This newsletter is brought to you with support from our featured partners: AMT, IMTS, Ironloop, Upkeep, and Industry 4.0 Club.

Putting AI in the Hands of Process Engineers🎙️

You scaled Boost AI to significant revenue before moving to industrial AI. What surprised you most about selling to factories versus selling to enterprise IT, and what unique Boost approaches have you found successful when selling to industrials?

At boost.ai, I spent seven years helping build one of Europe’s leading conversational AI companies. We grew from early-stage to a market leader, secured top-tier PE backing, and expanded internationally. That journey taught me a tremendous amount about scaling an AI company, but almost none of it prepared me for selling to factories.

The biggest surprise was how much the buying process is shaped by physical reality. In enterprise IT, you’re often selling to a CIO or a digital transformation team. The conversation is strategic and abstract - ROI models, integration roadmaps, vendor assessments. In manufacturing, you’re selling to people who can walk you to the production line and show you exactly where the problem is. A compressor that keeps tripping. A dryer that overheats. A quality deviation that costs them thousands per hour. The pain is tangible, and they need to see that your solution connects to their data, on their equipment, before they believe it.

What also surprised me is the trust dynamic. In enterprise software, trust is built through references, analyst reports, and procurement frameworks. In industry, trust is personal and earned over time. Plant managers want to know that you understand their operations, that you’ve been on a factory floor, that you won’t break something safety-critical. My earlier career at Emerson and NOV gave me a real advantage here because I already had that industrial language.

In terms of what carried over from boost.ai, the most transferable insight was the power of putting the tool directly in the end user’s hands. At boost.ai, we enabled customer service teams to build and manage their own virtual agents, freeing developers to focus on more complex integrations. At Intelecy, we do the same thing for process engineers: they build and deploy machine learning models themselves, while data science and IT teams can focus on the harder, bespoke challenges and ensure governance and security standards are met. That principle, that adoption scales when domain experts can act on their own knowledge, supported by technical teams rather than bottlenecked by them - has been just as powerful in industry as it was in enterprise software.

The vast majority of industrial AI pilots fail to reach production. What’s the single biggest reason you see projects die? Is it technical, cultural, or organizational? And how does Intelecy overcome those hurdles and get a 25-year veteran operator to trust a model they built in minutes over instincts they’ve built over decades?

The numbers are striking. Research consistently shows that the vast majority of AI projects - some studies put it at 80 percent or higher - never make it to production. And in manufacturing specifically, the failure rate is just as severe. But after working with industrial customers across energy, chemicals, water, mining, and food & beverage, I’ve come to believe the root cause isn’t technical. It’s structural.

Most industrial AI projects die in what I call the “handover gap.” A centralized team builds a model in a lab environment, hands over a proof of concept, and then moves on to the next initiative. But nobody on the plant floor was involved in building it, nobody owns it operationally, and nobody knows how to maintain or improve it when conditions change. The model sits on a shelf. It’s not a failure of the data science team - they did their part. It’s a structural problem: the people with deep process knowledge and the people with modeling skills are working in sequence rather than in parallel.

And then there’s the scalability problem, which is really the other side of the same coin. A data science team might build a brilliant model for one furnace or one production line - and it works. But that same organization has forty furnaces across five sites. Replicating that model manually, adapting it to each unit’s specific data, deploying and maintaining it - that doesn’t scale through a centralized team alone. With Intelecy, you can take what data scientists have already proven and scale it by a factor of 10 or 100 in a matter of hours, because the platform enables process engineers to replicate, adapt, and deploy models across assets and sites without having to start from scratch each time. That’s how you go from a successful pilot to real operational impact.

Intelecy addresses the handover gap by enabling process engineers to build models directly using data they already understand, while data science teams can focus on more complex, bespoke analytical work. Our platform is built for the front-line workers as it fundamentally changes who owns the day-to-day outcome. When a process engineer builds a model themselves, they can see which variables it uses, test it against known scenarios, and refine it as conditions evolve. That sense of ownership is everything.

As for the veteran operator, we don’t ask them to abandon their instincts. We position the model as a tool that extends those instincts. But here’s what I think is really underappreciated: those operators need time and space to actually think. What we see in so many plants is a culture of constant firefighting. People are running from one alarm to the next, fixing the same recurring issues, and there’s never time to step back and hypothesize about the root causes. That’s a tragedy, because those same operators often have the intuition to identify the big operational wins - the changes that would eliminate the fires altogether. Intelecy gives them that space. When a model monitors a process in real time and catches anomalies early, the operator isn’t scrambling to react; they’re free to ask bigger questions. Why does this keep happening? What if we adjusted this parameter upstream? That shift from reactive to proactive is where the real value lives. And once an operator sees that a model they helped build is catching things they would have missed during a night shift or a holiday weekend, trust follows naturally. It becomes a question of priority: do you want your best people permanently fighting fires, or do you want to give them the tools to prevent them? And there’s another reality we have to acknowledge: sometimes the people with the right process knowledge simply don’t have the time to work with the models themselves. In that case, they can hand that part over to us. We can build and maintain the models together with their operators, so they still bring the knowledge and judgment, without having to find capacity they don’t have.

That’s something we’ve recently started offering as Intelecy as a Service. It’s a slightly different way of thinking about industrial AI: the capability doesn’t necessarily have to sit entirely with the customer for them to benefit from it.

AI gets thrown around constantly today. How do you help customers parse the signal from the noise when choosing to work with Intelecy? And for a manufacturer still running on spreadsheets and tribal knowledge, what does the first 90 days with the platform actually look like?

The honest answer is: we start by being specific. A lot of vendors in this space sell “AI” as a broad promise - better efficiency, smarter operations, digital transformation. Those words don’t mean much to a plant manager who’s running a process 24/7 and needs to know whether a technology will solve a concrete problem. So we always anchor the conversation in their operational reality. What’s the biggest source of unplanned downtime? Where are they losing yield? What’s their energy cost per tonne? If we can’t connect AI to a measurable outcome they already care about, we’re just adding noise.

For a manufacturer that’s still heavily reliant on spreadsheets and tribal knowledge, the first 90 days are designed to be low-friction and high-impact. In the first few days, we connect to their existing data infrastructure - typically through our gateway that links to their control systems. There’s no rip-and-replace. We work with what they already have, inside the security and governance framework their IT team has established.

From there, we identify one or two high-value use cases together, often something like predicting a quality deviation or detecting anomalous equipment behavior. A process engineer who knows the process builds the first model on our platform, often within hours. That’s a powerful moment, because it replaces what would have been a months-long project with something the engineer built themselves in an afternoon, freeing up specialist resources for more complex analytical work.

By day 90, we typically have a model running in real-time against live production data, generating actionable insights that teams are already using. The spreadsheet doesn’t disappear overnight, but operators start to see that the platform gives them visibility they never had, and the tribal knowledge that used to live in one person’s head begins to get captured and scaled across shifts and sites.

Intelecy is known for streaming AI insights back into control systems for closed-loop automation, enabling adjustments in milliseconds. Where do you draw the line between letting the model act autonomously and keeping a human in the loop, and what change management is required to embed this into existing workflows?

This is one of the most important questions in industrial AI, and I think the industry broadly gets it wrong by treating it as binary - either the human is in control, or the machine is. The reality is much more nuanced, and the right answer depends on the decision’s risk profile.

At Intelecy, our platform supports a full spectrum. At one end, you have advisory models that surface predictions and anomaly alerts for operators to act on - a human sees the recommendation and decides what to do. At the other end, we can stream optimized setpoints directly back into DCS or SCADA systems for closed-loop control, where the model is making real-time adjustments. Most of our customers today operate somewhere in between, moving along that spectrum as their confidence grows.

The line we draw is informed by consequence. If the worst-case outcome of a model being wrong is a minor efficiency loss, there’s a strong case for automation. If the worst-case scenario is a safety incident, you keep a human in the loop and layer in constraints, guardrails, operating limits, and override mechanisms that the control system already enforces. We’re not replacing the safety logic. We’re adding intelligence on top of it.

The change management piece is significant and often underestimated. Technically, writing a prediction back to a control system is straightforward. Culturally, it requires trust - and trust is built incrementally. We typically recommend that customers start with an advisory phase: the model runs in parallel, operators can see what it would have recommended, and over weeks or months they develop confidence in its reliability. Only then do they move to closed-loop. Involving operations and control system engineers from day one is essential. If the people who run the plant weren’t part of the journey, they’ll resist the destination.

What’s the contrarian bet Intelecy is making that most industrial AI companies have wrong? And what are customers currently overlooking when it comes to scalable, impactful ROI from AI?

The contrarian bet is about where modeling capability should live in an organization. Most industrial AI companies are built around the assumption that machine learning must be centralized, that you need a dedicated data engineering pipeline, a specialist data science team to build the model, and a DevOps function to deploy and maintain it. That approach works well for high-complexity, one-off problems. But it creates an enormous bottleneck when you’re trying to scale AI across dozens of processes and sites.

Intelecy’s bet is that the people closest to the process, the engineers who understand the chemistry, the physics, the operating constraints, should be able to build and own the models for the routine, high-volume use cases: anomaly detection, quality prediction, energy optimization. And the platform should be sophisticated enough to make that possible without compromising on rigor. This doesn’t replace the data science function - it frees it up. When process engineers can handle the 80 percent of use cases that follow repeatable patterns, data scientists can focus on the 20 percent that genuinely require custom approaches. That’s a hard product challenge, but it’s the only path to scale. You can’t hire your way to deploying AI across dozens of sites. You have to enable the people who are already there.

As for what customers overlook, the biggest missed opportunity is thinking about AI as a project rather than a capability. Too many organizations treat it as a series of one-off initiatives: a pilot here, a proof of concept there. The real ROI comes when AI becomes part of how the organization operates day to day. That means investing in the platform, the people, and the governance to scale from one model to hundreds of models. The companies that will see outsized returns are the ones that stop asking “what’s the ROI of this one model?” and start asking “how do we build the muscle to deploy AI wherever it creates value?”

The other thing I see customers overlook is the data they already have. Most manufacturers are sitting on years of rich process data that has never been used for anything beyond basic trending and compliance reporting. That data is an extraordinary asset. You don’t need a massive AI infrastructure program to start extracting value from it - you need a platform that connects to it and puts it to work. That’s what Intelecy does.

To contact Camilla, reach out to her on LinkedIn here. She’s always open to chatting and sharing valuable insights.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Ironloop, Upkeep, and Industry 4.0 Club.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Ironloop, Upkeep, and Industry 4.0 Club.

Some Interesting Reads:

Anthropic released its Model Hardware Standard this past week. MHS lets AI agents run multiple lab and manufacturing instruments in parallel, such as microscopes, liquid handlers, and robotic arms, and perform intricate tasks ranging from routine drug discovery experiments to laser calibration on a quantum computer.

It’s interesting, but I think it will only further unlock and empower companies in the V&V space and other workflows that don’t touch product. If MCP drove agentic interoperability and pushed value toward those building workflows, MHS will do the same for companies building outside production. Nominal, for example. I don’t see any manufacturer ever letting this near real production workflows, not only for the lack of accuracy and the potential impact to downtime, but from a cybersecurity perspective as well. [Anthropic]

• Canada targets retaliatory tariffs on hundreds of steel and aluminum products [Manufacturing Dive]

• A pretty awesome guide for operations leaders on agentic AI [Arc Advisory]

• Another cyberattack, this time impacting Boston Scientific’s ability to manufacture products and to ship and process orders [Manufacturing Dive]

  • Here is a brief guide on OT Cyber from IIOT World

• US manufacturing is expanding, with 30,000 new jobs created due to demand for transformers, cooling equipment, and cement. [FT]

• John Deere increased its 2026 sales guidance by $250 million [Manufacturing Dive]

• Meta’s attempt to put robots inside data centers [Wired]

• An interesting perspective on why small batch manufacturing might be the key to reviving the domestic pharma ecosystem [Pharma Executive]

• CADGenBench Leaderboard [Hugging Face]

• Here’s an interview with the CEO of Generalist AI, Pete Florence [Joe Lonsdale].

• A pretty great interview with two of the smartest people talking about kernels, implicits, and CAD [X]

Blogs:

Some interesting takeaways from the Actuate conference from

Alpha Work

Actuate 2026 Takeaways

Last week I attended Actuate, a developer conference for the robotics and autonomy industry hosted by Foxglove, a data platform for physical AI…

Read more

11 days ago · 1 like · Kyle Getsiv

Scaling Robotics at Amazon with

The Industrial Experts

9. The Industrial Experts: Rajan Jha, Scaling Robotics at Amazon

Hello there - I’m Natan Reddy, a VC at Ironspring Ventures, where I invest in startups innovating across the industrial landscape (e.g. manufacturing, construction, transportation & logistics, and alternative energy). The most rewarding part of my job is learning from incredibly smart and talented founders, investors, and industry operators…

Read more

11 days ago · 3 likes · Natan Reddy

Products:

Skild has introduced in-context learning for robots where a single video prompt is enough to teach a new task, no fine-tuning required. It’s a real unlock, and one that should drive down the cost of deploying industrial robots while allowing far more flexibility across production workflows. [Skild]

It follows Generalist’s reveal of a multimodal model with a 30-second context window that combines video, language, and other sensor inputs to reach 59% accuracy on short-horizon atomic manipulation.[Generalist]

Together, these are robotics’ GPT-3 moment. Industrial work still demands reliability far closer to 99%, and getting there will require full post-training on task-specific data, but this is the clearest capability jump the field has had.

