3.4 Analysis¶
The five questions were answered by reading across every company study. This page gathers the cross-cutting analysis behind those answers: the capability matrices mapping who does what across the workflow, the engineering intelligence stack showing where the gap sits, the structural patterns that recur, and the corporate landscape rearranging the pieces. The answers themselves are drawn together in the Conclusion.
3.4.1 Capability matrices¶
Two matrices: the supply side (Platform Vendors) and the demand side (Service Providers, OEMs and Tier-1s).
Reading the row axis — sixteen stages, five phases. Both matrices are indexed down the left by the sixteen concrete stages a moulded part passes through, from customer requirements to mass production. These sixteen stages are not a rival standard: they are a fine-grained decomposition of the five APQP phases from §1.4 — each phase holds several stages. The finer resolution is kept deliberately, because the single most important result below — where AI actually lands — is only visible at the stage level (it concentrates at stage 8 and stage 16).
| APQP phase (§1.4) | Stages in these matrices |
|---|---|
| 1 · Plan & Define | 1 Market/customer reqs · 2 Vehicle reqs · 3 Part reqs · 4 Industrial design |
| 2 · Product Design & Development | 5 Concept · 6 CAD modelling · 7 Engineering review · 8 Simulation · 9 DFM |
| 3 · Process Design & Development | 10 Tool design · 11 Mould-flow analysis |
| 4 · Product & Process Validation | 12 Prototype · 13 Validation & testing · 14 Design changes |
| 5 · Launch, Feedback & Corrective Action | 15 Production release · 16 Mass production |
(The company reports' "Role in Workflow" tables use a condensed ~13-row version of these sixteen stages; the matrices below use the full sixteen.)
Mechanism legend (the honest-AI test made operational):
| Code | Mechanism class | Honest-AI verdict |
|---|---|---|
| R | Rules engine (checklists, KBE (knowledge-based engineering), feature checks) | Not AI — deterministic rules |
| P | Classical physics solver | Not AI — computes from equations |
| O | Optimization / DoE (often marketed as AI) | Not learned AI — automated search (see What is Optimization technology) |
| K | Retrieval / similarity search | Genuine data capability, unsupervised |
| ML | Genuine learned machine learning | Passes the honest-AI test |
| L | LLM / document AI (chat, RAG, copilots) | Learns language, not physics |
| — | Nothing found | Negative finding, first-class result |
Qualifiers — hover over an entry for its meaning (the qualifiers are also listed here): (p) preview · (r) research only · (t) thin, vendor-claimed · (a) academic. Two plain markers carry footnotes: *1 · †2.
3.4.1.1 Vendor matrix (Platform Vendors × 16 stages)¶
| # | Workflow stage | Dassault | Siemens (+Altair) | Autodesk | PTC | Moldex3D | SIGMASOFT | Hexagon (post-Feb 2026) | SIMCON | Cadence |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Market / customer reqs | L (Aura) | L (copilots) | — | L (Codebeamer Copilot, assistants) | L (MoldiBot, support) | — | — | — | — |
| 2 | Vehicle requirements | — | — | — | — | — | — | — | — | — |
| 3 | Part requirements | — | — | — | — | — | — | — | — | — |
| 4 | Industrial design | — | — | ML (Form Explorer) | — | — | — | — | — | — |
| 5 | Concept design | O (generative design) | ML (PhysicsAI Generate) | O (Fusion gen. design) | — | — | — | — | P + ML(p) (early-design fill check) | — |
