Designing, Embedding & Orchestrating Physical AI | Stockholm MLOps Event #34 | May 21 2026

Insights from the Stockholm MLOps community

AI is rapidly moving beyond chat interfaces and into physical systems.

At Stockholm MLOps #34, the discussion focused on what happens when AI leaves the browser and starts operating inside:

  • devices

  • vehicles

  • robots

  • factories

  • embedded systems

  • and constrained edge environments

The strongest signal from the event was clear:

Physical AI is not primarily a model problem.

It is an orchestration, runtime, infrastructure, and distributed systems problem.

Across talks from EMQ Technologies, Encube Technologies, and Embedl, the discussions repeatedly focused on:

  • communication infrastructure

  • orchestration layers

  • edge deployment realities

  • manufacturability

  • runtime reliability

  • distributed coordination

  • hardware-aware optimization

  • and operational AI in physical environments

This was not frontier-model theater.

It was infrastructure teams discussing what breaks when AI meets reality.

Summary — Key Insights from the Event

  • Physical AI is becoming a distributed systems problem

  • MQTT is emerging as a coordination layer for AI agents

  • Edge AI increasingly fails at deployment, not training

  • Hardware constraints reshape AI architectures

  • Manufacturability is becoming part of the AI feedback loop

  • Runtime visibility matters more than model demos

  • Physical AI requires orchestration infrastructure

  • Compliance and traceability are becoming operational layers

Key Insights from the Event

Physical AI Is Becoming A Distributed Systems Problem

“Physical AI becomes a distributed systems problem very quickly.” — Ivan Dyachkov, EMQ Technologies

One of the strongest themes throughout the evening was that once AI systems coordinate fleets of devices, the challenge rapidly shifts away from inference quality.

The real problems become:

  • orchestration

  • routing

  • synchronization

  • discovery

  • state coordination

  • reliability

  • recovery

  • and communication infrastructure

Ivan Dyachkov repeatedly framed AI coordination using classic distributed systems concepts:

  • brokers

  • routing

  • fanout

  • load balancing

  • and presence

The signal was clear:

AI engineering is increasingly converging with distributed systems engineering.

MQTT Is Emerging As An AI Coordination Layer

“Most agent frameworks rebuild things MQTT already solved decades ago.” — Ivan Dyachkov

The event strongly positioned MQTT as more than IoT infrastructure.

It was repeatedly framed as:

an orchestration substrate

  • a coordination layer

  • a communication fabric

  • and runtime infrastructure for physical AI systems

Instead of building direct mesh coordination between hundreds of agents and devices, MQTT brokers provide:

  • discovery

  • routing

  • fanout

  • workload distribution

  • and coordination

The talks repeatedly suggested that communication infrastructure itself is becoming part of the AI stack.

Physical AI Shifts The Problem From Inference To Orchestration

“When AI enters physical systems, orchestration becomes more important than inference.” — Ivan Dyachkov

One of the clearest operational insights from the event was that once AI controls physical systems, runtime coordination matters more than isolated model quality.

The discussions repeatedly focused on:

  • coordinating robotic systems

  • orchestrating device fleets

  • handling runtime state

  • distributing workloads

  • and managing real-world reliability

The challenge is no longer simply generating intelligence.

It is operating intelligence reliably across real-world environments.

Edge AI Breaks At Deployment Time

“Late-stage deployment failures are extremely common.” — Andreas Ask, Embedl

“You realize at the end that operations are unsupported or the model doesn’t compile.” — Andreas Ask

The Embedl discussion exposed one of the strongest operational realities of edge AI:

Many AI systems fail after development, during deployment.

The recurring issues included:

  • unsupported runtime operations

  • quantization failures

  • compiler incompatibilities

  • runtime mismatches

  • driver regressions

  • and fragmented hardware ecosystems

The strongest signal:

Deployment complexity increasingly rivals model development complexity itself.

Hardware Constraints Reshape AI Architecture

“Physical AI is constrained by hardware.” — Andreas Ask

The event repeatedly emphasized that physical AI systems operate under:

  • latency ceilings

  • power constraints

  • thermal limitations

  • memory limits

  • and hardware-specific runtime restrictions

This forces AI systems to become:

  • hardware-aware

  • runtime-aware

  • compiler-aware

  • and deployment-aware

The operational stack increasingly depends on:

  • quantization

  • runtime optimization

  • hardware profiling

  • and architecture-specific deployment strategies

Manufacturability Is Becoming An AI Feedback Problem

“There’s a difference between designing something and building something that works in reality.” — Yuri Gusev, Encube Technologies

“The goal is not text-to-CAD that looks good. The goal is manufacturable design.” — Yuri Gusev

The Encube discussion highlighted a major shift in AI-assisted hardware engineering:

Generative design is not useful unless validation loops exist inside the generation process itself.

This includes:

  • manufacturability analysis

  • stress testing

  • tolerance validation

  • and iterative engineering review

The event repeatedly emphasized that:
“looks plausible” is not operationally meaningful in physical systems.

Reality becomes the validator.

