Playback speed

Share post at current time

0:00

Transcript

Why Silicon Valley Got Industrials Wrong [Livestreams]

Meet Philipp Wehn, the a16z Founder using AI to improve Heavy Industries.

Feb 10, 2026

It takes time to create work that’s clear, independent, and genuinely useful. If you’ve found value in this newsletter, consider becoming a paid subscriber. It helps me dive deeper into research, reach more people, stay free from ads/hidden agendas, and supports my crippling chocolate milk addiction. We run on a “pay what you can” model—so if you believe in the mission, there’s likely a plan that fits (over here).

Every subscription helps me stay independent, avoid clickbait, and focus on depth over noise, and I deeply appreciate everyone who chooses to support our cult.

Help me buy chocolate milk

PS – Supporting this work doesn’t have to come out of your pocket. If you read this as part of your professional development, you can use this email template to request reimbursement for your subscription.

Every month, the Chocolate Milk Cult reaches over a million Builders, Investors, Policy Makers, Leaders, and more. If you’d like to meet other members of our community, please fill out this contact form here (I will never sell your data nor will I make intros w/o your explicit permission)- https://forms.gle/Pi1pGLuS1FmzXoLr6

This one’s a bit different from our usual conversations. We’re not talking about frontier models or training infrastructure—we’re talking about factories, rail networks, CNC machines, and the engineers who keep the physical world running. Philipp Wehn spent a decade inside Siemens and heavy industry before starting Nexxa.ai, and he has a thesis that cuts against most of what Silicon Valley has tried in this space: the horizontal SaaS playbook doesn’t work here. 90% of industrial problems are specialized problems, and until AI changed the economics of customization, there was no way to profitably serve them.

What I found most interesting was the specificity of where AI still falls short. Vision models can’t reliably count rooms on a floor plan. Computer use models can’t be fine-tuned for custom industrial tools. These aren’t abstract limitations—they’re the actual blockers preventing automation in environments where billions of dollars ride on getting it right. We also got into the forward-deployed engineering model, why Palantir’s approach works but doesn’t scale down-market, and what it actually takes to sell AI to industrial executives who’ve been burned by software promises for decades. If you’re building for any industry outside tech, there’s a lot here.

For a deep dive into why I think that industrials is a massive opportunity, read our deep dive into the low tech revolution here—

Before we get deep into the conversation, here is your regular reminder that we have a new foster cat that’s ready to be adopted. I call him Chipku (Hindi for clingy; his government name is Jancy), and as you might guess, he’s very affectionate. I’ve trained him to be better around animals and strangers, and he’s perfect for families that already have some experience with cats. We sleep together every day, and waking up to him is one of the nicest feelings. If you’re around New York City, adopt him here (or share this listing with someone who might be interested).

Companion Guide to the Livestream: on

This guide expands the core ideas and structures them for deeper reflection. Watch the full stream for tone, nuance, and side-commentary.

Guest: Philipp Wehn, Co-Founder & CEO of Nexxa.ai

Company: Nexxa.ai — Specialized AI agents for heavy industries. $14M total funding ($4.4M pre-seed led by a16z speedrun, $9M seed led by Construct Capital). Headquartered in Sunnyvale, CA.

1. The 90% Problem

The Event — Philipp Wehn, co-founder and CEO of Nexxa.ai, opened with a claim that reframes everything Silicon Valley has tried in heavy industries: 90% of industrial problems are specialized problems. Custom configurations, unique workflows, bespoke integrations. The standard playbook—build a horizontal SaaS product, sell it to everyone—doesn’t just underperform here. It fundamentally misunderstands the market.

Why decades of software sales failed — Industrial companies have grown fragmented legacy stacks organically over decades. Two automotive suppliers producing identical parts might run completely different ERPs, different MES systems, different parameter configurations. A standardized product cannot create equivalent value across these heterogeneous environments.

But that’s not the fatal flaw. The fatal flaw is the quality threshold.

Software startups have a luxury: they can ship scrappy. An API breaks, you reset it. A feature is buggy, you iterate. That tolerance doesn’t exist when you’re running 100 CNC machines producing parts for major automakers. If software fails and machines go down, you lose billions. The quality bar industrial companies require is simply higher than what most software startups deliver.

Insight — The standard Silicon Valley GTM assumes you can build once and sell many times. Industrial environments invert this: every deployment is substantially custom. The economics never worked until AI changed the cost structure of specialization.

2. Why Now: The Economics of Specialization Finally Pencil

The Event — Philipp named the inflection point directly: until one year ago, specialized software was not economically feasible. The only company that cracked it was Palantir—but they had Peter Thiel’s capital and eight-figure contracts out of the gate. Everyone else was locked out of the mid-market.

What changed — Two things, both AI-driven:

First, AI coding tools made building bespoke integrations dramatically cheaper. The labor cost of customization—the thing that killed the unit economics of serving six and seven-figure contracts—dropped enough to make the mid-market viable.

