Table of Contents
To run AI at a remote site with poor connectivity, put the model, the document index and the data it needs on a server at the site, so that answers never depend on the link. Use the link only for what can wait: sensor data for head office, logs and updates, sent in batches that survive interruptions.
Key takeaways
- Satellite links add delay. One large measurement study found a median access latency of 673.5 ms for geostationary (GEO) operators and 56 ms for Starlink (Raman et al., 2023).
- Questions and answers are small. Model files are not: a 70-billion-parameter model is about 140 GB at 16-bit precision (Hugging Face), roughly 31 hours over a 10 Mbit/s link at full rate.
- Design the site to work alone: local inference, local login, local logs. The link becomes a synchronisation channel, not a dependency.
- Store-and-forward is already part of industrial messaging. MQTT 5 can keep session state across disconnections (OASIS), and Sparkplug flags data sent late as historical (Eclipse Sparkplug 3.0).
- Power and space shape the hardware. An NVIDIA L4 card draws at most 72 W (NVIDIA); an H100 SXM up to 700 W (NVIDIA).
Why should AI run locally at a remote site?
Oil and gas fields in the Sahara, mines, pumping stations and pipeline facilities often depend on satellite, microwave or patchy cellular links that already carry SCADA traffic, voice and business applications. A cloud AI service adds three problems: every question and every agent step crosses the link and back; when the link drops, the assistant stops, often during the incident when people most need the procedures; and bulk AI traffic competes with operational traffic.
Data sent off site also leaves your control; our guide to sovereign, on-premise AI covers that side. This article covers the design of a site that keeps working on its own.
How slow are satellite links for AI?
Latency depends mainly on the orbit. A geostationary satellite sits at 35,786 km, while low Earth orbit is below 2,000 km (ESA). At the speed of light, each leg to a geostationary satellite takes about 0.12 seconds, and a request and its reply cross four legs: about half a second before any processing.
A CoNEXT 2023 study of satellite network operators, based on public speed-test data, measured these median access latencies (Raman et al., 2023):
| Orbit | Altitude | Median access latency measured |
|---|---|---|
| Geostationary (GEO) | 35,786 km | 673.5 ms overall; 620.4 ms for the best operator, 835.2 ms for the worst |
| Medium Earth orbit (MEO) | Between LEO and GEO | 279.36 ms (O3b) |
| Low Earth orbit (LEO) | Below 2,000 km | 56 ms for Starlink, 154 ms for OneWeb |
The same study found that over 80% of the GEO trace had jitter of 100 ms or more, against less than 20% for LEO. Half a second is tolerable for one question. An AI agent that calls a cloud model ten times in sequence would wait about seven seconds at the GEO median, on the link alone.
LEO services narrow the gap, but they depend on licensing. In Algeria, the regulator ARPCE launched a call for competition on 9 April 2026 for two licences to operate public networks using non-geostationary (NGSO) satellites, open to VSAT licence holders in Algeria and operators with a global constellation (Algérie Eco). Confirm which services are licensed for your site with ARPCE and your counsel. And even a fast link fails sometimes.
What needs the link, and what does not?
Most of what an AI system does can run without the link, and most of what must cross it can wait.
| Traffic | Typical size | Can it wait? | Where it should run or travel |
|---|---|---|---|
| Users' questions and answers | Kilobytes of text | No | On the site server |
| Search over the document index | Index of several gigabytes | No | On the site server |
| Sensor data for head office | Small, continuous | Yes, with buffering | Store-and-forward |
| Logs, usage data, user feedback | Small | Yes | Batched sync |
| New document revisions | Megabytes | Hours | Incremental sync |
| Software patches and container images | Hundreds of megabytes to gigabytes | Days, except urgent fixes | Delta sync or physical media |
| Model weights | Gigabytes to hundreds of gigabytes | Weeks | Physical media or off-peak transfer |
Model files dominate. Using Hugging Face's rule of thumb of about 2 GB per billion parameters at 16-bit precision (Hugging Face), and 1 byte or half a byte per parameter at 8-bit or 4-bit, our arithmetic gives these transfer times:
| Model package | Approximate size | At 2 Mbit/s | At 10 Mbit/s | At 50 Mbit/s |
|---|---|---|---|---|
| 8B parameters, 4-bit | 4 GB | about 4.4 hours | about 53 minutes | about 11 minutes |
| 8B parameters, 16-bit | 16 GB | about 18 hours | about 3.6 hours | about 43 minutes |
| 70B parameters, 4-bit | 35 GB | about 39 hours | about 7.8 hours | about 1.6 hours |
| 70B parameters, 16-bit | 140 GB | about 6.5 days | about 31 hours | about 6.2 hours |
These assume the full nominal rate with no other traffic or overhead; real transfers take longer.