| 6 | CAD modelling | O + ML (sketch GNN — graph neural network) | L (Design Copilot NX) | ML (AutoConstrain) + O | O (Frustum gen. design) | — | — | — | — | — |
| 7 | Engineering review | L | L | — | L (Windchill assistant) | K (iSLM Discovery) | — | — | — | — (BETA CAE, no AI) |
| 8 | Simulation | P + ML (structural surrogate) | P + ML(p) (PhysicsAI) + O (HEEDS) | P (Moldflow) + ML(r) | P* (CMA = Creo Mold Analysis — Moldex3D technology, reseller-branded) | P + O ("AI Optimization Wizard") | P | — (Digimat/ODYSSEE sold to Cadence) | P + ML(p) + O (Varimos) | P (Digimat) + ML (ODYSSEE) |
| 9 | DFM | R (DFMXpress) | R (DFM Advisor; DesignSim*) | — | R + ML(p) (Creo 13 beta seed) | R (DesignSim) | — | R (VISI) | — | O (Digimat gate/UQ — uncertainty quantification) |
| 10 | Tool design | — | R (Mold Wizard) | O (cooling-channel opt.) | R (EMX) | K (iSLM) + L(t) (cooling-channel LLM, unshipped) | P (full-mould model) | R (VISI CAD-CAM) | — | — |
| 11 | Mould flow analysis | P | P (Inspire Mold) + ML(p) (PhysicsAI-on-Mold, Jun 2026) | P + L(p) (Assistant Tech Preview) | P* | P | P | P (VISI Flow) | P + ML(p) (Cadmould AI Solver, Mar 2026, filling only) | — (owns no moulding solver) |
| 12 | Prototype | — | — | — | — | ML(t) (Process Discovery) | P (virtual moulding trials) | — | — | — |
| 13 | Validation & testing | — | — | P | — | K | — | ML (PC-DMIS AI Feature Scan) | P | P |
| 14 | Design changes | — | — | — | — | — | — | — | ML (Varimos Real, per-mould) | — |
| 15 | Production release | — | — | — | L (Arena AI) | — | — | — | — | — |
| 16 | Mass production | — | — | — | — | ML(t) | O (Autonomous Optimization = virtual DoE) | — | ML (Varimos Real, measured trials) | — |
What the vendor matrix shows: [Synthesis]
- Row 11 is the contested row and it is almost entirely classical. The three AI entries are all dated 2026 and all preview-grade. The most advanced (SIMCON) covers filling only, not warpage — the money problem (see SIMCON).
- The ML entries cluster far from the mould: CAD sketch models, styling generators, structural surrogates. Where the mould is, the AI is thinnest. This is the inverse-correlation pattern of Structural Patterns pattern #1 in matrix form.
- K appears exactly three times — all Moldex3D. Retrieval — the unsupervised, no-labels capability — has one vendor, ecosystem-locked, no CATIA path (see Moldex3D). [Evidence]
- SIGMASOFT's column has no relabelled-O marketing and no L — the vendor never says "AI" at all, which makes it the calibration column for every other vendor's marketing (see SIGMASOFT). [Evidence]
- Cadence's column is nearly empty in this workflow despite the study's deepest genuine production AI (Cerebrus, ChipStack — chip design, outside these 16 stages): capability proven, moulding intent unstated (see Cadence). [Evidence]
- The only capability in the vendor matrix that learns from measured moulding trials is SIMCON's Varimos Real — per-mould statistical fitting on machine trials, carried nowhere between projects (see SIMCON). (Varimos Real is coded ML rather than O because it fits a statistical model to measured trial data — it learns from measurements, unlike the purely search-based optimizers.) [Evidence]
3.4.1.2 Customer matrix (Service Providers, OEMs & Tier-1s × 16 stages)¶
| # | Workflow stage | HCLTech | Bertrandt | LTTS | Tata Tech. | BMW | VW Group | Hyundai | ZF | Forvia | Bosch |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Market / customer reqs | — | L (Bernd RFQ, Feb 2026) | L (PLxAI) | L (Discoveria) | — | L (Codebeamer Copilot) | — | L(t) | — | — |
