Why I’m (Mostly) Bearish on AI for Manufacturing – And Where I See Real Opportunities Hiding

Closed ecosystems, skinny TAM expansion, and the hard physics of parts make pure-play SaaS a tough sell. But owning the order flow and augmenting human expertise can still build durable businesses.

Source: Gemini

The hype-to-gravity ratio

No vertical attracts more AI pitches per unit of TAM than manufacturing. But for every headline about a “GPT-for-machining” demo, there are ten 50-year-old job shops still running on clipboards, tribal knowledge and a vertical mill from the Reagan era. Meanwhile, the poster-children of “AI-driven” manufacturing marketplaces, Xometry and Fictiv, are still wrestling with negative EBITDA and stubbornly high operating costs – even at nine-figure revenue scale. Xometry’s FY-2024 adjusted EBITDA was ($9.7) million on $545 million revenue (source), proof that smarter algorithms alone aren’t compressing the human overhead of quoting and QA the way slide decks anticipated.

Why is it so hard?

1. Walled-garden CAD monopolies. Dassault, PTC, Siemens and Autodesk control >80% of mechanical-design seats. This might sound like an upstream corner of the workflow, but design files are the source of truth for every downstream step – toolpath generation, tolerance analysis, inspection programming, even the purchasing spec inside an ERP. If an AI startup can’t read or write the native file without a brittle plug-in, it’s locked out of the “digital thread” and forced into a feature niche. The market value of plug-ins shows how little room there is: Sandvik’s 2021 purchase of Mastercam, the most popular CAM add-on, reportedly closed for well under $300 million despite decades of dominance.

2. TAM ceiling: the 5% problem. Feature startups automate a sliver of the workflow (tolerance extraction, DFMA hints) and bump effective TAM by single-digits; buyers shrug and stay put.

3. Integration friction > SaaS elasticity. Even “freemium” tools demand days (usually longer) of process change – deal killers in shops that still fax POs. (caveat: agentic integration is coming / already here for some areas so I expect this may change quickly)

4. Physics beats probability. One extra zero in a tolerance call-out (.0005 in → .00005 in) can increase cost 10x. A hallucinating LLM can wipe out the margin on an entire job.

Even tech-forward manufacturers learn fast that revenue scales almost 1-for-1 with headcount and capex. Whether you hold the spindle or wrangle suppliers, unit economics snap back to physics and labor.

Where AI can win

One note before diving in: some of the biggest wins in manufacturing tech will come simply from making existing tech actually usable. A lot of predictive tools fail not because the models are bad, but because the data they rely on is trapped (sitting in a warehouse system, never syncing cleanly with ERP inventory or product feeds). As in other verticals, a real unlock is reducing onboarding friction. The startups that solve this – by mapping messy spreadsheets, pulling signal from emails, or piping clean data from legacy systems – will close more deals and stick longer. It’s the boring glue that makes the rest of the stack work.

Machine-health copilots.
Unplanned downtime is still the costliest line item most plants can actually control, so edge-sensor stacks that stream vibration, temperature and power-draw data into AI anomaly detectors can pay for themselves in weeks. Pairing the hardware with cloud CMMS workflows shifts the spend from lumpy capex to a predictable subscription, while every avoided hour of shutdown produces a hard-dollar ROI that maintenance teams can sign off on quickly. The result is one of the few manufacturing beachheads where a hardware-plus-software bundle naturally matures into sticky, SaaS-like revenue – making the predictive-maintenance corner of Industry 4.0 a bright spot amid otherwise asset-heavy economics.

Owning the order for off-the-shelf parts.
Industrial distribution is a $500B+ U.S. market (and well over $1T globally) that still operates on late-1990s user experiences (think companies like Grainger). By applying AI to part-number and spec-similarity search, a platform can instantly swap obsolete or overpriced SKUs for readily available equivalents – capturing value at the moment a purchase order is cut, rather than charging a per-seat software fee. Because suppliers are eager to pay for incremental revenue, a transactional take-rate model is the most effective go-to-market approach.

Blueprint & model intelligence.
Extracting tolerances, surface-finish notes and other critical details from 2-D PDFs and correctly mapping them onto the corresponding 3-D geometry unlocks meaningful, data-driven DFM risk scores. The company that finally nails this 2-D/3-D linking problem will become indispensable middleware to every major CAD/CAM suite. A pragmatic path is to begin with high-touch services – proving time and cost savings on real jobs – then graduate to an API subscription once customers see clear ROI.

Risk & availability forecasting.
Accurately predicting BOM-level lead-time spikes or part shortages is a pain point with real budget behind it. Platforms that ingest procurement signals across the web and combine them with ECAD/MCAD data can alert OEM supply-chain chiefs (who control meaningful spend) before delays hit the production schedule. The winning play is to land on a “shared-savings” or cost-avoidance mandate, then expand by upselling suppliers and tier-1 vendors on revenue-generating visibility tools.

What I’m not betting on (yet)

• End-to-end quoting for job shops. Unless you control the spindle, you inherit the same principal-agent mess that drags Xometry’s unit economics. And even if you do control the spindle, you are then weighed down by the actual production and QA of the piece, which is still relatively opex intense.

• Generative CAD from voxels. True B-rep-native AI design would be a revolution but it likely requires novel math, not just bigger transformers. The holy grail is an engineer typing “lighten this bracket by 30%” and instantly receiving a watertight, tolerance-perfect CAD model that CAM, FEA, and inspection tools can run without manual cleanup. Today’s generative models mostly operate in voxels, meshes, or point clouds – good for visual ideation, but not manufacturable geometry. They don’t understand exact surfaces, tangencies, or micron-level tolerances that real machines require. Getting there will require geometric deep learning that reasons directly over edges, faces, and topological constraints. In other words: we don’t just need smarter AI, we need new computational geometry built for manufacturing.

• Point solutions confined to one OEM. Cutting scrap by 5% at Tier-2 aero shops looks great in a pitch, but the market is maybe a few hundred plants worldwide. Even if you win every one at $100K ARR, you’re staring at ~$30–50M tops – and every sale drags through months of audits and bespoke integrations.

A final word on timing

Manufacturing is heavy. The gravity of machines, legacy software, and precision constraints means most AI startups won't get liftoff with clever demos alone. And while defense budgets and on-shoring tailwinds will keep capital flowing into “Industry 4.0”, the biggest opportunities will be those that wrap software around real cash flow (parts, POs, distribution margin) and use AI to compress cost-of-sales or expand gross profit. If you’re building in those lanes, or think I’m wrong, let’s talk.