Which hardware fits a remote site?
At a remote site the constraints are power (often from generators), cooling in a shelter, space and the absence of IT staff. Match the hardware to the number of users and the model size your tests require.
| Option | Example | Memory | Power | Typical use |
|---|---|---|---|---|
| Laptop or workstation, CPU only | llama.cpp with a quantized small model | System RAM | Depends on the machine | One user, field kit, fallback |
| Embedded module | NVIDIA Jetson AGX Orin 64GB | 64 GB | 15 to 60 W | Vision at the equipment, small models |
| Low-power server GPU | NVIDIA L4, single-slot low-profile PCIe | 24 GB | Up to 72 W | Small or quantized models, a few users |
| Data-centre GPU | NVIDIA H100 SXM | 80 GB | Up to 700 W | Large models, many users; usually at a central site |
Figures come from NVIDIA's pages for the Jetson Orin, L4 and H100. llama.cpp is a C/C++ inference engine that supports 1.5-bit to 8-bit quantization and can split a model between CPU and GPU.
A smaller model is often the right trade at a remote site. Our article on small language models explains when a 7B model is enough, and our GPU sizing guide covers memory for many users. Check the server's operating temperature range against the shelter conditions, and put it on the same backup power as the systems it supports. For AI next to machines on a plant floor, see edge AI on the factory floor.
How do you keep data flowing when the link drops?
Use store-and-forward: the site keeps data locally while the link is down and sends it when it returns. Industrial messaging standards already support this.
MQTT sessions. In MQTT 5, QoS 1 means messages "are assured to arrive but duplicates can occur". If the Session Expiry Interval is greater than zero, the client and server must keep the session state after the network connection closes (OASIS MQTT 5.0).
Sparkplug edge nodes. The Sparkplug 3.0 specification describes an edge node that stores data at the edge until the primary host application comes back online, then flushes the historical data it stored while offline. A metric flag, is_historical, marks values sent late, for batching, store-and-forward or local backup during a network outage, so they are not treated as current values. Every metric also carries a UTC timestamp of when it was captured (Eclipse Sparkplug 3.0). Our unified namespace and MQTT guide explains the basics.
Apply the same rule to the AI system: log questions, answers, sources and user feedback on site, and sync them in batches. Size the local buffer from the data rate multiplied by the longest outage you expect, plus a margin. The same idea underlies "local-first" software, one of whose ideals is that "the network is optional" (Kleppmann et al., 2019).
How do you update models and software over a thin link?
Download centrally, not at the site. Fetch and check models and images at head office, and send the site only what it needs. Hugging Face's download tool can filter files by pattern and preview a download with --dry-run: in its example, the gpt2 repository has 5.6 GB to download across several formats, while the safetensors weights alone are 548 MB (Hugging Face).
Send only what changed. rsync's delta-transfer algorithm "reduces the amount of data sent over the network". Its --partial option keeps partially transferred files so an interrupted transfer can resume, and --bwlimit caps the transfer rate so updates do not starve operational traffic (rsync). Container images are made of layers, and Docker does not pull a layer that is already present locally (Docker), so a stable base image keeps updates small.
Ship what is too big. For a new large model, a hard disk through the site's controlled transfer process is often faster than the link. Our air-gapped deployment guide covers checking files before they cross.
Test before switching. Install on site next to the current version, run your test questions, and keep the previous version for rollback, as described in keeping an on-premise LLM up to date.
What else breaks when a site is cut off?
Login. If users authenticate against a server at head office, nobody can log in when the link drops. Microsoft designed its read-only domain controller (RODC) for branch offices with poor physical security and poor bandwidth to a hub site. Once a user's credentials are cached on it, it can handle that user's logons locally. By default it caches no user credentials; administrators choose which accounts it may cache, which limits exposure if the server is stolen (Microsoft).