| 2 | Vehicle requirements | — | — | — | human | human | human | human | — | — | — |
| 3 | Part requirements | L (XLM.AI, metadata) | — | — | human | human | human | human | L(t) | — | — |
| 4 | Industrial design | — | human studios | — | — | — | ML (Audi FelGAN — wheels, not moulded) | — | — | — | — |
| 5 | Concept design | — | human | — | — | ML (Synera load-floor surrogate) | — | — | — | — | — |
| 6 | CAD modelling | services (human) | human (CATIA/NX) | R (KBE / CAx automation) | R (eVMP parametric) | human | human | human | human | O (xGenerative Design) | human |
| 7 | Engineering review | R (DFMPro) | L (FMEA Agent, Apr 2026) | R (checkers) | human | human | human | human | human | human | human |
| 8 | Simulation | — | P (crash/CFD — no moulding sim) | P (300+ CAE engineers) | P + ML(t) | ML (Mistral LIM, crash) | ML (Porsche RL crash) | ML (Simcenter neural ROM (reduced-order model), chassis) | — | ML* (ODYSSEE crash ROMs) | P + ML(r) (moulding papers) |
| 9 | DFM | R (DFMPro inj.-moulding module) | — | R (+ Moldflow aspiration) | human (unmarketed) | — | — | — | — | — | human (BGSW service) |
| 10 | Tool design | R (partial) | human tool shop (750 t) | human (tyre moulds only) | human (with suppliers) | — | R/statistical (Braunschweig CAM KM, honestly non-AI) | O(a) (KAIST Bayesian rib study) | — | human mould shops | human (BMS tooling) |
| 11 | Mould flow analysis | — (no solver) | — (no solver named) | — (bare line item) | human (Moldflow-skilled staff) | delegated | internal, no public AI | undocumented | — | P* (Moldex3D) | P* (Moldflow, rented) |
| 12 | Prototype | — | MuCell / rapid tooling (physical) | — | — | — | — | — | — | — | — |
| 13 | Validation & testing | human (ASAP) | ML (computer-vision (CV) motion tracking) | ML (LTwin PINN — assets, not moulding) | — | ML (CT inspection) | — | — | ML (ZF Annotate — ADAS) | ML* (ODYSSEE) | — |
| 14 | Design changes | — | — | — | L (warranty GenAI) | — | — | — | — | — | — |
| 15 | Production release | L (PLM services) | — | — | ERP/PLM | — | — | — | — | — | — |
| 16 | Mass production | — | ML (vision — press shop, sheet metal) | ML (LTwin — assets) | ML (Visimatic CV) | ML (AIQX; Digital Moulds = IoT, not ML) | ML (spot-weld AI, 1.5M welds/shift) | ML (factory/robotics) | ML† (sensXPERT — vendor died 31 Jul 2025) | automation ("dark plants" — claim dates from older reporting) | ML (GenAI inspection — welding, not moulding) |
This matrix extends the six mechanism codes with plain-language markers — 'human' (work done manually), 'delegated' (pushed to a supplier), 'undocumented' (no public evidence either way) — because customer-side reality is often none of the six.
What the customer matrix shows: [Synthesis]
- Stages 9–11 are human or classical at all ten companies. The demand-side moulding middle exactly mirrors the supply-side one. This is the central symmetry (developed in The Engineering Intelligence Stack).
- Stage 8 and stage 16 carry nearly all the ML — the barbell, now visible as two dense matrix rows with an empty band between them.
- Every stage-8 ML entry is simulation-taught (crash ROMs, RL on FEM, neural ROMs on Amesim archives); every stage-16 ML entry is sensor/image-taught. No entry anywhere is taught by geometry-plus-production-outcome pairs. [Evidence of absence — reports passim]
- The one moulding-ML cell is marked †. ZF's sensXPERT deployment was real, quantified, and fleet-scale — and is orphaned (see ZF). Vendor viability is now part of the capability picture itself (see Structural Patterns pattern #8).
