For the past several years, I've had a front-row seat to a problem that nobody in the AI industry is talking about: the operators connecting rural America are drowning in operational complexity, and every AI tool being built for them has the same fatal flaw.
I spent years at ETI Software Solutions building the automation infrastructure that keeps Tier 2/3 ISPs compliant — FCC Broadband Data Collection tools, broadband label generation, the kind of unglamorous infrastructure work that doesn't get keynotes but keeps operators out of regulatory trouble. What I saw during that time was consistent: these operators were capable, resourceful, and chronically underserved by the software industry.
The Problem Nobody Names Correctly
When people talk about AI for broadband operators, they talk about chatbots. Customer service automation. Maybe some predictive maintenance analytics. The conversation stays shallow because most of the people building these tools have never spent time inside a rural ISP's NOC at 2 AM when an OLT goes down and the on-call engineer is 90 minutes from the nearest fiber splice point.
The real problem isn't a lack of AI capability. The real problem is that the AI tools being built for this market carry a hidden dependency: cloud infrastructure.
What ISP operators actually need is AI that works when the WAN connection is degraded, that processes subscriber data without routing it through a third-party data center, and that integrates with the specific OSS/BSS stack they've been running for a decade. None of that is a chatbot. All of it is an architectural decision.
What We Built — and What We Learned
XSI LodeStone is the answer we built after years of watching the problem from the inside. It's a sovereign agentic AI appliance — NVIDIA GB10 Grace Blackwell Superchip hardware, running the XSI LodeStone Orchestrator, entirely on-premises. No cloud dependency for inference. No subscriber data leaving the building without explicit operator authorization.
The lessons we learned building it are lessons the industry needs to hear:
Lesson 1: The FCC BDC is not a form. It's a data architecture problem.
Every ISP receiving BEAD funding has a continuous BDC filing obligation. The data that goes into those filings lives across network management systems, GIS platforms, service activation records, and customer databases. Getting it right requires reconciliation across sources that weren't designed to talk to each other.
An AI agent that can read across those systems, identify discrepancies, and generate compliant BDC submissions is not a nice-to-have. For operators under BEAD subgrant agreements, it is a compliance survival tool. We built one. It works. The gap between "it runs in staging" and "it handles a real submission cycle without human review" was six months of edge-case work that no vendor had bothered to do.
Lesson 2: Sovereign doesn't mean isolated.
One of the early misconceptions we had to design past: "on-premises AI" sounds like a regression to IT teams who've spent the last decade moving everything to the cloud. The pitch we had to learn to make wasn't "leave the cloud" — it was "run inference locally, integrate everywhere."
XSI LodeStone agents integrate with Google Workspace, Microsoft 365, existing OSS/BSS stacks. The data doesn't leave the building. The workflow integrations do. That distinction matters enormously to operators who are already invested in cloud collaboration tools but cannot afford for their network data to live in a third-party data center.
Lesson 3: The GB10 changes the math.
The NVIDIA GB10 Grace Blackwell Superchip — 1 petaflop of AI compute in a desktop-footprint appliance, 128GB unified memory — is the hardware event that makes this category possible. Before GB10, sovereign agentic AI at Tier 2/3 price points wasn't achievable. The inference hardware was either too expensive or too underpowered to run production multi-agent workflows.
XSI is a member of the NVIDIA Inception Program. Our appliance is built on GB10. While the rest of the industry chases cloud-scale agentic AI for hyperscalers, we've designed our stack for the operators that conversation tends to leave out — regional ISPs, regulated SMBs, and rural infrastructure businesses where sovereignty isn't optional.
What This Means for Tier 2/3 Operators Right Now
The BEAD deployment window is closing. Operators who are still manually managing FCC compliance workflows, still running reactive NOC operations, still doing network topology reconciliation by spreadsheet — they are accumulating operational debt that compounds with every subgrant dollar they accept.
The AI infrastructure layer is ready. The question is whether the right application will be in their hands before the window closes.
We ship Q3 2026. If you're building rural broadband — or investing in the operators who are — I'd like to talk.
Q&A with Rhyan
Extended questions from the LinkedIn discussion — answered in full.
Data processing agreements cover liability. They don't cover network latency, WAN reliability, or what happens when the AI system the operator depends on for fault resolution is itself dependent on an external data center that's experiencing its own outage.
Sovereign infrastructure for ISPs isn't a privacy position — it's an operational continuity position. A rural ISP's AI agent needs to keep running when the upstream internet connection is degraded. That's a structural impossibility for a cloud-dependent system, regardless of what the DPA says.
The regulatory question is separate and also real: BEAD subgrant cybersecurity requirements, FCC confidentiality obligations on BDC data, and an increasing number of state-level data localization requirements for telecom infrastructure. We can provide the full framework analysis in the whitepaper.
We built a base integration adapter architecture specifically for this kind of heterogeneous stack. The platform ships with pre-built connectors for Calix Cloud EMS, Nokia Altiplano, and standard OSS interfaces including ACS/TR-069 and TR-369. RADIUS integration is first-class.
For legacy BSS systems, we use a BaseLOBAdapter pattern that supports BEARER and SESSION_RPC authentication — the two most common auth patterns in older telecom BSS stacks. In practice, we've found that most Tier 2/3 BSS systems can be connected in under a week using these adapters.