Monitoring and support. The central team cannot see a disconnected site. Keep health checks and alerts local, and give one person on site a short runbook: restart the service, check the buffer, roll back.
Physical security. Unmanned sites need locked racks and encrypted disks, so that a stolen server does not expose your documents.
How do you deploy AI at a remote site, step by step?
- Measure the link: real throughput, latency and outages over several weeks, not the contract figure.
- Sort the traffic using the table above: what runs locally, what can wait, what must be shipped.
- Choose the model size from tests and site power, starting with the smallest serious candidate.
- Build and test at head office with the network unplugged, so hidden downloads and licence checks show up as errors.
- Ship a pre-loaded server with the model, index, images and local user accounts.
- Configure store-and-forward, with buffers sized for the longest expected outage.
- Set the update path: rate-limited delta sync off-peak, physical media for large models.
- Run a disconnection drill: cut the link for a full shift and see what fails.
What we have seen at KABAS AI
DALILE, our platform for industrial documents, runs completely offline on the client's own servers, with no internet access required. Virtual Lab also offers a local PC deployment that runs with no internet connection. Our mining solutions are designed for remote sites without connectivity, and our oil and gas solutions address knowledge scattered across drilling sites and facilities.
Designing for that constraint leads to three rules: decide at design time what the site must do alone, so answering, searching and logging never wait for the link; treat the link as a synchronisation channel, with a plan for what crosses it and when; and test disconnection before go-live, not after the first outage.
If you are planning AI for a remote site, our consulting team can work through the architecture with your IT, OT and telecom teams.
Sources
- Raman et al., Dissecting the Performance of Satellite Network Operators, CoNEXT 2023 (arXiv 2310.15808)
- ESA, Types of orbits
- Algérie Eco, L'ARPCE lance un appel à concurrence pour deux licences de communication par satellites NGSO, 9 April 2026
- Hugging Face, Optimizing LLMs for speed and memory
- Hugging Face, Download files from the Hub
- NVIDIA L4 Tensor Core GPU
- NVIDIA H100 Tensor Core GPU
- NVIDIA Jetson Orin
- llama.cpp on GitHub
- OASIS, MQTT Version 5.0
- Eclipse Foundation, Sparkplug 3.0.0 specification
- Kleppmann et al., Local-first software: you own your data, in spite of the cloud, Onward! 2019
- rsync manual page
- Docker, docker image pull
- Microsoft, AD DS: Read-Only Domain Controllers
Frequently asked questions
Can an AI assistant work at a site with no internet connection?
Yes. Once the model files, the inference server, the document index and their dependencies are on a server at the site, answering a question needs no external network. Some software tries to check for updates or send usage statistics, so each component must be configured for offline use and tested with the link unplugged.
Is a low Earth orbit satellite link good enough to use cloud AI?
It removes most of the latency problem: one large measurement study found a median access latency of 56 ms for Starlink against 673.5 ms for geostationary operators. It does not remove the dependency: when the link drops, a cloud assistant stops. Availability also depends on licensing; in Algeria, ARPCE opened a call for competition for two NGSO satellite licences in April 2026, so check the current status for your site.
How much bandwidth does a remote AI system need?
Very little for questions and answers, which are short text. The large transfers are model files and software updates. A 70-billion-parameter model is about 140 GB at 16-bit precision, which would take about 31 hours over a 10 Mbit/s link at full rate, so plan model updates as scheduled transfers or physical shipments.
What hardware suits a small remote site?
It depends on the number of users and the model size your tests require. Low-power options exist: NVIDIA lists its L4 data-centre card at 24 GB of memory and a maximum of 72 W, and its Jetson AGX Orin 64GB module at 15 to 60 W. Check power, cooling and backup supply at the site before choosing.
What happens to data collected while the link is down?
With store-and-forward, the edge system keeps it locally and sends it when the link returns. MQTT 5 can keep session state across disconnections, and the Sparkplug specification lets an edge node store data while the host application is offline and flag it as historical when it is sent later. Size the local buffer for your longest expected outage.