- The bundle-note companies would add almost nothing but "—": Motherson (131 moulding facilities) and Tata AutoComp (100+ presses) have empty AI rows; Magna and Denso add stage-16 factory AI only; Continental/Aumovio adds the single quantified customer engineering-AI win, at stage 1 (requirements, up to 80% faster against a baseline of up to 37,500 hours) (see Tier-1 Suppliers overview). [Evidence]
3.4.2 The engineering intelligence stack¶
The five-layer stack — Intent / Knowledge / Reasoning / Execution / Feedback — applied to all five groups at once. The question per layer: who owns it, for moulded-parts engineering?
| Layer | Vendor side (Platform Vendors) | Customer side (Service Providers · OEMs · Tier-1s) | Verdict for moulding |
|---|---|---|---|
| Intent (requirements → engineering tasks) | Copilots everywhere: Aura/Leo (Dassault), Design Copilot (Siemens), Creo/Windchill/Arena assistants (PTC) — all document-bound, all on rented LLMs | Bernd (Bertrandt), PLxAI (LTTS), Discoveria (Tata Tech.), Codebeamer Copilot (VW), and the one quantified win: Continental's requirements tool, 80% faster | Occupied, shallowly. Language AI over documents; none of it knows what a rib or a gate is. [Synthesis] |
| Knowledge (what the organisation already knows) | One product: Moldex3D iSLM Discovery (retrieval over mould history, ecosystem-locked). Everything else is RAG over manuals (MoldiBot) | Unmined assets everywhere: Forvia's ~2,500 CATIA seats of history, Motherson's 131-site archive, Bosch with "no geometry-retrieval capability found", VW's Braunschweig CAM system (real knowledge capture, frozen pre-AI) | Nearly empty. One vendor product, zero customer deployments. [Evidence of absence] |
| Reasoning (predicting consequences of design decisions) | Classical solvers, plus the 2026 cracks: SIMCON neural filling (preview), PhysicsAI-on-Mold (one month old), ODYSSEE at Cadence (structural, one step short of the mould) | Simulation-taught surrogates at the giants — BMW/Mistral LIM (crash), Hyundai ROM (chassis), Porsche RL (crash), Forvia ODYSSEE (crash); Bosch's moulding surrogates stayed research papers | Classical where it matters. Learned reasoning exists — for crash, never for fill/pack/warp in production use. [Evidence] |
| Execution (doing the work) | The strongest vendor layer: CAD/CAM automation, Mold Wizard, EMX, KBE, rules DFM | The strongest customer layer: thousands of engineers, LTTS KBE (12 → 0.5 man-days cases), Denso CAE templates (up to −80% analysis time) — automation, honestly non-AI | Owned and mature — by rules, templates and people. [Evidence] |
| Feedback (learning from what actually happened) | Absent. No vendor ingests scrap/warpage/tool-tuning outcomes. SIMCON's Varimos Real (per-mould trial fitting) is the closest, and it carries nothing between projects | Factory ML feeds acceptance, not design: AIQX, spot-weld AI, Bosch inspection, Magna vision. sensXPERT closed a real per-shot loop — and its vendor is dead | The empty layer, both sides. Measured reality never reaches the design chain. [Evidence of absence] |
3.4.2.1 The central finding: the middle is empty on both sides¶
Stack the two matrices and the same hole appears in each:
- On the vendor side, the Knowledge and Reasoning layers for moulding are one ecosystem-locked retrieval product and three preview-grade 2026 capabilities. [Evidence — see Capability Matrices §3.4.1.1]
- On the customer side, the same layers are empty outright: no company among the Service Providers, OEMs or Tier-1s has deployed retrieval over its part history, learned DFM, or a moulding surrogate. The only dated, quantified requirements/document-AI win in the Tier-1 tier sits at the Intent layer (Continental) (see Tier-1 Suppliers overview). [Evidence]
This is the study's most consequential symmetry. The market is not a supply problem alone (vendors haven't built it) nor a demand problem alone (customers haven't asked); both sides of the Knowledge/Reasoning middle are vacant simultaneously, while both sides' Intent and Execution layers are crowded. Where the stack is full, the mechanisms are language models and rules. Where it is empty is exactly where moulding-domain learning would live. [Synthesis]
This analysis deliberately stops at establishing the hole. What to build into the empty Knowledge/Reasoning middle is a different question from whether it is empty — and the analysis answers only the second.