Second, AI shifted from “tool” to “worker.” Pre-AI software was something humans operated to create value. Now AI can execute work autonomously. Industrial companies buy based on ROI, primarily cost reduction. A tool that requires skilled humans to operate doesn’t move the needle enough. A system that does the work directly does.

The mid-market gap — Palantir serves eight-figure contracts. SaaS companies serve with standardized products that don’t fit. The entire space between—what Philipp calls “the industrial backbone”—went untouched. Six and seven-figure contracts, companies too small for Palantir but too complex for horizontal software. That’s Nexxa’s market.

Insight — This is a classic “brownfield digitization” play. AI finally makes software economically viable for environments where the specialization cost previously exceeded the contract value. The same dynamic applies to restaurants, trucking, legal—thin-margin industries that software historically couldn’t profitably serve.

3. Tool vs. Work: The Value Creation Distinction

The Event — Philipp drew a distinction that matters: software before was a tool. You had to use the software to do something. AI can do something for the human.

Why this changes what industrial buyers will pay for — Industrial companies’ top line comes from hardware—machinery shipped, installed, managed. The cost side is humans, processes, workflows, paper, Word docs, Excel sheets. AI can’t easily impact top line immediately, but it can crush bottom line costs fast.

This is why the ROI conversation lands differently than typical enterprise software sales. You’re not selling productivity gains that require behavior change and adoption curves. You’re selling direct cost reduction through work automation.

The Palantir comparison — Someone in the chat asked how Nexxa differs from Palantir. Philipp’s answer: market segment and value mechanism. Palantir creates value through insights—data science, ontology, helping executives understand their business. Nexxa creates value through workflow automation—actually doing the work. And Nexxa targets contracts Palantir would never touch.

Insight — “Insight” vs “automation” is a useful frame for the entire applied AI space. Insight requires humans to act on it. Automation is the action. Industrial buyers, optimizing for ROI, will pay more reliably for automation.

4. The Forward Deployed Engineering Model (Done Right)

The Event — Philipp made a claim worth examining: “9 out of 10 companies that hire FDEs are actually hiring onboarding managers.”

His definition: true forward deployed engineering means 98% of the work is bespoke to the customer. If your product needs five business days of specialization, that’s onboarding. Not FDE work.

Why this distinction matters — The FDE model is expensive. You’re staffing engineers on-site for 6-10 weeks per deployment. That only makes sense if each deployment genuinely requires custom work that a product team couldn’t anticipate.

Nexxa’s platform, Nitro, is designed around this reality. Philipp described it as “a toolbox where you can build new tools very fast.” The FDE doesn’t just deploy pre-built solutions—they construct custom integrations on-site using a framework optimized for rapid specialization.

The three-step roadmap — Nexxa’s thesis on where this goes:

Integration — Connect legacy tools, add intelligence on top
Automation — Deploy agents and automations within existing stacks
Autonomy — AI models eventually understand legacy stacks well enough to rewrite and replace them (5-7 year horizon)

The foot-in-door strategy is integration. The long-term play is becoming the AI-native infrastructure layer as legacy systems get replaced.

Insight — The FDE model is high-touch and doesn’t scale like product. But if you’re right that each deployment is genuinely bespoke, there’s no alternative. The bet is that tooling (Nitro) eventually makes FDE work faster, and model improvements eventually automate parts of what FDEs do manually today.

5. The Technical Bottlenecks: Vision and Computer Use

The Event — Philipp flagged the two capabilities with highest potential that still don’t work well enough: vision models and computer use models.

Vision models — The example: upload a building floor plan to GPT or Gemini, ask how many rooms. Accuracy is 30-40%. The model miscounts rooms, misreads walls, fails at what seems like a basic spatial reasoning task.

This blocks massive automation potential. Industrial environments are full of technical diagrams—floor plans, rail systems, gas turbines, substations, P&IDs. If AI could accurately read and reason about these, entire categories of engineering review work become automatable. Right now, it can’t.

Computer use models — Industrial tools are built for humans. They don’t have good APIs or integrations. The natural way to automate work in these tools is through the UI—the same interface humans use. This is what computer use models promise: the logical extension of RPA (robotic process automation) that’s existed for two decades.

The problem: computer use models can’t be fine-tuned. All you can do is augment them with computer vision and prompt engineering. For custom tools the model has never seen, for complex engineering workflows, accuracy is still too low.

Insight — These are real technical constraints, not excuses. For builders considering industrial AI: don’t assume vision or computer use just works. The failure modes are specific and currently unsolved. Nexxa works directly with AI labs to get early access to releases, which suggests they’re betting on these capabilities maturing within their deployment timeline.

6. The Team as Moat

The Event — Philipp walked through five key hires and why each matters. The pattern: nobody came from a generic tech background. Everyone has specific industrial or adjacent experience.

