CDFAM Barcelona 2026 — Conference Report

CDFAM Barcelona 2026 was a different kind of engineering conference. No vendor booths competing on feature checklists. No "AI strategy" slides from incumbents. Instead, ~30 talks from founders, engineers, and researchers, almost all converging on the same argument: engineering is at a generational inflection point, and the tools we use to design products are being rewritten from the ground up.

Hat tip to Duann Scott and the CDFAM team for organizing what was easily the most intense — and most well-curated — symposium I attended this season.

The Exponential Gap

The framing talk of the conference, repeated in different forms by multiple speakers, was about what Wasil Rezk (BeyondMath, CCO) called the exponential gap.

Product complexity has grown exponentially over the last 30 years. Design variables, multidisciplinary constraints, and the cost of getting a wrong answer have all scaled with it. Simulation tools — and most of CAD — have not. They are still designed around a single engineer iterating one variable at a time.

Nico Haag (PhysicsX) put a hard number on what happens when you try to bridge that gap with a layer of AI bolted on top: 90–95% of engineering AI pilots fail to scale. Not because AI doesn't work, but because the surrounding workflow, data model, and domain reasoning weren't redesigned to support it.

The phrase that stuck: AI-native engineering. Not "AI assistance." Not "AI features." A reset.

Three Pillars of AI Integration

Javier Blanco (Quix) offered the cleanest framework I heard all week. Three things, all required:

1. Data accessibility — raw, centralized, addressable. Not "we have a data lake somewhere."
2. Structured domain knowledge — physics and engineering hypotheses embedded into the reasoning layer, not implicit in the engineer's head.
3. Autonomous execution — agents that iterate code against real data, without a human approving each step.

His case study was a European vacuum manufacturer that compressed a balancing-algorithm optimization from days to minutes. The headline number is impressive. The structural lesson is more important: it worked because all three pillars were in place. Bolting an LLM onto a broken data pipeline produces nothing.

Digital Co-Workers

The most consequential shift in the room was the move from "AI-assisted" to autonomous agents.

A speaker from Synera/SEAT predicted that 60% of current engineering time will be eliminated as digital co-workers absorb the repetitive cognitive work — variant analysis, regression checks, compliance validation, parametric sweeps. Not because the work goes away, but because a human stops doing it.

The framework most commonly cited: LLM core + company knowledge + specialized automation layers. The LLM provides language and reasoning. The company-specific knowledge graph provides domain grounding. The automation layers (CAD APIs, simulation runners, data extractors) provide execution. Each piece is replaceable; none of them is the product.

Industry Applications That Are Already Shipping

Several talks moved past the framing to show working systems:

• Maritime — Shahroz Khan (Compute Maritime) showed foundational models for ship design. The first offshore vessel "designed, simulated, and optimized" entirely via AI is now in the water. (Compute Maritime later presented at Threaded Warwick — see the deck linked from the conferences page.)
• AEC / Construction — Richard Zhang (Augmenta) showed "Functional Intelligence" coordinating architectural, structural, and electrical constraints automatically. Building geometry as an output, not an input.
• Consumer products — Yuan Mu (Nike) showed AI handling personal style and performance outcomes simultaneously — the kind of multi-objective optimization that breaks traditional CAD parameterization.

Trust, Provenance, and Determinism

The trust conversation at CDFAM was unusually serious. Two speakers stood out.

Rebeka Melber (Istari Digital) argued that the industry needs vendor-neutral, machine-readable artifacts with unique identifiers. Not "AI-generated reports." Provenance-bearing data objects that downstream tools — and humans — can verify. Without this, every AI output is a hallucination risk.

Rhushik Matroja (Cognitive Design Systems) added the regulated-industry view: deterministic workflows with validated pipelines are not optional in aerospace, defense, or medical. He demonstrated a 600-hour design project compressed to 200 hours — but only because the AI sat inside a deterministic harness that produced verifiable, repeatable outputs.

The lesson: speed is not the moat. Verifiable speed is.

The Engineer's Evolving Role

The most quietly profound thread was about what engineers actually do in this new world.

The before: engineers spend their cognitive budget figuring out which CAD command to run. Memorizing menus. Wrestling with constraints. Searching forums.

The after: engineers spend their cognitive budget on design intent — what the product needs to be, why, and what the success criteria are. The AI handles iteration. The human validates whether the output actually solves the problem.

This is not a small change. It rewrites what we hire for, what we train for, and what counts as engineering judgment.

The Hard Numbers

Two statistics worth carrying home:

• ~9% of engineering companies have mature AI implementations
• ~80% are currently in pilot/experiment mode

Read those numbers as a window. The AI-native engineering shift is real but not yet broadly distributed. The companies that move from pilot to production in the next 18 months will set the operational pace for the rest of the decade.

Companies and Speakers Worth Following

BeyondMath, PhysicsX, Quix, Synera, SEAT, Compute Maritime, Augmenta, Nike, Istari Digital, Cognitive Design Systems, Siemens, NVIDIA, and Generative Engineering all had substance on the program.

If you want the deeper dives, several of these vendors also presented at Threaded Miami and Warwick — their decks are available on the conferences page.

Read

CDFAM Barcelona was the cleanest, sharpest framing of where engineering software is going that I've seen this year. Not breathless. Not vendor-driven. Just engineers and founders looking at the gap between what tools do and what products now require, and arguing — with evidence — about how to close it.

If your PLM, CAD, or simulation roadmap doesn't have a serious answer for "what does AI-native look like in our stack," you are budgeting for the wrong decade.