3.4.3 Structural patterns¶
Four patterns distilled from the Platform Vendors (#1–#4 below), extended, plus four more from the customer side (Service Providers, OEMs, Tier-1s; #5–#8), and a ninth spanning every group.
3.4.3.1 Moulding depth and AI depth are inversely correlated — everywhere¶
Among the Platform Vendors: the deepest moulding vendors (SIGMASOFT, Moldex3D) have the thinnest genuine AI; the deepest AI companies (Cadence, and excluded Ansys with SimAI) have no moulding stake. The study's exclusions make the point plainly (see the ESI and Ansys scope notes). [Evidence]
Extended to the customer side, the pattern holds inside organisations: [Synthesis]
- VW owns a ~700-person mould shop and a €1bn AI programme — in separate worlds (see Volkswagen Group).
- Bertrandt owns a 750 t tool shop and two shipped AI agents — publicly unconnected (see Bertrandt).
- Bosch owns plants, a plastics research arm and BCAI — they met in two papers and never in a tool (see Bosch).
- Tata Technologies: moulding intelligence in people, AI in document tools — never met (see Tata Technologies).
Across the ~40 organisations this study examined, no single company holds deep moulding knowledge and deep learned AI in the same product or workflow. That sentence is the study's clearest empirical result. [Synthesis]
3.4.3.2 Everyone rents the LLM layer¶
No company in any group owns its language model: Dassault→Mistral, Siemens→Azure OpenAI, Autodesk→Azure OpenAI, PTC→Bedrock+Azure, Moldex3D→OpenAI (Platform Vendors, each report's §3/§8); LTTS→Anthropic Claude, Tata Tech.→Azure/AWS (Service Providers); BMW→Mistral, VW→Microsoft (OEMs); Bosch→five rented providers plus on-prem Aleph Alpha (Tier-1s). [Evidence] The only non-renters rent nothing because they have no LLM at all (SIGMASOFT, Hexagon, SIMCON). [Evidence]
The consequence, sharpened by the External & Startups note: the rent flows to a model layer that is itself consolidating — and exactly one frontier lab (Mistral) has crossed from supplier to domain owner, holding the ex-Emmi moulding-surrogate team and a direct OEM data deal (BMW LIM) (see External AI Entrants). Renting is safe with the rent-collectors (Microsoft, AWS, Google); it is strategic exposure with Mistral. [Synthesis + Inference]
3.4.3.3 All learning is simulation-taught or document-bound — now with a third class¶
The Platform Vendors' finding: every vendor "learning" is either trained on solver outputs (PhysicsAI, Cadmould AI Solver, Process Discovery) or bound to documents (all copilots). Nobody learns from factories. [Evidence]
The customer side confirms it and adds one class: sensor-taught in-process learning — sensXPERT at ZF, and the machine-side startup cluster (eMoldino, STASA/Kistler, Symate, plus10; see Startup Landscape). This third class learns from real physics but sees no geometry: it watches the press, not the design. The full taxonomy of what "learning" means in this field, as of 2026: [Synthesis]
- Simulation-taught (crash ROMs, neural solvers) — learns physics, never reality.
- Document-bound (every copilot) — learns language, never physics.
- Sensor-taught in-process (sensXPERT class) — learns reality, never design.