• Highbyte revealed agentic configuration capabilities for contextualizing industrial data [HighByte]

• OpenAI’s Jalapeño chip beat Nvidia’s GB200 and GB300 systems by 1.5x to 1.9x on throughput-per-watt and by up to 3.6x on latency across three open-weight models tested in-lab by [SemiAnalysis]

• Rockwell Automation and Augury are integrating their industrial AI and maintenance software platforms. [Assembly Mag]

Financing & Transactions 💵

Gatik raised $200M from QIA and Koch Disruptive Technologies
Physical AI / Robotics • Financing • USA

Cyclic Materials raised $75M from T. Rowe Price
Climate and Energy • Financing • Canada

Voya Energy raised $35M from Energy Impact Partners
Climate and Energy • Financing • USA

Transfyr raised $25M from General Catalyst and Lux
Physical AI / Robotics • Series Seed • USA

Agentrys raised $19.1M from Etna Labs
Design / Semiconductors • Series Seed • USA

Planned Downtime 🧑‍🔧

A deep dive with Heron Power, a company reinventing the transformer category [Weights and Biases]

Breaking the Bottleneck is brought to you with support from our partners.

• AMT & IMTS - The Association for Manufacturing Technology represents and promotes U.S.-based manufacturing technology and hosts the largest manufacturing technology trade show in North America, IMTS.

• Ironloop - Ironloop is a modern OT/ICS software platform that delivers configuration-level intelligence, providing manufacturers with ground-truth security visibility and confidence in compliance across air-gapped networks.

• Upkeep - CMMS Software helping maintenance teams manage work orders, assets, and preventive tasks efficiently from any device.

• Industry 4.0 Club - Leading cross-functional group accelerating the adoption of Industry 4.0, delivering better experiences to consumers, better profits to manufacturers, and better jobs to factory workers.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Ironloop, Upkeep, and Industry 4.0 Club.

Some Interesting Reads:

• A pretty awesome GitHub resource uploaded by Peter Corke, which covers a variety of topics: Kinematics + Dynamics, Inverse kinematics + Jacobians, Motion planning + control, Localization + SLAM, and has 50+ robot models.

• Cyberattacks in manufacturing and the lack of the CISO role [CyberNews]

  • A great breakdown of the ISA/IEC 62443 framework and where chemical manufacturers often fall short with existing risk posture [Automation.com]

• Excited to see the impact of Apple’s new Advanced Manufacturing Center offering free training and educational sessions for SMBs [Apple]

• The Ford Fathom is being built in Louisville at a 3-million-square-foot facility that was gutted and replaced with its Universal EV platform, which has separated component manufacturing, only for assembly to happen at the end, thanks to unicasting.

• Mesabi Metallics is building a new iron ore mine, and one of the first in the US in the last 50 years. [Industrial Sage]

• A perspective on humanoid deployments from Agility Robotics CEO, Peggy Johnson [WSJ]

• An interesting perspective from Bloomberg on the shifts in mentality that companies like Lear and Ford have had to talk to introduce more effective automation. [Bloomberg]

• McKinsey with a perspective on how legacy manufacturers can drive increased sales efficiency through pricing optimization (using AI to map product specs, find substitutable SKUs, and steer demand toward higher-margin options), automated lead generation and cross-sell mapping, and sales productivity (cutting administrative and quoting time so reps spend more of the day with customers). [McKinsey]

• Inside SK Hynix’s Factory [CNBC]

Blogs:

Manufacturing Talks with Jim Vinoski

Zamboni is Chapter 1 of *American Manufacturing* for a Reason!

Your host Jim Vinoski kicks off his series that will highlight each company in turn that he featured in his new book, *American Manufacturing: 22 Tales of Integrity, Ingenuity, and the Modest Heroes Who Built a Nation…

Listen now

22 days ago · Jim Vinoski

Products:

A deep dive with Rangeview, a company reinventing the casting industry [S3]

• Leo’s CAD assembly product that runs calculations, updates spec documents, and integrates into the company’s workflow before creating a 3D assembly using parts it creates or pulls from the user’s inventory. [Engineering.com]

• Path Robotics CEO on the new contract with HII [Path Robotics]

• Heron Power, a new advanced power electronics manufacturer, announced a new factory build [Heron Power]

• Fil Aronshtein with some great background on Dirac’s evolution into a AI Process Design platform [X]

Financing & Transactions 💵

Form Energy raised $750M led by T. Rowe Price
Energy • Series G • USA

Alloy Robotics raised $8M from Square Peg
Physical AI / Robotics • Series Seed • USA

Planned Downtime 🧑‍🔧

Rivian R2 Performance Test

Breaking the Bottleneck is brought to you with support from our partners.

• AMT & IMTS - The Association for Manufacturing Technology represents and promotes U.S.-based manufacturing technology and hosts the largest manufacturing technology trade show in North America, IMTS.

• Ironloop - Ironloop is a modern OT/ICS software platform that delivers configuration-level intelligence, providing manufacturers with ground-truth security visibility and confidence in compliance across air-gapped networks.

• Upkeep - CMMS Software helping maintenance teams manage work orders, assets, and preventive tasks efficiently from any device.

• Industry 4.0 Club - Leading cross-functional group accelerating the adoption of Industry 4.0, delivering better experiences to consumers, better profits to manufacturers, and better jobs to factory workers.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Ironloop, Upkeep, and Industry 4.0 Club.

Some Interesting Reads:

A nice breakdown on how to evaluate EDA tools enabling multiple types of PCBs beyond simple rigid boards. [Engineering.com]

• The Minnesota water systems cyberattack highlights the growing need for OT cybersecurity solutions. [NY Times]

• An interview with Adi Prasad, CEO of Matter, on what he learned from working for Elon and Travis Kalanick. [BI]

• MIT’s TechAMP makes a good case that America’s technology-adoption gap is an occupational-structure problem rather than a training-volume one. [Manufacturing Dive]

• An interesting article on how a streamlined path for behind-the-meter renewables and rate reform can reinvigorate the Industrial decarbonization in the US. [Mother Jones]

• How X Square built its robotic stack [IEEE Spectrum]

  

• The US banned foreign-made robots, creating a huge opportunity for companies like Standard Bots. [Morning Brew]

  • Here’s the official post from Brendan Carr [X]

  • Plus, the Standard Bots thesis from Robostrategy

• Olinia, Mexico’s $8,500 EV built for the dynamics that come with Mexican cities. [IEEE Spectrum]

• A neat 2016 MIT presentation on why Manufacturing is so difficult was found by !

• China has started to produce domestically developed immersion deep-ultraviolet (DUV) lithography machines, though some reports state that this is still four generations behind ASML's most advanced machine [Reuters]

• CATL has pivoted to AI, Grids, and ship production. I wonder if it’s tied to the slowdown in industrial output, which the FT revealed this week. [FT]

Blogs:

Breaking the Bottleneck

The CAD File Is Not a Static Artifact Anymore

🏭 Breaking the Bottleneck is a weekly newsletter and interview series covering manufacturing technology and physical AI. Want to chat? Reach out at aditya@machinafactory.org or connect with me on LinkedIn…

Read more

a month ago · 12 likes · Manufacturing Tech

Products:

Google announced Gemini Robotics 2, which is composed of a vision-language-action model (VLA) that converts vision and language input into motor control, an embodied reasoning (ER) model enabling an understanding of the physical world, and VLA optimized to run locally on robots.

• A look into NVIDIA’s World Model, Cosmos [NVIDIA]

• Arche Industries integrated hundreds of real catalog STEP parts directly into its Smith model, enabling customers to go from text to a structural and CFD-simulated part designed for manufacturing.

Financing & Transactions 💵

Antora raised $550M led by G2 Venture Partners and Eclipse
Energy • Series C • USA

ChipAgents raised $60M from B Capital
Semiconductors • Series A2 • USA

P-1 raised $50M from NEA
Physical AI / Design • Series A • USA

Foundational Industries raised $25M led by BoxGroup and Zigg Ventures
Autonomous Factories • Seed • USA

Intropy raised $11M from Felix Capital, Quiet Capital, and General Catalyst
Spare Parts Management • Seed • UK

Planned Downtime 🧑‍🔧

A Great Interview with Applied Intuition

Simulating Worlds with Fei-Fei Li

Breaking the Bottleneck is brought to you with support from our partners.

• AMT & IMTS - The Association for Manufacturing Technology represents and promotes U.S.-based manufacturing technology and hosts the largest manufacturing technology trade show in North America, IMTS.

• Ironloop - Ironloop is a modern OT/ICS software platform that delivers configuration-level intelligence, providing manufacturers with ground-truth security visibility and confidence in compliance across air-gapped networks.

• Upkeep - CMMS Software helping maintenance teams manage work orders, assets, and preventive tasks efficiently from any device.

• Industry 4.0 Club - Leading cross-functional group accelerating the adoption of Industry 4.0, delivering better experiences to consumers, better profits to manufacturers, and better jobs to factory workers.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Ironloop, Upkeep, and Industry 4.0 Club.

The CAD File Is Not a Static Artifact Anymore🎙️

You started your career as a Navigator and Gunnery & Missiles Officer, then spent two decades running digital transformations at Chase, OpenTable, and Yahoo before taking over Autodesk’s manufacturing business in 2018. Consumer internet and banking are slightly different than manufacturing, with software-native data and operations, and far lower integration and deployment overhead. What from those playbooks actually translates into how Autodesk approaches go-to-market, and what lessons could manufacturing software take from those markets?

I come from a family of factory workers and civil servants, so I had early exposure to how things get made and a real sense of wonder about making them. In its own way, a navy ship is a factory: steam propulsion, nuclear reactors, mechanical guns, radar, all of it running as interconnected systems.

Each industry I worked in taught me something different. Banking is about regulation and risk management; consumer businesses run on short cycles and the ability to move fast. But it is the commonalities that transfer. The first is ease of use. Nobody likes things that are hard to use, and running the mobile banking app at Chase taught me the relentless focus on simplification, making things easy for the customer. The second is efficiency: booking a restaurant table, ordering ahead, and the same theme kept recurring across every industry.

The third is a growing recognition of the importance of data. In banking, better data meant more accurate risk prediction and fraud detection. Manufacturing can benefit from this as well. A CAD file is no longer a static artifact; it is a living, breathing data system. Conceptual design, mechanical design, and manufacturing data are the foundation.

And the lesson beneath it all is that inertia can kill a company. The real question is how our software raises productivity relative to the old ways of working and the point solutions people have relied on. Consider what those cost today: data corruption, lost data, endless migrations between systems, or the sheer time spent producing 2D drawings, which can account for 40% of an engineer’s job.

You’ve framed AI as the third great shift in design and make, after parametric modeling in the ‘80s and Windows-based CAD in the ‘90s. For the median shop, the gap between paper-and-process and agentic orchestration with MCP integration is significant. How can machine shops and manufacturers begin that journey, and how do they balance this journey across different technical maturities, for example, design maturity with Fusion 360, with a legacy ERP for operations?

There is still some paper-and-pencil out there, but manufacturers have generally embraced technology, even before migrating to a modern stack. The gap you describe does sound intimidating; in my experience, the path is more practical than it looks.

It starts with data. Smooth, connected, flowing data is the currency that ties your entire process together, and that includes the processes themselves, the knowledge of how things actually get done. You cannot layer agentic AI on top of tribal knowledge.

That is worth stressing, because design and manufacturing are fundamentally different from text-first domains. They are more complex, and they demand precision. We spend a lot of time on how frontier models can augment design and make. Still, we also see their limits firsthand: to be useful on a factory floor, a model has to be trained specifically on real physics and real geometry. Engineers define intent that must hold up across geometry, tolerances, materials, toolpaths, and quality systems. Close enough is not good enough; it will break a machine and cost you a fortune.

So we usually begin with a more basic question: how do you connect your data? Most manufacturers are wrestling with fragmented legacy data, and much of our industry has built proprietary formats that do not work with anyone else’s systems. We don’t believe in that approach. What we have found is that AI becomes far more effective once engineering data has context and is connected across the product lifecycle, and, interestingly, AI can help get it there by interpreting and structuring legacy data and helping one system talk to another. I remember hard-coding systems to work together, and it was not that long ago; then came API connections, and now MCPs make it easier still.

That is really the answer to the maturity question. Interoperability is far easier in a more open ecosystem: you may have a legacy ERP and still connect to it. But I would add a caveat: if you are only using an MCP to push data into a legacy ERP, you may not be capturing the full advantage AI can offer.

Then it comes back to ease of use. We show customers how simple this is to adopt, and we hear it constantly with Fusion. Unlike some legacy CAD tools, you don’t need two decades of experience with Windchill, Teamcenter, or Enovia to make it work. From there, we talk to manufacturers about modernizing their end-to-end system so they can capture the full competitive benefit. Fundamentally, manufacturers that standardize and operationalize their engineering and design data will be best positioned to leverage AI as a competitive advantage.

One thing did surprise me. I think constantly about the mid-market, and I came in wondering how sophisticated those manufacturers would be. I have been struck by how much they have embraced technology; this is not my parents’ generation. But they do need help, and they need easy tools that connect everything for them.

In your February 2025 testimony to the House Energy & Commerce Committee, you argued that domestic reindustrialization fails unless manufacturers of all sizes become digitally enabled. The political reality is that reshoring dollars and ribbon-cuttings gravitate toward buildings, tooling, and steel. If the next wave of incentives once again funds machines but skips the “digital plumbing,” what more is needed to drive cost-competitive reindustrialization? And what role have you seen SBA loans play across the SMB ecosystem in reindustrialization so far?

