A few months ago, I was on a call with a robotics CTO — sharp team, serious funding, impressive demo. They'd built a warehouse inspection system on Jetson Orin that could identify structural anomalies human inspectors miss. Real product. Real differentiation. And they asked me a question I've now heard from dozens of Jetson teams:
"Can we just… never update the software?"
They weren't being lazy. They were terrified. Their last OTA update had bricked three devices. Each truck roll cost them $8k. They'd burned six weeks debugging a GStreamer conflict that surfaced only on devices running a specific JetPack point release. Their ML engineers (people hired to improve perception models) were instead maintaining Yocto meta-layers and answering questions about secure boot. And they were exhausted.
They'd done everything right. They chose NVIDIA. They prototyped on JetPack. The demo worked. But somewhere around their sixtieth deployed device, building the product stopped being the hard part.
Every team building physical AI on Jetson eventually crosses the same threshold. My co-founder Justin calls it the moment where "a bunch of data scientists get tricked into becoming embedded developers." Nobody plans for it.
JetPack gets you from zero to a working prototype faster than anything else. CUDA, TensorRT, DeepStream, Isaac — the ecosystem is unmatched. NVIDIA is four years ahead of the competition on edge AI, and it's not close. If you're training in the cloud on NVIDIA, deploying on Jetson is the obvious move. That's not changing.
What trips teams up is that NVIDIA made an intentional design choice with JetPack: optimize for the widest possible developer on-ramp. That means a mutable, general-purpose Linux distribution (a.k.a. NVIDIA-flavored Ubuntu). It's the right call for adoption. It gets millions of developers building. But a mutable, general-purpose OS is fundamentally the wrong foundation for fleet-scale production. That's not a flaw in JetPack. It's a design boundary. NVIDIA built a development kit, and it's the best one in the market. The problem is when teams try to ship it as a production OS.
We talk to teams every week who've discovered this the hard way. The pattern is consistent: the prototype works, the rollout gets funded, and then somewhere after 25 devices, infrastructure work that nobody scoped (or budgeted for) consumes the roadmap for twelve to eighteen months.
What breaks is predictable enough that I can usually tell a team their own story back to them.
Your ML team was hired to improve model accuracy and build better perception. Instead, half their cycles go to maintaining meta-layers, debugging CUDA dependency conflicts across JetPack versions, managing ROS containers, and fielding questions about secure boot. A robotics CTO once told me he'd stopped tracking what percentage of his team's sprints went to "infrastructure" because the number was demoralizing. The people you hired to build your competitive advantage are keeping the lights on.
JetPack's A/B partition scheme works on the bench. At scale, one bad update can cascade. There's no reliable migration path between major JetPack versions — 5 to 6 to 7 — and there are tens of thousands of Jetson devices in the field right now stuck on JetPack 4 with no path forward. The instinct to freeze everything is understandable. But with over 3,200 Linux CVEs surfacing every year and the Cyber Resilience Act raising the compliance floor across Europe, standing still isn't a strategy. It's a countdown.
Your buyer sends over a two-hundred-question security questionnaire. Reproducible builds? SBOM generation? Disk encryption key management? CVE patching cadence? JetPack plus Docker can't answer those questions. The deal stalls; not because your product isn't good enough, but because your infrastructure can't prove it's secure enough. If you lived through the Puppet and Chef era — SSH-ing into a hundred servers hoping they'd converge on the same state — this is that same problem, reborn on edge hardware.
The market already knows this. NVIDIA's own Edge AI team has shared that nearly 45% of Jetson customers have moved to Yocto-based systems for production. The shift isn't a theory, it's happening. The question is whether you build and maintain your own production OS (ie. hiring a Yocto team, integrating the BSP, owning the toolchain and CVE pipeline forever)or whether you use one that's already been hardened, validated, and deployed across hundreds of production sites.
This is the problem we started Peridio to solve. Avocado OS gives you a production-grade, immutable embedded Linux built for Jetson with all the hardware acceleration you depend on, none of the maintenance burden. Peridio Core handles fleet management: OTA with atomic rollback, deployment scheduling, fleet-wide visibility. If power fails mid-update, the device boots to the last known-good image. No bricks. No truck rolls.
The architecture is straightforward: we separate your OS from your application. We manage the OS layer: locked down, patched, compliant. Your team manages the application layer — models, ROS packages, configs — and ships updates independently, on their schedule, without touching the firmware underneath. Your engineers go back to the work they were hired to do.
We're an NVIDIA Partner Network Solution Advisor in Embedded Compute. We work directly inside the ecosystem with NVIDIA and with the leading Jetson hardware integrators. This isn't a bolt-on. It's purpose-built for this platform.
2026 is the year Jetson deployments scale from dozens to hundreds to thousands. The teams treating their OS strategy as a problem for later are building a migration crisis they'll have to unwind under pressure. Every month of technical debt compounds.
You didn't start your company to maintain Linux infrastructure. You started it to build something that works in the physical world. We exist so you can get back to that.
Our Co-Founder & CTO Justin Schneck is speaking at Embedded Vision Summit on this exact challenge — scaling computer vision deployments on Jetson from prototype to production. If you're in the Jetson ecosystem, come find us.
Or just talk to us.