The unoccupied fourth class — geometry + measured outcome, across a part library — is the one nobody in this study has built: the single most consequential absence the analysis surfaces. [Synthesis]
3.4.3.4 Specialists move faster than giants¶
Among the Platform Vendors: SIMCON (a Würselen specialist) released a public research preview of a neural solver before Autodesk or Siemens shipped any moulding ML; the enabling startup (Emmi) went founding→product→exit in ~18 months (see SIMCON; Startup Landscape). [Evidence]
Extended to the customer side: Bertrandt (distressed, mid-size) shipped two working agents while larger ESPs published frameworks; Hyundai — with a $3B AI factory — outsourced its one moulding-AI study to a university; the giants' pattern is to buy speed rather than build it (Emmi→Mistral, Navasto→Autodesk, Inspekto→Siemens, all within ~28 months) (see Startup Landscape). [Evidence + Synthesis]
3.4.3.5 (New — OEMs) The OEM barbell¶
AI at crash simulation upstream and factory quality downstream, zero at moulded-parts engineering, at every OEM studied — the seven-OEM sample (3 full reports + 4 bundle notes) has no exception (see Automotive OEMs overview). The barbell is not a sampling accident; it follows from delegation — the OEM does not own the moulding pain, so it does not buy moulding AI. Go-to-market corollary: sell to the supplier tier; the OEM logo is a lighthouse, not a first customer. [Synthesis]
3.4.3.6 (New — Service Providers) The ESP tier is white space, not a channel¶
Expected: a translation tier implementing vendor AI for customers. Found: no ESP anywhere sells moulded-parts AI, and none can be sold through off-the-shelf — the category would have to be co-created (EDAG and Quest Global are the most credible hosts) (see Service Providers overview). [Evidence + Inference] The tier's one product-grade asset cuts the other way: HCLTech's DFMPro is simultaneously the competitor to beat, a white-label channel, and a plausible customer for any learned-DFM entrant (see HCLTech). [Inference]
3.4.3.7 (New — Tier-1s) Tier-1 adoption is in-process-first¶
Where Tier-1s adopted moulding-adjacent AI at all, they adopted it at the press (sensXPERT at ZF) or the inspection station (Bosch, Magna) — never at design, DFM or mould flow. The reason is structural, not cultural: at the press, every shot labels itself; upstream, labels cost money nobody has budgeted. Adoption follows free labels. [Synthesis — Tier-1 Suppliers overview] The proven sales motion follows the same grain: instrumented pilot on one mould → measured percentage → fleet rollout, opex-priced (see ZF). [Evidence]
3.4.3.8 (New — Tier-1s / External) Vendor viability is now a capability criterion¶
sensXPERT — the one vendor a Tier-1 standardised a moulding-ML line on — ceased operations 31 Jul 2025; iMFLUX — the best-resourced process-control play, with P&G behind it — died in 2023 (see ZF; Startup Landscape). Consequence: every future Tier-1 AI purchase carries a vendor-death question, and every credible seller must answer it (escrow, on-prem, open formats). A capability that outlives its vendor is worth more than a better capability that may not — the customer side has learned this the expensive way. [Synthesis]
3.4.3.9 (Spanning all groups) Openness tracks distance from the mould¶
Across the nineteen companies studied in depth, exactly two have a substantial open-source footprint in engineering/CAD-relevant AI: Autodesk — the only Platform Vendor publishing CAD-ML code, papers and data — and Bosch — the only customer-side company publishing research code, including the physics-ML framework TorchPhysics. Everyone else's engineering-AI footprint is near-nil, and the near-nil findings were verified, not assumed. [Synthesis]
| Company | Group | Open-source AI posture | Key facts |
|---|---|---|---|