First and foremost, we have to invest in modernizing our design, data, and manufacturing systems, both the software and the hardware. I have said the same to a room full of chief information officers in India: you cannot simply throw bodies at this problem. If you are not modernizing your hardware, your software, and your processes, you will fall behind.

You are pointing to two problems we have been trying to reconcile. The first is that if reindustrialization reaches only the top of the pyramid or only certain sectors, we do not actually rebuild the industrial base, which is a long-term danger. The second, and more common, is that everyone thinks about hardware: let us offer incentives to go and build. But without rethinking your processes and your software, you will not be competitive.

Small and medium-sized businesses are not a nice-to-have here. They make up around 90% of manufacturing firms. They are the backbone of any manufacturing economy, not only in America, and a supply chain is only as strong as its weakest link. So, to your framing: if the next wave funds machines and skips the plumbing, we will have a problem.

That is why we are constantly reminding lawmakers that our global competitors are not investing solely in physical factories. We need incentives aimed specifically at digital transformation and AI adoption among smaller manufacturers. Loans from the Small Business Administration are one instrument; there are also tax credits, matching funds, and other targeted measures. Larger manufacturers can benefit too, though they already enjoy some advantages of scale.

In some ways, AI can be a great equalizer, giving a small shop access to capabilities that were once the preserve of large companies with deep pockets. As I mentioned, connecting old legacy data to a modern system, something that used to take a lot of people, is now far easier with MCPs, which can reduce the integration required.

SBA loans are an important source of capital for small manufacturers, but far from the only one. The very fact that we were invited to Congress to discuss competitiveness was encouraging, but we still need programs at both the national and state levels to expand access to low-cost capital. I said something in Congress that I firmly believe: manufacturers, certainly in America, and I suspect more widely, do not want handouts. They want to compete. Give them access to capital so they can invest and stay competitive. And these programs pay for themselves: a loan is not a grant, so it is repaid, there is a return on the investment, and firms then pay more in tax as they succeed. A broader, more resilient manufacturing base also brings a halo effect of jobs and domestic sales.

If I were king of the forest, I would set up directed programs, “directed” in the sense that firms choose from a menu. Perhaps a company has already modernized its software; perhaps it has just built a new factory and realized it must upgrade that software; perhaps what it really needs is skilled people. Let manufacturers decide what they need, but give them a menu, because success requires investment across all of these. Imagine a state or federal program that said: to compete in manufacturing, you will need to modernize your digital plumbing; you may need a new or retrofitted brownfield factory; and you may need to train your workforce. Here are three programs that support exactly that, choose what you need and apply.

Unifying design and manufacturing is a promise PLM vendors have been making for 25 years. How are you embedding AI? What changes are you making across the stack? Furthermore, how is Autodesk positioning itself with the MaintainX acquisition to further integrate manufacturing data upstream into the design process?

We have not closed the MaintainX deal yet, so I will keep that part hypothetical. But let me start with the premise, because I do not think PLM systems converge design and manufacturing on their own. They are one part of the process. We think in terms of CAD, CAM, PLM, and supply chain, and how all of those systems fit together. PLM performs a critical function, data and change management across the product lifecycle, but on its own, it stitches data together rather than truly converging it. What we are striving for with Fusion and our industry clouds is for CAD and CAM to communicate natively.

We believe in a far more integrated industry cloud that goes beyond point solutions. Those point solutions have served our industries extremely well for decades. American manufacturing was built on them, and they still work for many people. But manufacturing productivity gains have been plateauing, and to reach the next level of breakthrough, we believe you need AI within a converged industry cloud. It is not PLM alone.

So we embed AI at every step, CAD, CAM, and PLM, and we think about it on three levels. The first is task-based automation: we look for individual tasks that could be made more efficient. Automated drawings are a good example. We rolled that out a couple of years ago, using AI to spare designers and engineers the time required to produce them; we have done the same with sketch constraints, and we recommend machining strategies from a trained CAM model. These are the tasks customers tell us are low-value and inefficient.

We are now moving into workflow-level automation, using AI to power an entire workflow. One of the most exciting is how a design begins: going from a text description to 3D geometry, sketch to mesh to B-rep, invisibly to the user. I think AI will be game-changing in PLM in particular, because PLM systems have been so rigid and so dependent on tribal knowledge and bespoke rules. AI should be able to simplify the interface, capture the intent a user would otherwise enter by hand, understand it, and effectively build the PLM system for them, without anyone ever having to log in.

Last year, we ran an analysis with a large consulting firm and surveyed hundreds of PLM customers of all sizes. To a one, they hate their PLM system; switching it on is the worst part of their day. It should not be that way. They understand how vital lifecycle management is; the interface is cumbersome. You will hear more from us on this in the autumn.

The third level is system automation. We envisage a world in which entire systems are connected through AI, so that you can take a design change you are contemplating and predict its ramifications across the factory, across operational processes, even down to the logistics of the supply chain and what is actually available. AI connects the whole system and gives designers, manufacturers, and engineers visibility into the implications of their decisions at every stage.

On operations, as Andrew put it on our earnings call, we have long thought about the “design and make” process, but “operate” was always the logical next step, and it cuts across our architecture and construction business as much as manufacturing. People operate large buildings; people operate factories. When we can capture data coming off the factory floor and feed it back into manufacturing to make that system smarter, then back into design to make that smarter, with information flowing both ways, the value is immense. Some of our manufacturing tools have already moved into the operations group: we have done factory design work with Inventor for years, and we acquired FlexSim, which builds digital twins of a machine setup and makes a factory much easier to design.

Where this really compounds is context. We have been building a context graph: picture a single customer, every product they design, all the accumulated knowledge in their processes, and what does and does not work for a given product. We can capture that as a time series of context around design, manufacturing, processes, data, and what the customer cares about. With design and manufacturing in place, and now operations too, you build a far richer body of context, and it becomes much more feasible to interpret and consume it through a model tuned to a specific task. Large language models have none of that. Like everyone, we are impressed by what they can do; we also see firsthand where they struggle, and we think hard about how to use their strengths while giving customers what they actually need. When we apply that across multiple industries, we will unlock the power of AI throughout the economy.

As AI takes on tasks like drafting and G-code generation, you’ve argued that the role shifts from “button pusher” to problem-solver and systems orchestrator. Yet the upskilling and change management these roles demand is often very difficult, given entrenched ways of working, time constraints, and a changing workforce. How will jobs change with AI, and what steps should manufacturers take to overcome some of these operational hurdles?

There are two challenges here, perhaps three. One is the aging workforce; the second is upskilling. The third, which is really the backdrop, is something Andrew has said many times: there is a large backlog of work across many of our industries. Infrastructure projects are delayed for want of capacity; manufacturers would love to produce more but cannot, and construction is riddled with the same constraint.

Our view is that AI can augment these processes and unlock more capacity to get things done. In our industries especially, there is pent-up demand, and we see AI as a way to meet it despite the aging workforce, but it will require upskilling people. On our side, that begins with embedding AI in our products, so that adopting our tools makes people faster at the work they are already doing and helps clear the backlog.

For the next generation, we give our software away free to universities, vocational schools, and secondary schools. We believe part of our role in society is to train young people on modern tools so that the economy stays competitive, and we invest heavily there. We have now committed $350m over three years to provide access to technology, training, and certification for the AI-powered jobs that design and build the physical world, a continuation of what we have done for a long time. Part of that, $10m, is pledged to the RAISE program, a bipartisan effort to improve the training of future workers. It is a small fraction of the total, but it all adds up.

Specifically on the aging workforce, the single most important thing I tell manufacturers is to start documenting their processes. A great deal of knowledge is locked up in retiring senior people, and we can now capture it. One MIT professor I shared a panel with described using visual-learning systems to record what people do on an assembly line and learn from it, so that a new hire who lacks thirty years of experience still gets the benefit of that experience, because it has been captured and trained into a model. It starts with manufacturers deciding to document the existing processes they want to keep and then using AI to train new workers on them.

On change management, we try to make our AI tools intuitive enough that people can pick them up and take advantage of the task, workflow, and system automations, and we have people on our team who guide customers through deploying the software. But the single biggest factor is simply getting people to experiment. Even our own software engineers, some of whom have built the same products for ten or twenty years, are now having to learn entirely new ways of working, and what I have found is that once people start trying things, the capabilities tend to blow them away. One colleague who has been here for twenty years and in the industry far longer told me this is the most exciting period he has known in technology. Still, it began with him embracing the change and then getting out to experiment. He now does his own G-code generation on a little machine shop at the back of his house, testing and running it himself. I tease him that I worry about his marriage. That kind of enthusiasm is contagious. As with any change management, even before AI, you first have to get people to embrace the idea of change and the better outcome it can bring. Once they do and start experimenting, it becomes exciting and contagious.

________________________________________

Note: Edited and Paraphrased for Length

To contact Jeff, reach out to him on LinkedIn here. He’s always open to chatting and sharing valuable insights.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Ironloop, Upkeep, and Industry 4.0 Club.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Ironloop, Upkeep, and Industry 4.0 Club.

Some Interesting Reads:

A great profile on ASML’s new “high-NA” machine, which raises the numerical aperture of its optics (from 0.33 to 0.55) rather than switching to a new form of light. This lets it print features down to eight nanometers with three times the transistor density. [MIT Technology Review]

• TSMC is accelerating its build-out of its Arizona fabs because of the enormous demand for AI and plans to raise prices as well [CNBC]

• If you are in NY this week, stop by this awesome event hosted by Manufacturing Happy Hour in Brooklyn.

• Michael Kratsios released his version of Vannevar Bush’s Science: The Endless Frontier, called Science, A New Golden Age. [White House]

• A great perspective from Kyle Chan on the real end goal of Chinese subsidies with the country rolling back tax breaks on lithium EV [X]

• McKinsey’s perspective on Agentic AI across the Supply Chain [McKinsey]

• Psi Quantum, a company developing utility-scale, fault-tolerant quantum computers using silicon photonics and their factory. [Sachin and Adam]

• Air Liquide commits over $530M to U.S. Gas Plants [E+E Leaders]

• An interesting conversation with the VP of Advanced Manufacturing Innovation at Magna on AI and robotics in manufacturing. [RoadtoAutonomy]

• A new executive order preventing defense contractors’ waiver access to procure critical minerals and components from China [Manufacturing Dive].

• Meanwhile in Europe, China has started to wield its rare earth control over the chemical, electric vehicle, and defense manufacturers [NY Times].

  Blogs:

WattFactory

The Variety Trap

The West’s manufacturing strategy is changing…

Read more

a month ago · 8 likes · 1 comment · Glen Turley

Carlos Furlotti

Thoughts and Highlights on Building Vertical AI by BVP

Most AI playbooks were written for companies that don’t have to worry about what happens when their software fails. Industrial AI is different…

Read more

2 months ago · Carlos Furlotti

Products:

Generalist revealed its latest embodied foundation model, supporting a broad range of end effectors from 5-finger hands to specialized tools. Each hand is a different sensorimotor interface capturing geometry, contact, friction, forces, and dynamics into an end effector-specific language for physical interaction. The team was even able to modify the hands mid-rollout and let the same model keep running to great effect. [Generalist]

• Rockwell Automation and Augury partner [ARC Advisory]

• AMD unveils Kria module for real-time control, unified memory for robots [Robot Report]

• Really impressive from Black Forest Labs deploying video models in real factory use cases for Audi. The tasks spanned kitting parts into structured trays, inserting electronic control units into tight-fitting fixtures, assembling components together, and handling soft, flexible materials like seals and cables.

Financing & Transactions 💵

Atoms raised $1.7B from a16z
Robotics and Physical AI • Financing • USA

Sila Nanotechnologies raised $300M led by Atreides and Sutter Hill Ventures
Energy and Batteries • Financing • USA

Arrakis raised $30M led by Blossom Capital
Manufacturing AI • Series A • Europe

Planned Downtime 🧑‍🔧

The Secret of the Odyssey

Breaking the Bottleneck is brought to you with support from our partners.

• AMT & IMTS - The Association for Manufacturing Technology represents and promotes U.S.-based manufacturing technology and hosts the largest manufacturing technology trade show in North America, IMTS.

• Ironloop - Ironloop is a modern OT/ICS software platform that delivers configuration-level intelligence, providing manufacturers with ground-truth security visibility and confidence in compliance across air-gapped networks.

• Upkeep - CMMS Software helping maintenance teams manage work orders, assets, and preventive tasks efficiently from any device.

• Industry 4.0 Club - Leading cross-functional group accelerating the adoption of Industry 4.0, delivering better experiences to consumers, better profits to manufacturers, and better jobs to factory workers.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Jiga, Upkeep, and Industry 4.0 Club.

