Apr 4, 2025 · 38m · catalyst
Specialized AI brains for physical industry
0:00 / 0:00
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of Catalyst, host Shail Khan and investor Sam Smith-Eppsteiner examine why specialized vertical artificial intelligence is poised to transform heavy physical industries. They explore how domain-specific models, multi-agent architectures, and tailored workflows overcome legacy data silos and labor shortages across manufacturing, logistics, and engineering.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Shayle holds 34.6% of the talking time here. How this is scored →
Shayle as informed peer 4.3 Guest teaching 2.9 Guest disagreement 1.1 Shayle pushing back 2.1
speaking balance: gold is Shayle, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 36:44 Challenging the CapEx budget pricing thesis
Sam swiftly counters Shayle's proposal to price against CapEx budgets, pointing out that proving a facility lasts ten years longer cannot be validated for 30 years, whereas labor savings show ROI in two weeks.
Hardest push from Shayle ▶ 28:27 Questioning market size for siting and complianceShayle forcefully questions whether siting optimization tools can build generational venture-scale businesses or if they merely result in hundreds of small, fragmented point solutions.
Biggest teaching moment ▶ 19:20 Educating on cross-system semantic complexitySam dismantles Shayle's simplistic 'encyclopedia' framing by detailing how answering a simple job-site question requires resolving natural language across blueprints, ERPs, and procurement systems.
Shayle holds their own ▶ 9:59 Applying energy sector experience to data access barriersShayle brings sharp domain expertise on utility data privacy, explaining the historical failure modes of electricity software startups unable to access sensitive internal datasets.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Shayle as informed peer | Guest teaching | Guest disagreement | Shayle pushing back | Why |
|---|---|---|---|---|---|---|
| Host Monologue: Framing AI in Physical Industries | 3 | 2 | 1 | 2 | Shayle frames the thesis around specialized vertical AI versus general foundation models and asks whether legacy, siloed PDF data is truly unique to physical industries. Sam explains that physical industrial software stacks are uniquely antiquated, often dating back to the 1980s or 1990s. | |
| Overcoming Proprietary Data and Technical Visual Formats | 5 | 2 | 0 | 1 | Sam explains why proprietary, technical visual data like blueprints creates a moat against foundation models. Shayle connects this directly to his own venture experience in the electricity sector, noting how inaccessible utility data has historically constrained software startups. | |
| Addressing the Knowledge Gap and the Great Crew Change | 4 | 3 | 2 | 4 | Sam brings up the Great Crew Change as a key demand driver. Shayle pushes back, arguing that generational turnover is a slow-moving macro trend like climate change rather than an acute, burning pain point, though Sam notes oil and gas customers are already expressing immediate panic. | |
| Sponsor Messages: Bloom Energy, Engie, and Energy Hub | 3 | 4 | 1 | 1 | Following sponsor breaks, Shayle summarizes the knowledge-base category as a 'sentient encyclopedia'. Sam gently reframes this, using a detailed construction procurement example to show how multi-system reasoning goes far beyond simple information retrieval. | |
| Agentic Workflows and Automating Manual Supply Chain Tasks | 3 | 2 | 0 | 0 | The conversation shifts to agentic AI workflows. Sam breaks down logistics automation case studies like Hubflow and Conduit, highlighting how agents can automate tedious, error-prone manual scheduling tasks across enterprise boundaries. | |
| Engineering Co-pilots and Automated Design Verification | 4 | 2 | 0 | 0 | Shayle and Sam discuss engineering co-pilots and design verification. Sam details portfolio company Catstrom's approach to PCB verification, drawing parallels to how software engineering tools like Cursor delight users by automating the least enjoyable tasks. | |
| Compliance Automation, Siting Tools, and Market Defensibility | 6 | 4 | 2 | 5 | Shayle challenges the commercial viability and total addressable market of niche compliance and siting tools, questioning whether they can ever become venture-scale businesses. Sam counters by discussing data defensibility, expansion into transactional payment layers, and FDA-focused analogues. | |
| Business Models, Pricing Strategy, and Labor vs CapEx Budgets | 6 | 4 | 3 | 4 | The discussion turns to pricing models, inference margins, and tapping labor versus CapEx budgets. Shayle argues that extending physical asset lifetimes could unlock massive capital budgets, but Sam counters that proving CapEx lifetime extension takes decades while labor savings can be demonstrated in weeks. |
Statements from this episode (11)
Smith-Eppsteiner: Some law firms use Anthropic directly over Harvey
“Like I know law firms that are using Anthropic instead of Harvey or legal AI tools, so there's definitely anecdote on both sides.”
Smith-Eppsteiner: Last multi-billion hardware engineering software started in 1990s
“If you look at, I think we were looking at once you know, companies that have built sort of multi-billion category defining products selling to hardware engineers, and the last one was started in like the nineties. Maybe even the eighties.”
Smith-Eppsteiner: Foundation AI models struggle with blueprints and technical diagrams
“I think the reality is these models are just not that good today at sort of understanding you know, a very technical diagram, a blueprint, because they haven't seen enough of them.”
Khan: Industrial AI startups must train on customer data to scale
“If the training data is all private and walled, you got to be able to use your customer's data to train. They're not going to like that, and you got to figure out how to get over that hump, and if you can't get over that hump, there's kind of no way you end up…”
Khan: Industrial Labor Demographic Shifts Will Take Decades to Drive AI Adoption
“It's one of these things, though, that I do think is a big macro driver, but it's not like an, it's like climate change in some ways. It's like, it's going to take decades to play out, and that's like one of the problems with it, is that there's no immediate, …”
Smith-Eppsteiner: Prompt engineering reaches 70% to 90% of performance
“In talking to, you know, founders who are actually building, it seems like prompt engineering can get you quite far, right? So like, it's widely variable, but let's say you can get to like, 70, 90% of where you need to be from a performance perspective.”
Smith-Eppsteiner: DeepSeek proved low-cost, small models are viable for startups
“Like DeepSeek and others have shown, they're Is a path to sort of lower cost, smaller models that are viable for startups to build.”
Smith-Eppsteiner: Agentic AI works best on annoying, error-prone, or seasonal tasks
“What's a good fit is anything where the work is annoying, like, where the person who's doing it actually doesn't like doing it, finds it frustrating for whatever reason Is where the work is already error-ridden. I think that's very common where you have to ref…”
Smith-Eppsteiner: Hubflow automates trucker scheduling without altering receiver workflows
“Hubflow is automating that for the trucker, but the third party, the receiver, actually has the same exact workflow. Like, to them, it's invisible.”
Khan: Siting optimization for energy developers cannot build venture-scale businesses
“I do think there is value in that. I don't think you can build an enormous venture grade business just doing that.”
Smith-Eppsteiner: AI companies face 40% to 70% gross margins from inference costs
“Like, inference cost is non-trivial. I think for a bunch of the AI companies, both in our portfolio and that we've seen you know, this is, like, not going to be a 90% margin SaaS. Like, we may be talking about 40 to 70% gross margin based on inference costs.”