My cameras don’t report to anyone
When I built the security system for my home, I set it up with seven 4K cameras processed right there with Frigate, an open-source video recording software. Car, person and face detection runs on a local accelerator. I’m now replicating that same system in my new house.
The alerts are custom. If a car enters the driveway area at night, the garage light turns on for five minutes; the rule is limited to that zone to avoid false positives. No private video leaves the house, there’s no subscription and no third-party server.
The clearest contrast is Ring, Amazon’s home camera brand, which processes video in the cloud. Two recent events show what that means:
- Availability. On October 20, 2025, AWS suffered an outage in its US-EAST-1 region that lasted more than 15 hours (ThousandEyes, 2025). Ring was among the affected services, along with Snapchat and Venmo (Tom’s Guide, 2025). A local system kept recording and sending alerts.
- Privacy. In September 2025, Ring launched Search Party, an AI feature that connects a neighborhood’s cameras to look for lost pets and comes enabled by default (Biometric Update, 2026). The 2026 Super Bowl ad promoting it sparked a massive backlash over surveillance fears, and that same week the brand canceled its integration with Flock Safety, a camera network used by law enforcement (Fortune / AP, 2026). Two months earlier, it had launched Familiar Faces, a facial recognition feature for cataloging people who visit the home (TechCrunch, 2026).
My system does exactly the same thing as Ring. The difference lies in where the video stays and who decides what happens to it.
It fails too, of course. Once, the disk filled up and detection events stopped closing. With local, the problem is yours, but so is the fix: you don’t wait 15 hours for a provider to solve it.
If it’s so good, why weren’t we doing it before?
Because until two years ago, running your own model was a technical curiosity with mediocre results. That changed fast:
- Since January 2026, the best open-weight models have been, on average, about four months behind the most advanced closed models (Epoch AI, 2026).
- With a single high-end gaming GPU, a graphics card with 8 or 12 GB, you can already run models locally that match the frontier of 6 to 12 months ago (Epoch AI, 2025).
There’s an important nuance: the open model that truly catches up with the frontier usually has hundreds of billions of parameters and needs a server, not a laptop (Modelfit, 2026).
For business, that’s more than enough. Removing backgrounds, classifying, summarizing, detecting objects or generating variants doesn’t require the most powerful model in the world. A very good one running on your own equipment does the job.
What your company gains with local AI
Three benefits support the decision to bring an AI workload in-house: data control, operational continuity and predictable cost.
1. Your data stops traveling to places you don’t know
Leaks often start with a well-meaning employee who pastes confidential information into a public tool because it solves the problem faster. According to IBM and the Ponemon Institute’s annual Cost of a Data Breach report, a high level of shadow AI (employees using unapproved AI tools) added USD 670,000 to the average cost of a breach, and 63% of the 600 organizations surveyed had no policies to govern AI use (IBM, 2025).
Banning it rarely works. What does reduce the incentive is giving the team an internal AI that is just as useful, running on your own infrastructure, where the contract, the customer database or the code never leave the company.
Protecting customer data is also a commercial advantage. Among the companies surveyed by Cisco, 94% say their customers wouldn’t buy from them if they didn’t protect their data properly (Cisco, via Govtech Review, 2024), and 95% say the benefits of investing in privacy outweigh the costs, with an average return of 1.6 times the investment (Cisco, via Outlook Startup, 2024). For a brand, being able to explain with certainty where its customers’ information is processed translates into trust, less friction with its own clients’ legal and procurement teams, and a concrete argument against the competition.
2. Your operation keeps running even if the provider goes down
The October 2025 AWS outage had the greatest impact of the 38 that IncidentHub, an independent incident aggregator, recorded that year, and it affected hundreds of SaaS providers at once (IncidentHub, 2025).
Outages are the most visible dependency, but there are others: usage limits, rising prices, models the provider changes or discontinues overnight. In our previous article, we saw how a measurement provider could change its methodology and force an entire industry to recalibrate. The same thing happens with AI models. A local model keeps working the same way for as long as you decide.
3. Your cost stops growing with every use
A Cloudian survey of 203 IT decision-makers found that 93% of companies have already moved AI workloads off the public cloud, are doing so or are evaluating it, and almost half point to unpredictable cloud costs as a barrier (Cloudian, via StorageNewsletter, 2026). The figure should be read with care: Cloudian sells on-premises infrastructure.
The rule of thumb is simple. If the workload runs all day, every day, your own hardware pays for itself. If it’s sporadic, the cloud still wins.
Speed shows up as an extra benefit in some cases: a camera alert, a kiosk or an event experience needs an immediate response, and without connectivity the cloud simply doesn’t exist. It shouldn’t be sold as a universal advantage, though: a small local model responds fast, and a large cloud model still reasons better.