Some Interesting Reads:

Parsec launched its State of the Manufacturing Industry report, interviewing 1,200 manufacturing leaders across a variety of personas. Here are some of the highlights. [Parsec]

• 37% have a unified, data-driven strategy in place, but 60% are in the implementation or planning phase

• 72% of manufacturers have adopted AI in some form (up from 53% in 2024), with 10% at scale across their operations, 22% actively implementing, and the remainder piloting or in early use

• Top AI use cases include quality control (50%), IT operations (46%), and supply chain management (45%)

• (70%) have completed or are in the process of reshoring; just 12% have no plans to reshore, a significant difference from 2024, when those numbers were 33% and 38%, respectively

• "Supporting long-term success will come down to two key pillars: people and technology. Leaders should invest in workforce capabilities and support upskilling wherever possible. A strong data foundation, a robust execution layer, and a focus on scaling will drive technology success

  

• Some recommendations from Bain on scaling production in Aerospace and Defense. [Bain]

• A case study on tariff efficacy with American Steel [WSJ Opinion]

• Hitachi Vantara achieved a 50% reduction in inventory and a 77% reduction in order-to-ship lead times at a US storage array manufacturing site by using its internal HMAX agentic AI suite. [Blocks and Files]

• The Tesla Semi Factory [Core Memory]

• FedEx CEO Raj Subramanian is seeing operational flows rewrite themselves, calling the current tariff-driven trade reconfiguration the most significant shift he has witnessed in his career. [Fortune]

• The nearly $1.8T in US factory investment announced since January 2025 is paired with a labor market without workers. [Manufacturing Mag]

• The bottleneck in alloy forging and casting across components has caused the persistent gas turbine shortage. [WSJ]

• A breakdown of automotive companies and their humanoid bets [X]

• According to the IEA's Global Critical Minerals Outlook, if China's export controls are fully enforced, manufacturers in several strategic sectors could face significant supply disruptions affecting around $6.5 trillion in downstream value. [IEA]

  

• The impact of a failing American EV ecosystem on the broader automotive industry. [NYTimes]

• A great breakdown of the emerging yet massive chip ecosystem from at Chipstrat.

  Chipstrat

  The Next Trillion-Dollar Chip Company

  Two AI chip startups had huge financial outcomes in the past year. Groq was acquired for $20B, and Cerebras IPO’d to a $40B+ market cap. Huge, but not trillion-dollar huge. And in this environment, an inference-first system that runs frontier models and scales to gigawatts of compute has a real shot at a trillion-dollar market cap if it can scale, grow, and IPO…

  Read more

  2 months ago · 31 likes · Austin Lyons

Products & Research:

Upkeep launched Learn, a full AI course-creation feature that builds fully tailored training material (script, slides, video, voiceover) from the customer’s actual SOPs and related material. “The companies with operational excellence are connecting their institutional knowledge with what they’re actually doing on the floor. UpKeep Learn helps teams bring training into the same system where maintenance and safety work already happens.” The company is also hosting a webinar on SOP Automation, and AI-generated training grounded in actual procedural workflows.

• Autodesk announces AI features embedded across Fusion to help teams automate a series of tasks. [Autodesk]

• Fulcrum launched Archie, its native AI platform, to let shop owners ask questions, build custom apps, and deploy autonomous agents. [AdvancedManufacturing.org]

• Xometry launched a new AI architecture designed to improve process recommendations, pricing accuracy, and supplier matching. [Xometry]

• An interview with Overview AI CEO, Christopher Van Dyke. [Vision Systems]

• Walden Robots launches, combining Large Behavior Models with real-world operation to drive quicker industrial adoption. [Walden Robotics]

• Archer announces Zee, an aviation-specific foundational model built on ADS-B, ATC communication, terrain, aircraft state, and weather data. [Archer]

Financing & Transactions 💵

Walden Robotics raised $300M from Toyota and Deviation Capital
Robotics and Physical AI • Seed • USA

Astromechanica raised $250M
Aerospace and Defense • Series B • USA

Senra Systems raised $65M from Lowercarbon Capital and Interlagos
Wire and Harness Manufacturing • Series B • USA

Microagi raised $55M from Hummingbird Ventures
Robotics & Physical AI • Seed • Germany

Hyperion Robotics raised $7.4M from Course Corrected
Robotics & Physical AI • Seed • Finland

Planned Downtime 🧑‍🔧

World Cup Final Highlights

Breaking the Bottleneck is brought to you with support from our partners.

• AMT & IMTS - The Association for Manufacturing Technology represents and promotes U.S.-based manufacturing technology and hosts the largest manufacturing technology trade show in North America, IMTS.

• Ironloop - Ironloop is a modern OT/ICS software platform that delivers configuration-level intelligence, providing manufacturers with ground-truth security visibility and confidence in compliance across air-gapped networks.

• Jiga - Is a platform that connects hardware teams with vetted manufacturers for custom parts. Instead of black-box instant quotes, you see exactly who’s making your parts and can talk to them directly.

• Upkeep - CMMS Software helping maintenance teams manage work orders, assets, and preventive tasks efficiently from any device.

• Industry 4.0 Club - Leading cross-functional group accelerating the adoption of Industry 4.0, delivering better experiences to consumers, better profits to manufacturers, and better jobs to factory workers.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Jiga, Upkeep, and Industry 4.0 Club.

UNS Only Solved Half the Problem🎙️

Inductive Automation's origin runs against many of the conventions the venture industry proposes for the manufacturing software ecosystem. How did you and the founding team think about the early versions of FactorySQL and FactoryPMI, the transition to the modernized standard in Ignition 8.3, and i3X today? What were some of the key lessons you learned building a business in the industrial software category?

Inductive Automation was not built according to any conventional playbook, and that was entirely intentional. Steve Hechtman spent more than two decades as a systems integrator before founding the company in 2003, and the problems he was solving with FactorySQL and FactoryPMI were ones he had lived personally. The first version of FactorySQL was a straightforward OPC-to-SQL data bridge: get the data off the PLC and into a relational database so you could actually do something with it. FactoryPMI followed as a Java-based visualization layer on top of that data. Both were solving a real friction point that every integrator in the industry understood viscerally.

What made those early products unconventional was the underlying philosophy, and it went beyond licensing. At a time when every major industrial software vendor was charging per client, per tag, and per screen, and building on proprietary stacks, Steve built an unlimited-licensing model sold by the server and grounded the entire platform in open IT standards. Using SQL databases for industrial data, leveraging Java for cross-platform compatibility, embracing open protocols rather than building proprietary ones: these were not obvious choices in 2003. The OT world was deeply skeptical of IT practices on the plant floor. But Steve came from the integrator world, and he had lived the pain of systems that couldn't talk to each other, couldn't scale without a purchase order, and couldn't be maintained by anyone outside a narrow specialist community. The industry said the unlimited model wouldn't work. Competitors at trade shows told us we were doomed. But we understood something they didn't: when you remove the economic and technological barriers from a powerful platform built on open standards, you unlock innovation you couldn't have anticipated. Integrators could just say yes to customers. They could experiment. They could build systems that weren't constrained by the fear of overages. That thinking has never changed.

In 2010, consolidating FactorySQL and FactoryPMI under the Ignition platform was driven by a specific technical opportunity and a clear strategic goal. When OPC UA emerged in 2009, it gave us the foundation to become fully cross-platform. We rewrote everything in Java, which meant Ignition could run on Windows, Linux, and Mac from a single codebase. That was genuinely unusual for industrial software at the time, and it removed an entire class of infrastructure constraints for our customers. Ignition 8.3, which we released at ICC in September 2025, is our most substantial release to date. Beyond the architectural improvements to the gateway, historian, and security, it delivered something the industrial software world has largely lacked: first-class support for version control, DevOps workflows, and infrastructure as code. Engineers can now treat Ignition projects the way modern software teams treat any other codebase, with branching, pull requests, and repeatable deployments. Colby described it publicly as a new beginning, not a destination, and the move to an annual major-release cadence, starting with Ignition 2027, reflects the same philosophy: industrial software has to move at the pace of the industry it serves.

The biggest lesson across all of it is that building in industrial software requires a long time horizon and an obsessive commitment to openness. Every time we bet on an open standard, SQL, OPC-UA, MQTT, Python, we were rewarded, and more importantly, so were our customers. Open standards compound in value over time because they allow customers to build on a foundation that is not controlled by any single vendor. Today, we are seeing that play out in customer procurement decisions. Organizations are explicitly refusing proprietary architectures. They want systems that can be maintained by a broad talent pool, extended without vendor permission, and integrated without custom middleware. That is a cultural shift that took years to arrive, and it validates everything we built toward from the beginning.

You've been one of the most visible champions of the Unified Namespace as the single source of truth for the connected factory, but i3X is, in a sense, an admission that UNS only solved half the problem. So what did the industry get wrong about UNS that i3X is now correcting, and given the long graveyard of competing industrial standards, what stops i3X from becoming just another API or endpoint folks migrate to? How do you ensure it drives measurable operational value on the factory floor?

The Unified Namespace was a genuine architectural breakthrough, and we have been among its loudest advocates because it fundamentally changed how people think about data organization in industrial environments. Moving from point-to-point spaghetti integrations to a centrally organized, context-rich data bus was the right answer to a real problem. But it is worth being precise about what UNS solved and what it did not, because those are actually two different layers of the problem.

The industrial data stack has always relied on specialized standards, each doing one thing well. OPC-UA is the right answer at the device level: it provides secure, reliable, semantically rich communication between controllers and SCADA systems. MQTT with Sparkplug is the right answer for publishing context-rich data using report-by-exception, enabling scalable architectures that efficiently connect remote sites and deliver data to the cloud without flooding the network. Each of those standards did its job well. What was missing was a standard for what happens when an application wants to consume that data in full, not just the live values, but the historical record, the object relationships, the type definitions, and the contextual metadata, all through a single interface. That is the gap i3X fills.

i3X is a REST-based industrial interoperability API developed through CESMII alongside many of the leading industrial hardware and software companies in the world. It can discover namespaces, explore object types, query real-time and historical data, navigate relationships, and subscribe to events, all through a single standardized interface. What is significant about that is that before i3X, those capabilities lived in separate places. Live data came through one standard, historical data through a historian API, and relationships and context through a separate data model layer. An application that needed all of it had to build integrations against each of those independently. i3X collapses that into a single surface, which means Ignition can expose the full power of its tag database, including UDTs, alarms, history, and relationships, to any application that speaks i3X. We are not replacing the standards below it. OPC-UA and MQTT still do what they do best. i3X is the application-facing layer that sits above them, making the whole stack consumable.

There is also a supply chain dimension here that does not get enough attention. When industrial data can be shared through a standardized API, it opens the door to structured data exchange across organizational boundaries. A manufacturer sharing operational data with a supplier, an OEM sharing machine data with an end customer, a system integrator delivering a project with standardized handoffs: all of these become dramatically simpler when the consuming interface is a standard rather than a custom integration negotiated on every engagement.

Your question about the graveyard of failed standards is a fair one, and I take it seriously. The reason most industrial standards fail is that they are developed in standards bodies removed from commercial reality, or they get adopted at the specification level but never at the deployment level. i3X avoids those failure modes because every member of the working group is a company with real deployments and real customers who need this to work, and because we are not asking customers to adopt an abstract specification. Native i3X support will be built directly into Ignition. Customers get a working implementation within the platform they already run, which bridges the gap between a standard that exists on paper and one that actually shows up on the factory floor.

The MCP module is your productized answer to bringing LLMs onto the factory floor: read access to live tags, alarms, and history, with hard guardrails against write-back to the PLC. How do you see this impacting the OT technical stack, and where do you see the value evolving as the agent, alongside a human, becomes the user? Furthermore, how does Ignition stay the platform rather than getting commoditized?

The MCP Module is the most important thing we have announced since Perspective, and I want to be precise about why. It is not because AI assistants on the factory floor are interesting as a novelty. It is because MCP is another open standard, and open standards are what allow people to innovate without asking for permission. The more open standards we build into Ignition, the more possibilities the platform unlocks. Every time we add a standard, SQL, OPC-UA, MQTT, Sparkplug, and now MCP, we expand the universe of things customers and partners can build on top of Ignition without requiring custom integration work. Systems just work together. That is the core value proposition, and MCP extends it into the AI layer in a way that no one else has yet.

What we are building does two things. First, it allows Ignition and its modules to expose tags, alarms, historian data, configuration, and design capabilities via the MCP protocol, enabling any AI system to access them in a standardized way. Second, and this is the part that matters most for the ecosystem, it makes Ignition the easiest platform for users and integrators to build their own MCP servers. You define tools, resources, and prompts inside the Ignition Designer, author tools in Python, and connect AI agents to domain-specific OT data and workflows without writing integration code from scratch each time.

The guardrails question is the right one to ask. We are starting with read access to live tags, alarms, and history, and the boundaries around write-back to the PLC are intentional and firm. Industrial safety is not a product feature; it is a constraint we design around, not through. As the agent model matures and as the industry develops the governance frameworks to support it, that boundary will evolve. But we are not going to rush it. The "ABCs" that our team articulated at ICC 2025- accuracy of information, being mindful of data access, and coordinating actions with a human in the loop- reflect how we think about the near-term deployment model: AI as a powerful co-pilot for the operator, not as an autonomous actor.

As for how Ignition stays the platform rather than getting commoditized, the moat is not any single feature. It is the combination of depth in the OT domain, the trust of 69% of the Fortune 100, an integrator network of over 5,000 companies, a free training university with over 10 million video views, and a licensing model that invites experimentation rather than penalizing it. AI companies building industrial applications need a platform that speaks the language of PLCs, historians, alarms, UDTs, and OPC-UA. That is not a generic software problem. That is Ignition's problem to solve, and we have been solving it for over two decades. Ignition 8.3 was explicitly designed with the architectural elements needed to take advantage of MCP and whatever protocols come after it. We are not retrofitting AI onto a legacy platform. We built the foundation with this in mind.

You've built a deliberate partner architecture with the Ignition Technology Ecosystem Program across hardware, software, and systems integrators over the past couple of years. From where you sit, what actually drives pull-through for a partner? What separates the ones that turn the Ignition install base into a real distribution channel from the ones that get a logo on a slide and nothing else? How can startups leverage your ecosystem?