| Autodesk | Platform Vendors | Substantial — the vendor outlier | AutodeskAILab on GitHub: UV-Net (MIT), UVStyle-Net, BRepNet, Fusion 360 Gallery dataset (non-commercial research license). Shipped features and trained weights stay proprietary (§11). [Evidence] |
| Dassault Systèmes | Platform Vendors | Near-nil | 4 graphics repos; no models, no datasets, no Hugging Face org. AI monetized as sovereign-cloud service (§11). [Evidence of absence] |
| Siemens (+Altair) | Platform Vendors | Meaningful, but not engineering-AI | ~212 repos (hypervisors, Yocto, ROS#); OpenPBS (AGPLv3, HPC scheduler). physicsAI/romAI fully closed (§11). [Evidence] |
| PTC | Platform Vendors | Near-nil | ~33 repos of SDKs/extensions, no AI/ML; much of it for businesses PTC sold (§11). [Evidence of absence] |
| Moldex3D | Platform Vendors | None verified | Proprietary solver, proprietary measured material data, proprietary iSLM knowledge platform; the Python API is a product feature, not open source (§11). [Evidence of absence] |
| SIGMASOFT | Platform Vendors | Nothing found | Consistent with a 50-person closed-solver vendor (§11). [Evidence of absence] |
| Hexagon | Platform Vendors | Near-nil | Nothing attributable; "hexagon" ML repos are Qualcomm's DSP — a name collision (§11). [Evidence of absence] |
| SIMCON | Platform Vendors | None — but its partner is open | Emmi AI open-sourced the AB-UPT architecture and NeuralDEM; the trained moulding model is closed (§11). [Evidence] |
| Cadence | Platform Vendors | Near-nil | GitHub org exists with zero public repos. (The "Cadence" workflow engine is Uber's, not theirs.) (§11). [Evidence of absence] |
| HCLTech | Service Providers | Near-nil for engineering AI | Open-source presence is IT-side, not DFMPro/CAD-ML (§11). [Evidence of absence] |
| Bertrandt | Service Providers | Nothing found | One blog post advocating OpenEMS (energy, not engineering AI) (§11). [Evidence of absence] |
| LTTS | Service Providers | Near-nil | GitHub org with zero public repos; IP sold in engagements, not released (§11). [Evidence of absence] |
| Tata Technologies | Service Providers | Effectively nil | AI assets are proprietary service IPs in sales decks (§11). [Evidence of absence] |
| BMW | OEMs | Real, but production/vision only | BMW-InnovationLab repos + SORDI dataset; zero engineering-AI or CAD/simulation releases; crash corpus closed (raw material of the Mistral deal) (§11). [Evidence] |
| Volkswagen Group | OEMs | Thin | ML lab publishes papers; nothing moulding-relevant released (§11). [Evidence — thin] |
| Hyundai | OEMs | Nothing moulding-relevant found | Limited search; 42dot not audited (§11). [Evidence of absence — limited] |
| ZF | Tier-1s | Nothing found | ML embedded in products or bought in; the community learns nothing from its deployments (§11). [Evidence of absence] |
| Forvia | Tier-1s | Nothing significant found | Light search; provisional (§11). [Evidence of absence — light] |
| Bosch | Tier-1s | Substantial — the customer outlier | github.com/boschresearch: ~399 public repos (driving/perception/control research); TorchPhysics (Apache-2.0) for mesh-free neural PDE solving. No moulding- or CAD-geometry project visible (§11). [Evidence] |
- Openness is inversely correlated with moulding proximity. The two open companies (Autodesk's research arm, Bosch's research arm) open-source upstream building blocks. Every company whose revenue depends directly on moulding know-how — Moldex3D, SIGMASOFT, SIMCON, the Service Providers, the Tier-1 product lines — is fully closed. The closer to the mould, the tighter the fist. This is pattern #1 seen from the code-repository angle. [Synthesis]
- The Emmi AI split is the modern template: open the architecture (AB-UPT code on GitHub), keep the trained domain model (NeuralMould) proprietary. Openness markets the science; the weights carry the value (see SIMCON; Startup Landscape). [Synthesis]
3.4.4 Corporate landscape¶
Full relationship diagrams.
