I gave a talk at DevConf.CZ 2026 in Brno, arguing that Podman Quadlets are a better default deployment unit than Kubernetes for small, fixed industrial edge devices. The argument held up. But the more interesting question was why the broader ecosystem schedule looks so much like the stack I care about?
By the end of day two I had an answer: the schedule was not five separate trend clusters — it was one converging argument.
The Four-Stage Pipeline#
The programme collapsed into four dominant topic clusters: agentic and local AI, software supply-chain security and EU CRA compliance, image-mode Linux with Podman and Quadlets, and observability with OpenTelemetry. Each had its own speakers and vocabulary; put them in order and you get a pipeline.
Stage 1 — Produce. AI agents are producing and packaging software, and the distribution substrate has converged on OCI. The Ramalama project ships local AI models as OCI containers using the same ORAS-based registry infrastructure as any application workload. MCP and A2A are treated as open protocols — the Unix-philosophy pattern of small, composable interfaces applied to AI tooling. An edge platform that already speaks OCI gets AI workloads largely for free.
Stage 2 — Prove. The EU Cyber Resilience Act is no longer a future item. A full programme track covered it: SBOMs, artifact signing with Sigstore and PGP, confidential containers, admission-time verification. The signal was precise: mandatory vulnerability-reporting obligations arrive September 2026; full compliance December 2027. The industry is past the debate — it is building the tooling to make signed-by-default operationally viable. Unsigned artifacts are treated, increasingly, as admission failures rather than policy gaps.
Stage 3 — Ship. The Quadlets story fits here. A well-attended talk — Declarative Podman Setup? Just use Quadlets — confirmed that the “OS as a versioned, signed container image; workloads as systemd-managed Quadlets” model is mainstream practice, not a Fedora-specific experiment. CI/CD patterns long associated with application delivery now apply to the base OS itself via bootc and image-mode Linux. On an 8–10 year device lifecycle, betting on a deployment unit the broader Red Hat and Fedora ecosystem is standardising on materially lowers platform risk. One clarification: the community uses “Quadlets” generically for systemd-native declarative container management. The Quadlet Margo deployment profile — the subject of my own talk and the Intro to Podman Quadlets series — is a proposed, not-yet-TWG-ratified contribution. Ecosystem validation and the standards-body proposal are not the same claim.
Stage 4 — Observe. OpenTelemetry has won the instrumentation debate — the live conversation at DevConf.CZ was about cost, cardinality, and overhead, not whether to adopt OTEL. Mimir for storage, eBPF for cheaper collection on constrained hardware. The honest caveat for the industrial edge: on 256 MB-class hardware, the OTEL Collector — a Go binary ranging from roughly 40 MB for a minimal build to well over 100 MB for the full contrib release — is non-trivial overhead that must be budgeted per device. The eBPF/bpftrace track points at a lighter collection path for the most constrained device tier.
Trend 5: The Connective Tissue#
The fifth cluster — open-source sustainability, governance, and developer communication — was fewer sessions but woven through the whole programme. As AI accelerates code production, the bottleneck moves to human review, governance, and maintainability. For any platform whose differentiation rests on shaping an open standard, governance and community health are not soft concerns — they determine whether the pipeline runs in five years or collapses under maintainer churn.
The Margo community learned this directly — and the DevConf.CZ governance talks reinforced it. Participatory upstream engagement is the posture that earns standards-body credibility over time.
The Flywheel#
The four-stage pipeline — Produce → Prove → Ship → Observe — is not a waterfall. Telemetry from the Observe stage feeds back into the next Produce cycle: what you observe about workload behaviour in production informs the next build, what SBOMs track, how the next image is composed.
These trends aren’t five stories — they’re one pipeline, and the next step is a loop that feeds itself. This is the architectural shape of the next generation of industrial systems — the pattern I’m watching most closely.
The loop is bounded to what OpenTelemetry signals can carry: runtime performance, resource utilisation, failure modes. The goal is not autonomous self-modification — it is giving platform maintainers the observability to make the next build better than the last.
That converged stack — OCI-native, signed-by-default, immutably shipped, continuously observed — is the substrate that physical AI (AI inference workloads running on physical industrial edge devices, not the embodied-robotics sense the term sometimes carries) will run on. Not in theory — as an architectural consequence of decisions the ecosystem is making right now.
The Close: Sovereignty Through Convergence#
The industrial edge is one domain where this plays out — not the only one. Sovereign infrastructure — operating across air-gapped environments, under export controls, on hardware whose lifecycle outlasts any single cloud provider’s service availability — faces the same four-stage problem regardless of the specific domain.
The Triangle of Compromises argument is a specific instance of a general choice: when you cannot depend on a cloud orchestrator, the deployment substrate must be something the OS manages natively. Quadlets give you that. The Declarative Without the Elasticity Tax post works through why the tradeoff exists and when it matters.
What DevConf.CZ showed is that the choice is no longer niche. The ecosystem is converging on OCI-native, signed-by-default, systemd-managed primitives — not because any single vendor pushed it, but because four independent pressures (AI distribution, EU regulation, OS immutability, observability standardisation) all point at the same stack.
The future is not a single component — it is solutions that deliver value through OT/IT convergence, built on open primitives that no single vendor can pull from under you. That is what the Margo standard is trying to formalize. And why an agentic-AI architect was paying close attention to an edge and OS conference in Brno.
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.