The honest answer is that the partners who turn the Ignition install base into a real distribution channel are the ones who show up with a clear answer to a specific customer problem and a genuine willingness to go deep. The presence in the Technology Ecosystem Program and the listing on our site are earned when you can demonstrate the "better together" story and show how your product works with Ignition in a real deployment. That is the baseline. What separates the ones who build real business from the ones who stay at the logo level is whether they have a real problem to solve and whether they are willing to do the work to solve it for our shared customers.

The program reflects this with its tiered structure: registered, verified, gold, and premier. The higher a partner goes, the more value they deliver to customers, and the more Inductive Automation invests in making that partnership visible and effective. That structure is not bureaucratic; it is how we distinguish depth from surface-level integration.

The ecosystem spans three broad categories. Hardware companies that Ignition runs on, software companies that integrate with Ignition, and companies that have built modules directly on the Ignition platform using our SDK. Each category has a different motion and a different value story, but the common thread is that the best partners in every category are thinking about the customer outcome first.

Sepasoft and Kanoa are good examples of the module model done right. Both built MES capabilities natively on top of the Ignition platform using our SDK, which means their customers get a unified experience rather than a bolt-on integration. Integrators who are already doing SCADA work for a manufacturer can add MES functionality from either company as a natural extension of the project they are already running. The value is clear, and the delivery path is already established. That is what pull-through actually looks like.

On the hardware side, Opto 22 is a strong example of a different kind of depth. They make devices that embed Ignition Edge directly, which means a single piece of hardware can bridge the greenfield and brownfield worlds. A customer deploying at a remote site or retrofitting legacy equipment gets a device that is secure, reliable, and already running Ignition at the edge. The integration story is not something you have to figure out; it is built into the product. That is a fundamentally different conversation than showing up with a data sheet and a logo.

Barbara is another example worth calling out in the software layer. They tackle one of the harder problems in OT deployments: managing and orchestrating Ignition across distributed environments. As customers scale from one site to dozens or hundreds of sites, the operational complexity of managing Ignition instances grows significantly. Barbara brings a DevOps and infrastructure-as-code mindset to that problem, which aligns well with where 8.3 is taking the platform. For startups, the lesson from all of these examples is the same: find the specific place in the Ignition deployment lifecycle where your product creates clear, demonstrable value, build the integrator relationships to deliver it, and earn your way up the program tiers from there.

Finally, if i3X and MCP both land the way you're betting, and LLMs continue to evolve, how does the ecosystem evolve? Is there a world where drivers can be built quickly, the middleware layer is largely commoditized, and value accrues only to the application layer? Furthermore, looking out five years, what does the Ignition and Inductive Automation product suite look like, and what role will it play in the agentic industrial stack?

If i3X and MCP both land the way we are betting, you will see a meaningful compression of integration cost at every layer below the application. Drivers will be faster to build because the data models are standardized. Middleware will face real commoditization pressure because the value of a custom integration layer diminishes when both sides of the integration speak a common protocol. That is not speculation; it is the same dynamic we saw play out with SQL and then with MQTT. The abstraction layers that once required significant expertise to build become utilities, and value migrates to whoever is closest to the outcome.

Where I think people get the analysis wrong is in concluding that this means the platform layer also gets commoditized. The platform that controls the data context, the schema, the security boundary, the historian, and now the MCP interface is not being abstracted away. It is what the application layer depends on. Ignition sits exactly at that layer. We are the system that understands what a Modbus register 40001 actually means, what its historical pattern looks like, what alarm conditions are associated with it, and what the UDT definition implies about the machine it belongs to. An AI agent without that context is useless on the factory floor regardless of how good the underlying model is. That context lives in Ignition.

Looking ahead, we tend to think in one-year increments rather than five because the pace of change in this industry demands it, but there are three clear vectors we are building along. The first is native AI integration in Ignition, both from a developer experience standpoint and from an agent standpoint. People will be building Ignition applications in ways that look a lot like vibe coding today, using AI to accelerate development, generate configurations, and scaffold projects. But they will be doing it on the Ignition platform precisely because of everything that makes Ignition trustworthy: the security model, the support infrastructure, the unlimited licensing, the community. The MCP Module is the foundation of the agent-facing layer, and we will steadily build on it. The second vector is the annual release cadence. Moving to Ignition 2027 and annual major releases makes us meaningfully more agile, and Ignition 8.3 lays the architectural foundation for a lot of what comes next. We will be showing a significant amount of what that looks like at ICC 2026, with announcements about Ignition 2027 features that I think will make clear how seriously we are taking this moment.

The third is the ecosystem. The Technology Ecosystem Program, i3X interoperability, and the integrator network are all working toward the same outcome: an industrial platform that compounds in value because every new partner, standard, and agent capability makes the entire installed base more capable. The long-term bet is that the companies that win in the agentic industrial stack will be the ones that own the trusted data layer on the factory floor, have the community and the integrators to deploy it, and move fast enough on the software side to be the platform AI applications are built on rather than the system they have to work around. That is the position we have been building toward since 2003, and Ignition 8.3 and the MCP Module are the most direct expressions of it to date.

To contact Travis, reach out to him on LinkedIn here. He’s always open to chatting and sharing valuable insights.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Jiga, Upkeep, and Industry 4.0 Club.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Jiga, Upkeep, Ironloop, and Industry 4.0 Club.

IMTS Investor and Innovation Summit 🗞️🔬 📚

I’m excited to share something I’ve been working on for the past few months with the Association for Manufacturing Technology (AMT), tied to IMTS in Chicago on September 15th. Reshoring, the rebuilding of the defense industrial base, and the first wave of AI-native tooling have turned manufacturing into one of the hottest sectors in technology.

The Investor & Innovation Summit at IMTS is a one-day event for the investors, operators, and corporates building and funding industrial technology. On Tuesday, September 15, we’ll spend the day digging into what’s actually driving the sector and Breaking the Bottleneck is helping convene the room.

With U.S. manufacturing technology just posting a 22.5% jump in machinery orders in a single month, reshoring incentives working their way through the industry, and a defense industrial base taking shape, there’s never been a better moment for an event like this at IMTS. We’re bringing together the best in the field to speak to what’s actually working in go-to-market, what broke in production on the shop floor, and what reindustrialization really requires in capacity, capital, and contracts. The Summit is an effort to get corporates, investors, and startups into one room and those conversations out in the open.

We’ll kick off with AMT’s Ryan Kelly and Chris Chidzik opening with a read on where U.S. manufacturing stands in 2026, followed by five panels with great moderators Leland Chamlin, Julian Counihan, Annelies Gamble, and Em Moshouris.

• Hyperscalers to the Shop Floor - Praveen Rao, Shiv Trisal, and Les Karpas

• The New Defense Industrial Base - Aaron Slodov, David Caputo, and two other speakers to be finalized

• Building Software for the Factory Floor - Devin Bhushan, Sunny Han, Igor Marinelli, and Nick Haase

• Reinventing CAD - Brad Rothenberg, Sai Nelaturi, Hardik Kabaria, and Taylor Young

• Robots Get Real - Saman Farid, Edward Mehr, Andy Lonsberry, Ariyan Kabir

I’ve haven’t been this energized about something adjacent to the newsletter in a while, and I’m excited to share it with you! Registration instructions are listed below!

Details

McCormick Place, Chicago, Tuesday, September 15, 2026 ·

• Sessions 9:00 am – 4:30 pm ·

• Reception 5:00 pm– 6:30 pm

[Register Now] Select Investor Summit as Ticket Option to Sign Up

Breaking the Bottleneck is brought to you with support from our partners.

• AMT & IMTS - The Association for Manufacturing Technology represents and promotes U.S.-based manufacturing technology and hosts the largest manufacturing technology trade show in North America, IMTS.

• Jiga - A platform that connects hardware teams with vetted manufacturers for custom parts. Instead of black-box instant quotes, you see exactly who’s making your parts and can talk to them directly.

• Upkeep - CMMS Software helping maintenance teams manage work orders, assets, and preventive tasks efficiently from any device.

• Industry 4.0 Club - Leading cross-functional group accelerating the adoption of Industry 4.0, delivering better experiences to consumers, better profits to manufacturers, and better jobs to factory workers.

• Ironloop - Ironloop is a modern OT/ICS software platform that delivers configuration-level intelligence, providing manufacturers with ground-truth security visibility and confidence in compliance across air-gapped networks.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Jiga, Upkeep, and Industry 4.0 Club.

Extreme Analysis Market Map with Threadmoat🗞️🔬 📚

Enjoyed building this with Michael Finocchiaro at Threadmoat. This is the second edition of a new series featuring market maps of engineering and manufacturing startups shaping the landscape today. For more analysis of this ecosystem and more on each of these startups, check out ThreadMoat and visit Fino’s profile! As always, let us know if we’re missing anyone!

Some Interesting Reads:

• Accenture has agreed to buy a majority stake in Dragos and the entire stakes in runZero and NetRise, generating some $208M in ARR for roughly $4.175bn. How this reshapes Dragos, given its heavy reliance on managed security providers and Accenture’s rivals such as Deloitte, will be worth watching in a category that, to this observer, looks ripe for innovation. [Dragos]

• Slate drove down BOM costs by focusing on traditional welding rather than stamping, and on a rear suspension design called the De Dion, dating back to the late 1800s. [Heatmap]

• Cyber specialists traced the Jaguar Land Rover hack to a Russian group, with Microsoft identifying the perpetrators and alerting the carmaker. The intrusion was a $2.5B blow to the British economy and cost the company roughly $350M. [NY Times]

• OpenAI and Broadcom get spicy with Jalapeno, which was co-developed from initial design to manufacturing in nine months, an extremely fast ASIC development cycle [OpenAI]

• Ford rehired engineers after quality control AI fell short [Bloomberg]

• Last week, Agility announced plans to go public through a SPAC, citing roughly $300m in bookings for its Digit v5 robot. It is tempting to read this as the first crack to appear in the humanoid-robotics market. [Agility]

• Another Physical AI market map [Bessemer]

  

• Another EV company conducts layoffs, a direct effect of dwindling subsidies, an opening for China's dominant EV ecosystem, and an opportunity for Mexico to develop its own EV [CNBC, Gizmodo]

• Industrial M&A deal values reached a record $173 billion over the past year, with average deal sizes operating at $155 million in FY24, $288 million in FY25, and $375 million in the latest annual period. [PwC]

• The much-touted helium shortage left chip-makers' margins largely unscathed because high margins let them outbid rival buyers. [WSJ]

• McKinsey Partners with a perspective on Robotics [McKinsey]

Products & Research:

An awesome breakdown of what is working and where we are headed in AI for CFD and FEA by Mark !

Burhop

AI Surrogates in CFD and FEA

Introduction: Why CFD/FEA Engineers Care About Surrogates Now…

Read more

2 months ago · 2 likes · Burhop

• Cobot launched the second iteration of Proxie, redesigning the initial version after distilling more than 500 design insights from nearly 13,000 operating hours accumulated across 28 Proxie robots in logistics, manufacturing, and healthcare. [Robot Report]

• A breakdown on Geometric Deep Learning in CAE [Siemens]

• A deep dive with Jacomo Corbo into PhysicsX, which can now accelerate simulation for industrial engineering by between 10,000 and 100,000 times. [Intrepid Growth Partners]

• Fuuz’s VS Code Extension [Fuuz]

• Digital Metal launches a quoting engine [Digital Metal]

• An interesting conversation with Jim Belosic and SendCutSend [Core Memory]

Financing & Transactions 💵

HyperLight raised $80M from MediaTek
Semiconductors • Series C • USA

Isometric raised $40M from AVP
Industrial AI (Certification) • Series A • UK

Partly raised $50M from DST Global Partners
Industrial AI (Supply Chain) • Series B • USA

Copia Automation raised $26M from AE Ventures and Squadra Ventures
Industrial Automation and OT • Funding • USA

Podium Automation raised $18M from Construct Capital
Industrial Automation • Series A • USA

Planned Downtime 🧑‍🔧

Chip Design with MatX

Breaking the Bottleneck is brought to you with support from our partners.

• AMT & IMTS - The Association for Manufacturing Technology represents and promotes U.S.-based manufacturing technology and hosts the largest manufacturing technology trade show in North America, IMTS.

• Jiga - Is a platform that connects hardware teams with vetted manufacturers for custom parts. Instead of black-box instant quotes, you see exactly who’s making your parts and can talk to them directly.

• Upkeep - CMMS Software helping maintenance teams manage work orders, assets, and preventive tasks efficiently from any device.

• Industry 4.0 Club - Leading cross-functional group accelerating the adoption of Industry 4.0, delivering better experiences to consumers, better profits to manufacturers, and better jobs to factory workers.

• Ironloop - Ironloop is a modern OT/ICS software platform that delivers configuration-level intelligence, providing manufacturers with ground-truth security visibility and confidence in compliance across air-gapped networks.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Jiga, Upkeep, and Industry 4.0 Club.

If You've Industrialized Bad Data, You Haven't Solved the Problem, You've Scaled It🎙️

You spent three decades on the operator side, leading global process control at Covestro, then overseeing 100+ discrete sites at Stanley Black & Decker, before joining APERIO. What’s a problem that was structurally invisible to you as a Fortune 500 SVP but became obvious from the startup seat? And conversely, what does the startup side consistently get wrong about how operators actually buy?