Runtime Visibility Matters More Than Model Quality

“The hardware behavior needs to be visible already in the PyTorch stage.” — Andreas Ask

A recurring operational signal throughout the event was that organizations increasingly lack visibility into:

  • runtime behavior

  • compiler transformations

  • deployment lineage

  • edge execution

  • and hardware compatibility

The discussions repeatedly emphasized that runtime systems fundamentally reshape model behavior during deployment.

That means observability and visibility increasingly become part of operational AI infrastructure itself.

Humans Remain Inside The Loop

“Humans are still expected to stay in the loop.” — Yuri Gusev

Despite the heavy focus on agents, orchestration, and automation, none of the speakers advocated for fully autonomous systems.

Instead, the event repeatedly emphasized:

  • human validation

  • engineering review

  • revision comparison

  • manufacturability inspection

  • and operational oversight

The strongest signal was that physical AI systems require:
structured collaboration between humans, infrastructure, and AI systems.

Compliance And Traceability Are Becoming Infrastructure

“The engineering evidence itself becomes operational infrastructure.” — Andreas Ask

One of the strongest operational themes was that regulated AI systems increasingly require:

  • lineage tracking

  • reproducibility

  • audit readiness

  • deployment traceability

  • and evidence generation

The discussions repeatedly highlighted how:

  • MLflow results become fragmented

  • evidence collection remains manual

  • and deployment coordination spans multiple organizational teams

The implication was clear:

Compliance itself is increasingly becoming an operational infrastructure layer.

Operational Signals Emerging From Meetup #34

Several larger operational shifts emerged across the event:

  • AI infrastructure is moving into physical environments

  • orchestration is becoming a foundational AI layer

  • communication infrastructure is becoming strategically important

  • edge deployment complexity rivals model development complexity

  • runtime visibility is becoming operationally critical

  • hardware-aware optimization is becoming mandatory

  • manufacturability validation must exist inside AI generation loops

  • distributed systems engineering is becoming core AI engineering

  • compliance and traceability are moving into runtime infrastructureThe strongest overall signal:

Physical AI is evolving into an infrastructure discipline.

Not just a model discipline.

📊 Insights by Stockholm MLOps

Signals From The Room

Throughout the event, attendees shared live operational perspectives, infrastructure priorities, and ecosystem insights through interactive audience discussions and live polling during the meetup.

The signals below reflect recurring themes, operational concerns, and infrastructure priorities emerging directly from practitioners, founders, engineers, architects, and operators attending Stockholm MLOps.

Physical AI Is Increasingly Viewed As An Infrastructure Problem

The strongest audience reactions consistently centered around:

  • orchestration

  • runtime coordination

  • edge deployment

  • hardware constraints

  • observability

  • and distributed systems complexity

The operational signal from the room was clear:

As AI moves into physical systems, infrastructure and orchestration increasingly matter more than standalone model performance.

AI Infrastructure Continues To Dominate Community Interest

Recurring audience discussions repeatedly focused on:

  • AI Infrastructure

  • Agentic Systems

  • AI Security

  • Runtime Governance

  • Evaluation & Observability

  • Inference Optimization

This strongly reinforces the infrastructure-heavy and operationally mature profile of the Stockholm MLOps ecosystem.

The Community Continues To Shift Toward Operational AI

One of the clearest recurring patterns across the evening:

The conversation is steadily moving away from:

  • model demos

  • prompting

  • standalone copilots

and toward:

  • operational AI systems

  • orchestration

  • governance

  • deployment reliability

  • runtime infrastructure

  • and infrastructure ownership

What This Event Signals

Stockholm MLOps #34 strongly reflected the next operational phase of AI systems.

The industry focus is shifting from:
“Can AI control physical systems?”

toward:
“How do we operate those systems reliably at scale?”

The event repeatedly showed that future competitive advantage in physical AI may depend less on frontier models and more on:

  • orchestration layers

  • communication infrastructure

  • runtime tooling

  • deployment systems

  • hardware-aware optimization

  • validation loops

  • and operational reliability

The core signal from the event:

When AI enters the physical world, infrastructure becomes the product.

Speakers & Companies Featured

  • Ivan Dyachkov — Field CTO (EMEA), EMQ Technologies

  • Yuri Gusev — CTO, Encube Technologies

  • Andreas Ask — Deep Learning Researcher & Product Owner, Embedl

Topics covered:

  • Physical AI

  • AI orchestration

  • MQTT

  • edge AI

  • embedded systems

  • hardware-aware AI

  • AI infrastructure

  • distributed systems

  • generative CAD

  • manufacturability

  • runtime reliability

  • edge deployment

  • AI agents controlling devices

  • AI communication layers

Related Insights

Upcoming Events

👉 Unlocking Enterprise AI Productivity — Stockholm MLOps #35

👉 Air-gapped MLOps — Stockholm MLOps #36

Event Details

📍 HiQ, Stockholm
📅 May 21, 2026
🤝 In collaboration with EMQ Technologies & HiQ
👥 Stockholm MLOps Community

👉 Stockholm MLOps Community:
https://www.meetup.com/stockholm-mlops-community/

Previous
Previous

Sovereign AI & Control AI

Next
Next

AI Agents and the Future of Health | Stockholm MLOps Event #33 | May 18 2026