3.4.4.1 The 2024–26 wave, in one table¶
| Deal / event | Date | Moulding-AI consequence |
|---|---|---|
| ESI → Keysight | complete Jan 2024 | No moulding impact — ESI never had a thermoplastic injection solver (see ESI scope note) |
| Inspekto → Siemens | Feb 2024 | Moulded-parts vision QA absorbed by a giant (see Startup Landscape) |
| Navasto → Autodesk | Dec 2024 | Surrogate-AI talent now inside the Moldflow owner (see Startup Landscape) |
| Altair → Siemens | closed 26 Mar 2025 | Created the one giant owning solver + ML + install base: Inspire Mold + PhysicsAI (see Siemens) |
| Ansys → Synopsys | closed 17 Jul 2025 | Best-in-class CAE AI (SimAI) now owned by an EDA company; still no moulding use case (see Ansys scope note) |
| sensXPERT ceases operations | 31 Jul 2025 | The only deployed Tier-1 moulding-ML vendor gone; orphaned fleet at ZF (see ZF) |
| Hexagon Digimat/ODYSSEE/MSC → Cadence | closed 23 Feb 2026 (€2.7bn) | The best "real ML next to moulding" case changed hands; Hexagon keeps the metrology pipes, Cadence gets the ML — the loop is cut (see Hexagon; Cadence) |
| PTC sells Kepware/ThingWorx → TPG | closed Mar 2026 | PTC exits factory data — a retreat from the Feedback layer (see PTC) |
| Emmi AI → Mistral | 19 May 2026 | The neural-moulding team now belongs to a frontier LLM lab; SIMCON's AI partner sits inside Mistral (see SIMCON; External AI Entrants) |
| Customer side: Forvia Interiors → Apollo (Apr 2026); Continental → Aumovio spin (Sep 2025) + ContiTech → Lone Star (Jul 2026); Marelli Chapter 11 (Jun 2025); ZF divestitures | 2025–26 | The moulded-parts customer base is itself restructuring — PE owners now sit over major moulding estates (see Tier-1 Suppliers overview; Forvia) |
3.4.4.2 What the wave means for who can execute¶
Siemens is, as far as this study found, the only giant with all three pieces. After Altair: a native moulding solver (Inspire Mold), genuine trainable ML (PhysicsAI), and a global install base. It shipped the combination in June 2026 — at the designer tier, unproven with customers (see Siemens). If any incumbent closes the moulding-AI gap by itself, the structure says Siemens. [Synthesis]
Cadence bought the AI but not the mould. It now owns the ODYSSEE/Digimat layer that sat one step from moulding — and no moulding solver to attach it to. Its proven AI bench is chip-side. Capability proven, moulding intent unstated (see Cadence). [Evidence]
Hexagon kept the loop's other half. Metrology data pipes without the ML it sold — the measure→learn→correct loop is now split across two companies (see Hexagon). [Synthesis]
The moulding specialists stay independent — and private. Moldex3D (privately held), SIGMASOFT (family-held via Dr. Flender Holding), SIMCON. The deepest moulding physics is thus structurally outside the consolidation wave: not for sale on public markets, and partner-dependent for AI (SIMCON's neural partner is now inside Mistral). They are the natural partnership targets for anyone executing a moulding-AI strategy — and the natural acquisition targets if a giant decides to buy the domain. [Synthesis + Inference]
The frontier-lab wildcard. Mistral holds the ex-Emmi team, the Dassault partnership, and the BMW data deal — three depths no other lab has (see External AI Entrants). A moulding "Large Industry Model" is one procurement decision away if someone supplies process data; the mitigation is that moulding outcome data is fragmented across thousands of moulders — the long tail frontier labs are structurally bad at harvesting (see External AI Entrants). [Inference]
Net execution map: the ability to execute AI-for-moulding is concentrating into (a) one giant with all the pieces (Siemens), (b) one AI-rich buyer without moulding (Cadence), (c) three independent specialists without AI muscle, and (d) one frontier lab with the team and no stated intent — while the customer base restructures under PE owners who reward legible engineering-hour savings. Nobody currently occupies the middle that the stack analysis showed empty; the consolidation wave has so far rearranged the pieces without assembling them. [Synthesis]