The thing that was structurally invisible to me as an SVP at Covestro or VP at Stanley Black & Decker wasn’t the technology; it was the cost of organizational inertia on every decision. When you’re inside a Fortune 500, you assume friction is normal. From the startup seat at Aperio, I suddenly saw it clearly: large operators aren’t slow because they’re unsophisticated. They’re slow because they’re accountable in ways startups never are. A bad call on a 16-site DCS rollout isn’t a pivot, it’s a safety event, a capital write-off, and a career moment all at once. One factor in my long-term success was claiming accountability for technical decisions. That relieved the workforce from the burden and let us move faster.

The flip side, startups consistently treat the technical champion as the buyer. They win the plant engineer’s enthusiasm, only to wonder why the deal stalls. The COO or SVP who controls the capital budget thinks in terms of risk reduction, not feature lists. Startups pitch OEE improvements and ROI calculators; operators are asking, “What happens to my site if this fails at 2 am and your three-person support team is asleep?” Getting that answer right, the risk story, the integration path, and the operational continuity commitment are what separate a pilot from a rollout. That’s where pilot purgatory lives.

OT owns the sensors, IT owns the cloud, and data science owns the models. Accountability is distributed, but someone still has to sign the contract. Who’s APERIO’s typical internal champion today, a plant-level OT lead, a corporate digital transformation VP, a CIO? And once they’re bought in, how do you align the other two seats at the table when each one measures success on a different KPI?

At Aperio, our most consistent internal champion was the end user, typically the OT data historian lead or unit engineer. Not the digital transformation VP, not the CIO. The person closest to the process, who understood the data and felt the pain of not having better tools. This tells you something important about where industrial AI adoption actually lives today: it’s being pulled from the shop floor up, not pushed from the boardroom down. My lifelong experience is that the person whose hands are closest to the process knows more about it than anyone else.

The alignment challenge across IT, OT, and data science remains unsolved industry-wide. OT cares about uptime. IT cares about security and governance. Data science cares about model performance. Those aren’t naturally compatible, and the contract signer is trying to satisfy all three simultaneously with a vendor who only speaks fluently to one of them. Layer on top of that a C-suite chasing the latest technology hype cycle, and you have budget and mandate flowing toward whatever got the keynote at Davos that year, not necessarily what the plant actually needs. When alignment works, it’s because someone with enough organizational authority makes it their problem to broker. When that person doesn’t exist, you get pilot purgatory from the inside. Extremely frustrating for everyone involved.

On a Tuesday afternoon at a plant running APERIO, what is the system actually doing differently from a traditional historian, and at which point in that chain does the error get caught: at the sensor, at the edge, inside the historian, or in the lake? What’s required from a change management perspective to ensure the error is captured, noted, and resolved before it reaches the model?

On a Tuesday afternoon at a plant running DataWise, the system is doing something a traditional historian wasn’t designed to do: actively evaluating OT data quality across the entire enterprise in near real-time. A historian records. DataWise watches.

A board operator has deep experience but finite attention. They’re tuned into a handful of critical tags. DataWise is evaluating anomalies enterprise-wide, from sensors to the cloud, including sensor failures, IT/OT network issues, and process deviations. The error gets flagged at the point it occurs, before it ever reaches the model.

That matters more than people realize. Models are only as good as the data they’re built on. DataWise flags bad data from the earliest stages of deployment, keeping the historian cleaner and ensuring bad data that does get through is identified and resolved rather than quietly corrupting model output over time.

The change management requirement is significant. The technology can flag the anomaly, but someone has to own the resolution. That means clear accountability for data quality, not just process performance, and building the habit of treating a DataWise flag as seriously as a process alarm. In most plants, that’s a cultural shift, not a technical one. And the honest reality is that many industrials simply don’t have the time or manpower to address every anomaly. The system can surface the problem. Only the organization can fix it.

The 80% AI failure rate is often cited, and “data quality” has become the convenient post-mortem. But pilots also fail because of the wrong scope, buyer, KPI, or business case. How much of what gets blamed on data is actually a buying and scoping problem dressed up as a data problem, and how do you know APERIO is solving the real bottleneck rather than the most visible one?

The 80% failure rate citation is real, but “data quality” has become a catch-all that lets everyone off the hook. Pilots fail for all the reasons you listed: wrong scope, wrong buyer, wrong KPI, and data quality is often the most visible symptom rather than the root cause.

That said, for industrials specifically, data quality is genuinely foundational. You cannot train a reliable predictive maintenance or process model on bad data. Industrials are sitting on millions of sensors, years of inconsistent tagging, and an IT infrastructure never designed with model training in mind. That enterprise data mess is real.

But most pilot failures I observed weren’t primarily data problems. They were buying and scoping problems wearing data’s clothes. The champion didn’t own maintenance or operations and couldn’t influence the work that needed to change when the model surfaced an issue. You’d have a model correctly flagging an asset anomaly, with no one authorized to act on it. Or what looked like a model failure was actually a bad sensor throwing the output sideways, and the organization had no process for telling the difference.

Scope is the other silent killer. If the pilot doesn’t connect to what leadership actually measures, it dies quietly regardless of technical performance. The data problem is real. The buying-and-scoping problem is just harder to admit.

Industrial foundation models trained on operational time-series data are starting to appear from the Siemens, GE, and Rockwell sides of the market. Does that future make a dedicated data trust layer less important, because models can learn around noise, or more importantly, because biased data at scale produces confidently wrong outputs across an entire fleet?

The “models can learn around noise” argument is mostly hype, and it’s a dangerous framing for industrials specifically. A foundation model trained on operational time-series data is still being trained on something. If that something includes years of bad sensor data, inconsistent tagging, and unresolved instrumentation failures, the model doesn’t learn around it. It learns from it. At fleet scale, biased data doesn’t produce uncertain outputs. It produces confidently wrong ones consistently across every asset the model touches.

But the more important point is entirely missed in the model conversation. In an industrial environment, a bad sensor can trigger a control action or a safety event. The stakes aren’t a wrong recommendation on a dashboard. They’re a process upset, an unplanned shutdown, or worse. Field instrumentation reliability and data trust aren’t model training problems; they’re operational integrity problems.

The Siemens, GE, and Rockwell foundation models are worth watching. But they don’t eliminate the need for a data trust layer; they make it more critical. You still need a human in the loop who can distinguish a genuine process anomaly from a failed transmitter. Foundation models don’t know the difference. They just scale whatever they’re given. If what you’re giving them is unreliable, you’ve industrialized the problem, not solved it.

To contact Jane, reach out to her on LinkedIn here. She’s always open to chatting and sharing valuable insights.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Jiga, Upkeep, and Industry 4.0 Club.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Jiga, Upkeep, and Industry 4.0 Club.

Never Let the Technology Precede the Problem🎙️

You’ve held leadership roles at IBM, AWS, and now Google Cloud, each during a different phase of industrial digitization. What’s the single biggest lesson each stop taught you that the previous one couldn’t?

It’s fascinating to reflect on how rapidly industrial digitization has evolved throughout my career. My experiences at these three companies each taught me a unique lesson about transformation, with one building directly on the next.

My time at IBM was a masterclass in the sheer scale and complexity of enterprise transformation. When I joined to lead technology strategy, the market was flooded with point solutions, early cloud, mobile, and social vendors, each pitching their specific technology as a hammer looking for a nail. At IBM, we focused instead on the concept of convergence benefits. This was about looking at the holistic problem and working backward from the desired business outcome. The big lesson there was that enterprise transformation requires understanding how people, processes, and technology must intricately weave together to achieve that final result.

If IBM taught me how enterprises think, AWS taught me how they can execute with agility. AWS radically challenged the traditional approach of bringing in big teams to tackle a project all at once. Instead, they pioneered the “two-pizza team” philosophy—the idea that an effective project team should be small enough to be fed by just two pizzas. By keeping teams tight, you minimize lines of communication and eliminate the bureaucratic overhead that slows down decision-making. This emphasis on speed, nimbleness, and constant experimentation for the customer is exactly what fueled AWS’s explosive growth. By focusing on highly targeted, small-scale pilots that could be scaled rapidly, we proved that you don’t have to sacrifice velocity to solve enterprise-grade problems.

Now at Google Cloud, the lesson has been all about intelligence and the true power of data. As a data and AI company, our core mission is to organize the world’s data and make it universally useful, a mission that translates incredibly well to the manufacturing industry, which sits on a goldmine of both IT and OT data. In that data resides the narrative of everything happening within a company, why it’s happening, and what you can do about it. Google has made me realize that the ultimate phase of digitization isn’t just about understanding the enterprise blueprint or moving fast. It’s about listening to what the data is telling you, uncovering actionable insights, and making the entire enterprise fundamentally smarter.

One thread that runs through all of those roles is the IT/OT divide, something you’ve called the “enemy of profits.” After years of the industry talking about convergence, why is that wall still standing, and what does Google Cloud’s Manufacturing Data Engine do differently to tear it down?

The IT/OT wall still persists due to the inherent complexity of manufacturing. Unlike retail, for example, which only has to manage buying and selling, manufacturing operates on three distinct imperatives: first, you have to invent a great product, then you have to make it faster, cheaper, and better than the competition, and finally, you have to navigate a complex supply chain and logistics to sell it to the right people. This involves a massive, fragmented web of people, processes, and machines. Historically, the industry managed this complexity by using traditional standards like ISA-95, which explicitly separated shop-floor OT from corporate IT systems. Because these systems speak entirely different languages, the divide became deeply structural.

While that divide is starting to blur, Industry 5.0 and AI agents are completely changing the game. To an AI agent, data is just data; it doesn’t care whether information originates on the shop floor or in the cloud. The real challenge is the heavy lifting required to connect those machines, extract their data, and harmonize it. True harmonization means building a digital thread that can instantly tell you which customer ordered a product, which plant made it, during which shift, and on which specific machine.

This is exactly where Google Cloud’s Manufacturing Data Engine and Manufacturing Connect come in. We can seamlessly connect to more than 250 OT data sources, such as PLCs, HMIs, and SCADA systems, and automatically integrate that machine data with your IT systems of record, such as CRM or PLM platforms. By bridging these two worlds into a unified data foundation, we break down historical walls and give companies the single thread they need to build a truly intelligent enterprise.

A key piece of that convergence story is the Cortex Framework, designed to bridge heavy ERP data and SAP environments with real-time analytics. For a mid-market manufacturer running on legacy systems, what’s the honest timeline and investment to get from where they are today to a functioning “digital thread”? How should leaders benchmark whether the framework is driving measurable ROI?

To understand the honest timeline and investment required, you first have to look at how long this process traditionally takes. In manufacturing, you are constantly dealing with a “museum of machines”–hundreds of protocols and assets that are 30 to 50 years old and were never designed to talk to each other. Traditionally, a company would spend six to eight months just on data extraction, stretching the total project timeline out to 14 months. It’s a slow, frustrating process because if you ask five different legacy systems for today’s production numbers, you will get five different answers. They all have their own metadata and versions of truth.

Google Cloud Cortex Framework and our Industry 5.0 architecture fundamentally change this dynamic by acting as a translation layer. Cortex provides pre-built data models for complex environments like SAP, allowing you to connect legacy systems and extract data in days or weeks, rather than months. We sit on top of everything using a Unified Name space (UNS) with a pub/sub model, meaning all your legacy systems subscribe to a single source of truth rather than creating their own. By getting this data layer right, we compress the 14-month timeline into weeks. In fact, we’ve seen customers deploy more than 800 AI agents into production in under 10 months because their data foundation was finally seamless.

When it comes to benchmarking ROI, leaders shouldn’t view this as a multi-year gamble. Our global survey of over 500 manufacturing executives shows that generative AI is already delivering sizable returns, with 75% of leaders reporting immediate improvements in productivity. The most successful leaders are achieving these gains by following a “buy to integrate” philosophy. They don’t try to build everything from scratch, nor do they buy rigid, closed systems. They let Google Cloud handle the heavy data lifting, freeing them up to focus their time and budget on the specific use cases that drive real productivity and growth.

Even with the right data infrastructure in place, none of it matters without people to act on it. With 30% of manufacturing jobs going unfilled and 1.9 million projected vacancies by 2033, that’s a real problem. You’ve framed AI as a tool to capture retiring workers’ expertise and deliver it to new hires. How do you actually extract decades of tacit knowledge from a seasoned machinist and encode it in a way an AI agent can deliver reliably?

The manufacturing talent gap is undeniable, but the nature of the workforce is also undergoing a massive shift. The traditional image of a factory worker carrying wrenches is giving way to a new reality of “technocrats” who navigate the shop floor with tablets, configuring machines remotely and using computer vision to diagnose issues. Because of this shift, we can’t bridge the talent gap using old-school methods. If you ask a veteran worker to write down everything they’ve learned over 40 years, the project will fail miserably. As humans, we have decades of experience, which makes us excellent at catching mistakes but not necessarily at documenting them.

AI can play a profound role in capturing that hidden expertise by letting veteran workers validate information rather than author it from scratch. Instead of writing a massive manual, an expert can simply be recorded performing a complex repair. AI can instantly translate that footage into a step-by-step instruction guide or a synthetic training video. The veteran worker simply reviews and approves it, acting as the expert validator behind the AI.

With that verified data foundation in place, companies can deploy AR and VR tools to guide new hires through complex tasks in real time. The AI walks the technician through the process step by step, using automated validation to confirm each task is completed correctly before allowing them to proceed. Ultimately, this isn’t just about preserving legacy knowledge; it’s about dramatically compressing onboarding timelines and empowering companies to help generalist workers confidently perform highly specialized tasks.