Custom provisioning systems take longer, but the architecture supports it. This is where the hands-on onboarding process matters — we want to know your stack before you sign anything.
The core team is small and deliberately chosen. I spent years at ETI Software building the FCC compliance and ISP operations tools that are now in production at Tier 2/3 operators across the US. Our CTO Georgii Galych has a decade of production Kubernetes and cloud infrastructure engineering across distributed systems at scale. Our CIO Fakhri Huseynov runs our cloud/DevOps infrastructure from Riga — Azure, AWS, Terraform, deep production experience.
XSI is a member of the NVIDIA Inception Program, which gives us direct access to NVIDIA's developer resources and early access to GB10 hardware. The ambitious part isn't the ambition — we've all built production systems at this complexity level before. The breakthrough is that the hardware economics finally make it viable for operators who couldn't previously afford it.
Pricing details are available through the access request process — we're pricing per-appliance with a platform subscription that includes agent updates, integration maintenance, and support. The design philosophy is that it should be significantly cheaper than staffing an AI engineering team and dramatically more capable than any cloud subscription tool you'd deploy in its place.
On the "build vs. buy" question: building a production agentic AI platform on GB10 hardware that integrates with your specific OSS/BSS stack, maintains FCC compliance accuracy, and handles multi-agent fault resolution would take a specialized team two to three years. We have that team. We've done that work. The time-to-value difference is why operators are talking to us.
Common Questions About Sovereign AI for ISPs
Search-ready answers to the questions we hear most often.
Sovereign AI refers to AI systems where inference runs entirely on the operator's own hardware, under the operator's policy controls, with no dependency on external cloud infrastructure for core functionality. For broadband operators, this means AI agents that process network data, subscriber records, and compliance information without that data leaving the operator's premises.
For ISPs specifically, sovereign AI matters because: (1) FCC-regulated data has confidentiality obligations that complicate third-party cloud processing; (2) rural network operations require AI that functions when WAN connectivity is degraded; and (3) BEAD subgrant cybersecurity requirements create compliance risk around cloud-based AI deployments that operators haven't fully assessed.
The FCC Broadband Data Collection program requires ISPs to file detailed coverage and availability data twice annually, with ongoing obligations tied to BEAD funding agreements. The data must reflect actual served locations, technology types, and speeds — and the FCC has significantly increased its verification scrutiny since 2022.
BDC automation involves AI agents that (1) read across network management systems, GIS platforms, and service activation records; (2) identify and reconcile discrepancies between these sources; (3) generate compliant BDC submission files in the FCC-required format; and (4) flag edge cases for human review before submission. The goal is to replace a process that currently takes many operators dozens of staff-hours per filing cycle with an automated workflow that runs in hours with minimal human oversight.
The XSI LodeStone Orchestrator is the agentic AI coordination layer we built for the LodeStone appliance. It routes work between domain-specific agents, enforces Shadow Mode and four-eyes approvals on anything that touches live equipment, and runs entirely on-premises on the NVIDIA GB10 Grace Blackwell Superchip.
For ISPs, the significance is the combination: the GB10 Superchip delivers roughly one petaflop of AI compute in a desktop form factor with 128GB of unified memory, and the LodeStone Orchestrator turns that compute into an auditable multi-agent workflow engine tuned for Tier 2/3 network operations. That combination makes production-grade agentic AI economically and operationally viable for regional operators for the first time — compute and orchestration both sit in the building, governed by the operator, not by a third-party API.
The Broadband Equity, Access, and Deployment (BEAD) program is the $42.45 billion federal initiative to expand high-speed broadband to unserved and underserved areas. Administered by NTIA and distributed through state broadband offices, BEAD subgrants come with substantial compliance obligations that continue for the life of the deployment.
The AI compliance implications are underappreciated: BEAD subgrant agreements incorporate NTIA cybersecurity guidance that may constrain where and how operators process network and subscriber data. Operators deploying AI tools that route operational data through third-party cloud infrastructure should conduct a formal assessment of whether those deployments satisfy BEAD compliance requirements. Sovereign, on-premises AI deployment eliminates this risk category entirely.
Three structural differences: First, XSI LodeStone is an appliance, not a SaaS subscription — it ships as pre-configured hardware with the AI platform pre-loaded, requiring no AI engineering skills to deploy. Second, inference is fully sovereign — it runs on-premises on the GB10 hardware, not in a cloud data center, which matters for both compliance and operational resilience. Third, the platform was designed specifically for Tier 2/3 ISP operational realities, including pre-built integrations for Calix Cloud, Nokia Altiplano, FCC BDC workflows, and common Tier 2/3 BSS stacks.
Most AI tools being sold to ISPs are general-purpose SaaS products with ISP use cases bolted on. XSI LodeStone is the reverse: built for ISP operations from the ground up, deployed as sovereign hardware infrastructure.
XSI LodeStone is targeting a Q3 2026 general availability launch. We are currently accepting access requests from Tier 2/3 ISPs and broadband investors through the form on xsilodestone.ai. Early access participants will work directly with our team on integration and onboarding ahead of GA launch.
You can also download the full data sovereignty whitepaper — "The FCC BDC Is Not a Form: It's a Data Architecture Problem" — by entering your information in the sidebar form. The whitepaper covers the full regulatory framework, GB10 economics, and operational architecture in detail.