything into a first-party stack. Hakl said his goal coming out of GTC is to identify four or five more software providers — beyond Linker Vision — that can build on the platform's combination of deterministic networking, edge compute and zero-trust security.
Pushing AI applications deeper into the real world
AI Grid is part of an overall push at the Nvidia conference to bring AI applications deeper into the real world.
Akamai announced it is expanding its Akamai Inference Cloud across more than 4,400 edge locations with thousands of Nvidia RTX PRO 6000 Blackwell Server Edition GPUs, operationalizing the Nvidia AI Grid reference design at global scale. Its intelligent orchestration platform matches each inference request to the right compute tier, optimizing what Akamai calls "tokenomics" — balancing cost per token, time-to-first-token and throughput across its distributed footprint.
Spectrum (Charter Communications) is deploying Nvidia AI Grid for GPU-accelerated edge workloads, using Blackwell GPUs to enable low-latency remote rendering for CGI production studios across its fiber network. T-Mobile and Nvidia are also pushing physical AI to the network edge.
According to Nvidia's GTC announcement, telcos and distributed cloud providers collectively operate around 100,000 distributed network data centers worldwide — from regional hubs to mobile switching offices and central offices — with enough spare capacity to offer more than 100 gigawatts of new AI capacity over time. AI grids turn that existing real estate, power and connectivity into a geographically distributed compute platform.
For AT&T, the argument is that network ownership — not cloud adjacency — is the defensible differentiator in the edge AI era. Whether enterprise buyers agree at scale will determine how quickly this architecture moves beyond its early pilots.