You’ve spent two decades watching industrial tech waves: IoT, blockchain for supply chain, digital twins, and now agentic AI. Which of those earlier bets do you think the industry over-invested in, and what lesson from those experiences should manufacturers bring to agents?

Rather than a case of simple financial over-investment, it’s more useful to look at where the industry chased hype that ultimately under-delivered. From that list, blockchain and digital twins are the two clear examples where I believe the technology was applied to the wrong problems.

With blockchain, the technology was heavily hyped because people saw its success in crypto and started treating it like a hammer in search of a nail. They tried to force it into standard transactional supply chain problems without realizing it was built to solve specific multi-party trust issues, not to handle real-time, compute-heavy enterprise tracking. Similarly, early digital twins lacked a clear, unified definition. Many companies poured resources into building complex 3D CAD models of their shop floors just to have a visual representation on a screen, without tying those models to a specific operational problem or a measurable ROI.

The lesson for the era of agentic AI is clear: never let the technology precede the problem. Those earlier tech waves stalled because companies invested in the tool first and the use case second. With AI agents, the focus has to remain entirely on the business outcome. Leaders shouldn’t deploy an agent just because the capability exists; they need to identify a specific bottleneck and build the data foundation to solve that precise issue. When you work backward from the problem rather than chasing the tool’s hype, you ensure your technology delivers immediate value while protecting your transformation from reputational risk.

To contact Praveen, reach out to him on LinkedIn here. He’s always open to chatting and sharing valuable insights.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Jiga, Upkeep, and Industry 4.0 Club.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Jiga, Upkeep, and Industry 4.0 Club.

On Autodesk + MaintainX 🗞️🔬 📚

This past week, Autodesk announced an agreement to acquire MaintainX, a developer of maintenance and operations software, for approximately $3.6 billion. It is an impressive outcome for Chris, Nick, and the MaintainX team, a credit to what they have built over the past few years, and to the manufacturing software ecosystem more broadly.

The deal is funded with roughly $1.6 billion on hand and $2 billion of new debt. It is the largest acquisition in Autodesk’s history, with the previous high being the $875 million paid for PlanGrid in 2018. This is a major strategic land grab, and it signals Autodesk’s conviction in its expansion from design and make into operations.

Two things stand out. The first is around the deal structure. In a category where five-year revenue multiples for engineering and manufacturing technology sit around 11.9x and where even the five-year peak for broader SaaS tops out near 26.0x, an implied ~27x forward revenue multiple feels excessive. Furthermore, $2 billion of new debt is not free money. At today’s rates, the interest bill alone sets a bar that the cross-sell into Autodesk’s installed base has to clear before any of this counts as a good use of capital.

This raises questions about who was driving this liquidity event, especially given that the company raised a $150M round at $2.5B barely a year ago and is now flipping at $3.6B. It echoes the peak SaaS exuberance of 2020–21, an era of near-zero rates that, in my view, bears little resemblance to today’s environment. In fairness, a category leader growing by more than 50% is a strong asset, especially if there is a clear strategic fit. That, however, brings me to my second point.

On strategic fit, I am not sure I follow, at least as it pertains to the manufacturing sector. The logic of acquiring real-time telemetry (asset failures, work orders, maintenance patterns) is clear enough, particularly given that operations and maintenance command 70–80% of an asset’s total lifetime cost. What is less clear is how that operational data feeds back into Autodesk’s design and engineering functions or into the generative design AI that Autodesk sells. To close the design-to-manufacturing loop, an MES, QMS, or OEE platform would, in my view, have been the more natural fit. It is worth noting that on the construction and facilities side, the strategic fit is stronger, and I see greater alignment.

For all that, this is good news for the manufacturing software ecosystem, especially after years in which venture and private capital have poured into the sector. Companies like Tractian, Squint, and Augury should only benefit, as all three just watched their category get repriced. As Autodesk further expands into the operations category, it will be interesting to see who moves next.

Some Interesting Reads:

This came out a few weeks back, but McKinsey calculates that building the capacity to produce products and their upstream inputs in the US could cost on the order of $2 trillion. The analysis is fantastic, providing insight into the ramp-up needs, a capacity and utilization sensitivity analysis on import volumes, and job and labor requirements. [McKinsey]

• An awesome breakdown of the functional taxonomy of World Models from Fei Fei Lei [X]

• Honeywell’s Aerospace spinoff plans to generate at least $6.5 billion in earnings and $4 billion in cash flow by 2030 [Honeywell]

• A breakdown of the impact the manufacturing and robotics community has had on Columbus, Ohio [NYTimes]

• BCG with its vision for the Factory of the Future [BCG]

• A great analysis from the FT, echoed by Axios and Fortune this week. Manufacturing investment is sputtering, as corporate commitments to build factories fail to translate into construction. Furthermore, an interesting quote from Katie Farmer, CEO of BNSF, states that she has “certainly seen a revival” in “certain commodity areas”, including steel, but that “the flip side of that is that in the other industrial segments, we’ve really seen a plateau”. [FT]

• GE Aerospace invests another $1B in the US [GE]

• A high-level analysis on the state of process manufacturing [IoT Analytics]

• An interesting assessment of the structural advantages of the modern robotics systems integrator [Avi Zurlo]

• Energize with another Industrial AI Market Map [Energize Capital]

Sponsored by Jiga

Build Real Supplier Relationships With Jiga - Most hardware teams are stuck choosing between two bad options: instant-quote platforms that hide suppliers behind a black box, or local shops that hit capacity walls right when you need them most. One trades away relationships for speed. The other offers trust but can’t scale. Jiga gives you both. Direct access to vetted manufacturers with reliable capacity from prototype through production. You speak directly to the people making your parts. And when you come back, they already know your specs and technical needs. Same shop, same knowledge, consistently reliable quality without starting over every time.

Products & Research:

Protos is an AI-native co-engineering workspace built by Aris Machina (Sweden) for teams designing physical systems from first principles, batteries, pharma, humanoids, advanced materials, optics, and much more. It collapses the coordination tax across data, simulation, and domain knowledge, becoming the R&D operating system for deep tech. Check out more here at the Protos Homepage [Aris Machina]

• Siemens announced the Intelligence Center X, combining Mendix with Siemens’ Graph and AI Studio to help customers build applications and deploy agents with operational context. [Siemens]

• A pretty awesome breakdown of WestMag, one of the coolest newish startups building the American Actuation Supply Chain by both and . [WestMag]

  Core Memory

  This Is Why America Can’t Have Robots And Other Nice Things

  At a hardware happy hour in the Mission — I dragged our social media editor Armaan along, half-expecting to be the only woman there — I met a different kind of tech bro. You know the usual kind: youn…

  Read more

  3 months ago · 119 likes · 6 comments · Kylie Robison
  Not Boring by Packy McCormick

  America Spins on Westmag

  Welcome to the 730 newly Not Boring people who have joined us since our last essay! Join 268,518 smart, curious folks by subscribing here…

  Read more

  3 months ago · 129 likes · 2 comments · Packy McCormick

• NVIDIA announced Cosmos3, an open Omni-model for Physical AI Reasoning and Action [Hugging Face]

• Generalist announces $400 million in new funding [Generalist]

• CoLab announces Operator to analyze, search, generate, and automate tasks across your engineering data [CoLab]

• Intrinsic launches its own Vision Model and 3D-Object Perception Transformer [Intrinsic]

• AWS launches Kiro, an AI-powered IDE that accelerates iteration across provisioning, policy training, and simulation validation with full contextual awareness of the Physical AI lifecycle. [AWS]

Financing & Transactions 💵

Agile Robots plans to raise $800M from Softbank and others
Robotics / Physical AI • Funding • Europe

Generalist raised $400M from Radical Ventures, 8VC, USV, and Norwest
Robotics / Physical AI • Funding • USA

Mecka raised $60M from Framework Ventures
Robotics / Physical AI • Funding • USA

Endra raised $50M from a16z and Notion Capital
Industrial AI and MEP • Funding • Europe

Layup Parts raised $42M from MarlinSpike
Composites • Series A • USA

Invisix raised €20M from Hitachi Ventures
Semiconductors • Seed • Europe

Planned Downtime 🧑‍🔧

A Great Lecture with Base Co-Founder Justin Lopas

The 20 Greatest Soccer Players of All Time

Breaking the Bottleneck is brought to you with support from our partners.

• AMT & IMTS - The Association for Manufacturing Technology represents and promotes U.S.-based manufacturing technology and hosts the largest manufacturing technology trade show in North America, IMTS.

• Jiga - Is a platform that connects hardware teams with vetted manufacturers for custom parts. Instead of black-box instant quotes, you see exactly who’s making your parts and can talk to them directly.

• Upkeep - CMMS Software helping maintenance teams manage work orders, assets, and preventive tasks efficiently from any device.

• Industry 4.0 Club - Leading cross-functional group accelerating the adoption of Industry 4.0, delivering better experiences to consumers, better profits to manufacturers, and better jobs to factory workers.

• Ironloop - Ironloop is a modern OT/ICS software platform that delivers configuration-level intelligence, giving manufacturers ground-truth security visibility and compliance confidence across air-gapped networks.

• T51 - T51 is an AI-native planning, scheduling, and optimization platform that integrates securely with your existing ERP, MES, and data stack.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Jiga, Upkeep, and Industry 4.0 Club.

AI Is Accelerating Design Faster Than We Can Validate It

You spent years at PARC and Carbon doing deep research on accessibility constraints for multi-axis machining and lattice design before founding C-Infinity. What was the specific moment or failure mode you kept seeing in the field that convinced you the problem had to be solved from scratch, rather than fixed inside an existing platform?

For fifteen years, I watched the same failure mode repeat from the outside. At PARC and Carbon, I was often sitting at the boundary between advanced process capability and the engineering organizations trying to use it. I watched Fortune 500 manufacturers, aerospace primes, medical device companies, and consumer hardware teams hit the same wall: a design could be complete in CAD and still be nowhere close to production-ready.

The specific domains changed from multi-axis machining, lattice design, metal additive manufacturing, and assembly, but the underlying bottleneck was the same. Function-oriented product data had to be transformed into manufacturing-ready process intelligence, and that transformation was still done almost entirely by expert humans. CAD models are authored by design engineers with the implicit assumption that the next consumer is another human: a manufacturing engineer, process planner, quality engineer, or supplier. That works when timelines are forgiving. It does not scale when companies are under pressure to accelerate time to market and respond quickly to change.

That is the missing computing layer we started C-Infinity to build. Existing platforms were not designed for it. CAD defines the product, PLM manages the product record, MES executes the production workflow, but the reasoning layer that turns product definition into a feasible manufacturing plan has largely been missing. If AI is going to meaningfully accelerate time to market, it cannot assume engineering data was created for automation. It has to work with human-oriented engineering data while reasoning over geometry, physics, motion, constraints, and manufacturing process logic.

We started with assembly because nearly every advanced product is assembled, and assembly planning exposes the core challenge clearly: translating product structure into process structure. It is a combinatorial, geometry-rich, physics-constrained problem embedded in a complex enterprise workflow. Solving it requires a purpose-built foundation.

AutoAssembler treats the EBOM and MBOM as dual representations of the same configuration space rather than two separate handoffs. In practice, where does that model break down first is it change management across variants, geometry validation, or organizational resistance to collapsing two workflows into one?

The honest answer is variants under continuous change, but it is worth being precise about what “dual representation” actually covers.

The EBOM and MBOM share a part structure, and at that layer they can be treated as projections of a common configuration space. AutoAssembler uses that fact to reason from the EBOM to the assembly sequence. But a full MBOM is more than its part structure. It also carries process content: tools, fixtures, consumables, work centers, sequence, and process logic that have no direct analog in the EBOM.

So the projection holds for parts. It does not hold for the MBOM as a whole.

The model starts to strain first under variant proliferation. A single product family can generate thousands of MBOM permutations driven by configurability (regions, customer options, supplier substitutions, etc). Each variant inherits much of the assembly logic of its siblings but diverges in specific places. The hard part is tracking which planning decisions are shared across the family and which are variant-specific.

The second pressure is continuous change. Every ECO forces a recomputation: what changed in the EBOM, which parts of the existing MBOM and assembly plan remain valid, what has to be re-planned, and what can be reused. When a change touches a part that participates in multiple variants, the two pressures compound. You are re-validating the projection across a family, not a single product.

Geometry validation is a real failure mode, but it is usually more well-defined, e.g. this fastener cannot be reached, this sub-assembly cannot be sequenced as drawn. Those problems become much harder when variants and ECOs are also in motion, because the system has to know which configuration and which version of the plan it is validating.

Organizational resistance is real, but in our experience it tends to recede when the model holds up under variants and change. When it does not, no amount of organizational alignment fixes the underlying problem.

So the model breaks down first where the projection is most stressed: high-mix products under continuous engineering change. That is why EBOM-to-MBOM translation stops being a data-management problem and becomes a manufacturing-reasoning problem.

When you’re selling into an OEM with an entrenched PLM stack, who do you actually close first: the manufacturing engineer, the manager, or the VP of Engineering running the program? What is the initial product wedge?

The wedge is the ECO (Engineering Change Order) process.

That is where the pain is immediate, measurable, and already tied to the customer’s existing PLM workflow. OEMs process thousands of engineering changes, and every change creates downstream manufacturing questions: what changed, which process steps are still valid, what has to be re-planned, what work instructions need to be updated, and what risks have been introduced?

The manufacturing engineer is usually the primary user because they feel that pain directly. They are the ones working through the backlog: interpreting design changes, assessing assembly impact, reusing what they can from the existing plan, and manually updating instructions.