The roster:

Philipp (CEO) — Ex-Siemens, ex-automotive, ex-management consulting. Worked in defense, manufacturing, rail, construction—always at board level. Speaks the language of industrial executives.
CTO — 10 years older, Sunnyvale native, master’s in cybersecurity. Knows on-prem deployment, data compliance, running AI in regulated environments.
Head of Creative — Came from music and media. Building what Philipp calls a “luxury tech brand”—differentiated experience in a space where B2B software is notoriously ugly.
Head of Go-to-Market — Two exits, always selling new technology into legacy markets. Ex-Siemens, ex-Airbus.
Chief Product Officer — From Shield AI, where he was VP of Product and Customer Engineering. Built their forward deployed team from zero.

Why this matters more than usual — Industrial sales are consultative. You’re not selling technology; you’re selling the value from technology. The executives buying need to trust that you understand their problems—not that you have a cool demo.

Philipp’s claim: “I’m not going to show you my brilliant demo of all the cool stuff you can do with my product. They don’t care.” Saying this to industrial executives is like “somebody finally gets my problem.”

Insight — Founder-market fit is overused as a concept but underweighted in execution. In industrial AI specifically, domain credibility isn’t a nice-to-have—it’s table stakes for the consultative sale. The team composition is the moat until the product matures.

7. The GTM Motion

The Event — Philipp described a classic consultative enterprise sales process, adapted for AI:

Free consultation — Qualify the opportunity
Paid 2-day on-site assessment — Examine tools, processes, workflows
Deliverable — 2-3 AI use cases with 7-8 figure impact each
Project — 6-10 week deployment with engineer on-site
Expansion — Trust built, leads to additional projects

Who they sell to — Business executives with P&L responsibility. Not CTOs, not CIOs, not VPs of engineering. GMs of billion-dollar business units who understand that AI can transform bottom line (and eventually top line).

Why not product-led — The sales motion mirrors the product thesis. If 90% of problems are specialized, you can’t demo your way to a sale. You have to diagnose first, then prescribe. The 2-day on-site isn’t a cost center—it’s how you earn the right to propose a solution.

Insight — This is Palantir’s model scaled down for mid-market economics. The risk is obvious: it’s labor-intensive and doesn’t scale like SaaS. The counter-argument is that for this market, there’s no alternative. The product-led motion that works in software doesn’t work when every deployment is bespoke.

8. The Pattern Across Livestreams

The Meta-Observation — This is the third recent livestream where the core problem is fragmentation and buried knowledge. Yi from Evercurrent talked about hardware teams losing context across siloed tools. Emilio from Wafer talked about GPU optimization requiring scattered profiling data. Now Philipp on industrial stacks being fragmented across ERPs, MES systems, Excel, and paper.

What this suggests — The current wave of applied AI isn’t primarily about model capability improvements. It’s about AI finally being good enough to integrate, synthesize, and act on information scattered across legacy systems that were never designed to talk to each other.

The constraint isn’t “can AI reason well enough?” It’s “can AI access and unify the information it needs to reason about?” The companies winning in applied AI are solving the integration problem as much as the intelligence problem.

The Bet

Nexxa is betting that:

The mid-market industrial segment is massive and underserved
AI economics now make bespoke work profitable at 6-7 figure contract sizes
Being first to build trust in this segment creates durable positioning
Vision and computer use models will improve enough to expand what’s automatable
Eventually (5-7 years), models can rewrite legacy stacks entirely

The risk is the high-touch model. Every deployment requires on-site engineering. That’s expensive, hard to scale, and depends on hiring people with rare skill combinations (AI engineering + industrial domain knowledge).

The counter: if specialization is genuinely required, there’s no shortcut. You either do the work or you don’t win the contract. And if you build the tooling (Nitro) well enough, you can make FDEs more efficient over time.

2026 goal per Philipp: stop singing their own praises, start having Fortune 100 customers tell the story. Case studies as the GTM lever. That’s the right instinct—in consultative sales, social proof from peers is worth more than any demo.

Nexxa website: nexxa.ai

They’re hiring — Toronto hub, Bay Area hub. Looking for ML engineers, computer vision specialists, and “founder mode” integration engineers who can build fast in messy environments.

Philipp Wehn: LinkedIn

Subscribe to support AI Made Simple and help us deliver more quality information to you-

Flexible pricing available—pay what matches your budget here.

Thank you for being here, and I hope you have a wonderful day.

Dev <3

If you liked this article and wish to share it, please refer to the following guidelines.

That is it for this piece. I appreciate your time. As always, if you’re interested in working with me or checking out my other work, my links will be at the end of this email/post. And if you found value in this write-up, I would appreciate you sharing it with more people. It is word-of-mouth referrals like yours that help me grow. The best way to share testimonials is to share articles and tag me in your post so I can see/share it.