AutoAssembler helps them do that work faster and more systematically. We identify what changed, determine which planning logic can be reused, generate the missing portions of the assembly plan, and produce updated work instructions.

That makes ECO a strong initial wedge because it is not an abstract AI use case. It is a concrete operational bottleneck with direct impact on manufacturing readiness, engineering throughput, and program execution.

Over time, the goal is to shift this intelligence left. The same manufacturing reasoning that helps process planners evaluate ECOs should also become available to design engineers before changes are pushed downstream. That allows them to measure the manufacturing impact of engineering changes earlier, when the cost of iteration is lower, and the design is still flexible.

The economic buyer is usually higher up: a VP of Engineering, VP of Manufacturing, or digital transformation leader. They care about throughput, change velocity, program delays, and the cost of engineering rework.

How much of AutoAssembler’s architecture today is model-based versus statistical, and where do you expect that balance to shift as the product matures? To ensure seamless adoption, what’s your philosophy on integration depth versus breadth, and which platform are you targeting next?

AutoAssembler uses a neuro-symbolic architecture: model-based and deterministic at the core, with machine learning applied where it creates leverage.

Manufacturing planning is not a pure prediction problem. A system cannot simply infer a process plan statistically and treat it as correct. It has to reason explicitly over geometry, motion, contacts, constraints, accessibility, sequencing, and feasibility. That requires a deterministic planning foundation.

Machine learning becomes powerful around that foundation. It can learn customer-specific preferences, recurring planning patterns, reuse decisions, and organization-specific process logic. Over time, the system becomes more informed by each customer’s environment while preserving deterministic reasoning where correctness matters.

That architecture also shapes our deployment model. We deploy inside the customer’s VPC and do not cross-train across customers. Each customer’s system is tuned to its own products, processes, and know-how — and as the system learns over time, that institutional knowledge compounds inside their environment rather than leaking into a shared model. The customer’s manufacturing know-how stays with them, and the system gets more valuable to them with every deployment, ECO, and planning decision.

On integrations, we are API-first. We support all major CAD platforms today, and we design our endpoints to connect with critical workflow triggers: ECOs, design releases, variant updates, and manufacturing-readiness checks.

That focuses our integration strategy. We prioritize workflow depth over connector count. We integrate where decisions happen and where automation can remove friction from the workflow. The goal is for AutoAssembler to operate as an event-driven manufacturing reasoning layer across the customer’s enterprise stack.

In practice, many customers operate mixed CAD and PLM environments, which makes interoperability essential. Our tightest integrations today are within the PTC ecosystem. We’re deliberate about not pre-announcing the next platform; we let customers pull, and commitments drive that sequencing rather than picking targets in advance. The goal is to meet customers where they already work while giving them a reasoning layer that connects product definition, manufacturing planning, and production readiness.

Most people in the CAD/PLM world frame the opportunity as “AI will accelerate design.” Your thesis seems to be the opposite, AI is already accelerating design faster than humans can validate it, and the real bottleneck is now verification. Who in the industry do you think is most wrong about where the value actually gets created?

The design-acceleration thesis is not wrong. It is incomplete.

As design generation and physics simulation get cheaper, validation across the stack becomes the scarce resource. AI will help engineers generate more concepts, explore larger design spaces, and run simulations faster — that part is real. But the harder question is no longer whether AI can generate more candidate designs. It is which of those designs can actually be manufactured, assembled, changed, inspected, sourced, and scaled in the real world.

Manufacturing has always had this asymmetry: process capabilities often advance faster than our ability to design for them. The manufacturing stack can do more than the surrounding engineering and planning workflows can reliably support. AI-assisted design accelerates one side of that asymmetry without addressing the other.

So the larger opportunity is not pushing more designs downstream into manufacturing. It is pushing manufacturing intelligence upstream into design, so engineers can make targeted decisions that minimize feedback loops before production. The systems that create the most value will not simply generate more options. They will help engineers generate options that already understand manufacturing feasibility, assembly logic, change impact, and production readiness.

The next bottleneck is not imagining more designs. It is proving which ones can become real products.

To contact Sai, reach out to him on LinkedIn here. He’s always open to chatting and sharing valuable insights.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Jiga, Upkeep, and Industry 4.0 Club.

Diagon: How a Simple Spreadsheet Became the World's Most Comprehensive Equipment Marketplace 🎙️

You were one of the first formal buyers Tesla hired specifically to source factory machinery. What did “professionalizing” that role actually look like during Model 3 production hell? And how has that experience, coupled with your work at Astra and Bright Machines, driven the early vision for Diagon?

I actually started this role in 2013, during the Model S ramp and the design and build of the Model X production lines, which is a really interesting place to begin the story.

When I joined, my job was to intake requirements for every equipment project across the company. The first thing I needed was an organized system, a single dashboard where every project’s requirements were documented. From there, I began identifying qualified suppliers, which meant building spreadsheets and Confluence pages cataloging every equipment manufacturer that sold the machines and systems we needed. Then came standardizing how we negotiated contracts and got spending under terms. When I started, very few projects were under any kind of general T&Cs. By the time I left, nearly all of them were.

Diagon started as a spreadsheet! That spreadsheet of manufacturers and machines became the company's single source of truth. Something special happened along the way. Engineers were constantly coming to me for supplier recommendations, and I could quickly answer questions about which companies made which machines, what they typically cost, what the lead times looked like, and what the specs were. The same dataset eventually tracked warranties, financing options, and other supplier details. Engineers, the finance team, and even the CFO used it to build forecasts and cost models for future factory projects.

That idea has stuck with me. At Diagon, we’re building the largest open dataset of manufacturing equipment ever attempted, and we’re layering AI on top of it to create the most powerful equipment sourcing tools available anywhere.

Albus is designed to give a single SME engineer the procurement sophistication of a Tesla or Boeing sourcing team. But those teams carry decades of tribal knowledge. How does an AI agent replicate supplier judgment, not just supplier search?

We’ve built a fully autonomous equipment sourcing agent named Albus, and a few things give him real judgment, not just search.

First, he has access to our proprietary dataset of equipment and suppliers, so he simply knows more about the industry than the average equipment buyer. Second, because he’s autonomous and connected to platforms like LinkedIn, Practical Machinist, Reddit, and other forums, Albus learns from many experts and SMEs at once. A typical SME only has the experience they’ve personally lived through. Albus can synthesize across thousands of practitioners to make far better-educated predictions about the right machines for a given application. Third, he has access to a deep pool of used and pre-owned equipment, which helps him surface genuinely great financial deals for customers.

The “One Big Beautiful Bill” offers 100% immediate write-offs on equipment purchased after January 2025 and 35% tax credits for semiconductor gear. Are manufacturers actually pulling purchases forward in response, or is the industry too slow to capitalize?

My honest read is that companies are treating this as an accounting benefit. It helps clean up financial statements and improves how their position shows on paper. But at the end of the day, actual financing dollars are what drive purchase decisions. So the write-offs have only helped the industry moderately. The real movers are venture capital, private equity, and debt financing. As that financing has become more available, that's been the major driver of purchasing activity I'm seeing.

You’ve said the real bottleneck in industrial software is culture, not technology, and that Diagon focuses on building software “frontline workers actually want.” What does the onboarding conversation sound like with a skeptical plant manager who’s been buying machines the same way for 20 years? And once they’ve made a purchase, how do you keep them engaged on the platform?

Everybody hates new tools. Nobody uses Diagon because they want an equipment sourcing platform. They use Diagon because they need a CNC machine or a laser cutter. The platform is a means to an end. It’s one of the fastest ways to go from a concept of what you want to source to a list of viable machine options and live listings.

For skeptical plant managers, what’s worked best is just doing a demo, especially a live sourcing session. If they were recently in the market for, say, a waterjet cutter, walking them through that flow is the best pitch we have. Diagon is free for buyers, so there’s really no risk in trying it. They get the accuracy of going directly to manufacturer websites, combined with the flexibility of an AI platform like Claude. Albus (Diagon’s CAPEX Sourcing Agent) invites them to collaborate on the Diagon platform and evaluate the options that best suit their needs.

Five years out, does Diagon look more like an industrial marketplace, a procurement operating system, or a financing platform, or is the thesis that those three things are inseparable? How will this vision change the workflows and operating model for your customers?

Honestly, none of the above. Five years out, I think Diagon becomes an agent-friendly equipment information API, and the mode of interaction looks much more like that of a trusted advisor or even a colleague. Customers will move seamlessly from identifying the machines they’re looking for (which feels more like a procurement operating system), to finding what’s actually available on the market (the marketplace piece), to lining up financing and shipping. It will be one of the most powerful tools any manufacturer can have on the floor.

Today, we’ve built Diagon to fit within the existing process: engineers define specs, a procurement leader identifies vendors, an equipment buyer reaches out for quotes, and then compares, negotiates, and selects. In the future, I think this looks much more like iterative design. Engineers will be far better informed upfront on the machine options available, which means smarter decisions on equipment specs and even product design, and ultimately less capex spend. Equipment buyers will come to see Diagon as the place where the work currently happening in inboxes and on phone calls actually gets executed.

To contact Will, reach out to him on LinkedIn here. He’s always open to chatting and sharing valuable insights.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Jiga, Upkeep, and Industry 4.0 Club.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Jiga, Upkeep, and Industry 4.0 Club.

A Design and Simulation Market Map 🗞️🔬 📚

Enjoyed building this with Michael Finocchiaro at Threadmoat. This is the start of a new series featuring market maps of engineering and manufacturing startups shaping the landscape today. For more analysis of this ecosystem and more on each of these startups, check out ThreadMoat and visit Fino’s profile! As always, let us know if we’re missing anyone!

Some Interesting Reads:

Keychain surveyed 1,000 U.S.-based CPG manufacturers, with 52% planning to add new production lines and 25% planning to open new facilities. In terms of AI adoption, 12.3% have already deployed it, 18.4% are piloting it, and 22.1% are planning to implement it.

• Cerebras, which went public this week, warned that the US Chip manufacturing rebuild could take up to 15 years [WSJ]

• An interesting breakdown and analysis from Joris Feels on the AI CAD tools best suited for 3D Printing. [3DPrint]

• Fanuc announced a partnership with Google, unsurprisingly shares surged [Bloomberg]

• Atlas shared the Q1 2026 State of Clean Energy Manufacturing report. Surprised by the significant drop in battery investment given the energy bottleneck, though unsurprised by the cancellation of 48,000 announced manufacturing jobs. [Atlas Public Policy]

• KPMG released its industrial manufacturing tech report. 48% plan to significantly increase cybersecurity investment, and 70% are taking a centralized approach, with IT leading AI implementation. I’m personally skeptical of this approach, AI’s value should be democratized to business owners. [KPMG]

  

• Emerging approach? Stellantis and Ford announce partnerships with Chinese carmakers to survive in a European market squeezed by high costs and weak demand [FT]

• China’s BYD leads in silicon-carbon battery development, a technology that the US invented, and companies like Sila Nanotechnologies have been working on for years. A great breakdown of another technology developed by the West, and productionized by China. [The Information]

• An AWS solution breakdown for capturing multi-modal embeddings from manufacturing-specific data, such as engineering diagrams, CAD drawings, inspection photographs, thermal analysis plots, and fatigue curves. [AWS]

• a16z shared that 20% of manufacturing firms are using agents [a16z]

• BCG’s CEO Guide to Physical AI [BCG]

Products & Research:

Figure’s livestream drew 1.5 million views on X over its first eight hours. While exciting, I’d be curious to see how this performed relative to the first generation of warehouse automation, and it looks like the Agility CEO agrees. Also interesting that the robot didn’t optimize for the barcode being up on a lot of the parcels. [Figure]

• Config launches building the data layer for robotics foundation models [TechCrunch]

• Alcon case study, with HighByte serving as the foundation for predictive maintenance and defect-monitoring use cases. [HighByte]

• Siemens and Xometry announced a $50 million partnership, including the integration of Supplyframe, which provides design-to-source information for electronic components, and Thomas’s sourcing data. [Engineering.com]

• Mind Robotics raises another $400M for Manufacturing robots [Robot Report]

• Microsoft updated MatterSim, a multi-task foundation model for in silico materials characterization [Microsoft]

• Chef Robotics now automates secondary packaging and kitting automation for CPG manufacturers and customers. [Chef Robotics]

• Perceptron released MK1, a purpose-built system for video understanding and embodied reasoning

  

Finance & Transactions 💵

Anduril raised another $5B from Thrive Capital and Andreessen Horowitz
Aerospace & Defense • Series H • USA

Mind Robotics raised $400M from Kleiner Perkins, Meritech, and Redpoint
Robotics / Physical AI • Funding • USA

RangerAI raised $8.4M from Bonfire Ventures and 25madison
Industrial AI • Series B • Canada

Planned Downtime 🧑‍🔧

A24 and Bourdain, Sign Me Up

Breaking the Bottleneck is brought to you with support from our partners.

• AMT & IMTS - The Association for Manufacturing Technology represents and promotes U.S.-based manufacturing technology and hosts the largest manufacturing technology trade show in North America, IMTS.

• Jiga - Is a platform that connects hardware teams with vetted manufacturers for custom parts. Instead of black-box instant quotes, you see exactly who’s making your parts and can talk to them directly.

• Upkeep - CMMS Software helping maintenance teams manage work orders, assets, and preventive tasks efficiently from any device.

• Industry 4.0 Club - Leading cross-functional group accelerating the adoption of Industry 4.0, delivering better experiences to consumers, better profits to manufacturers, and better jobs to factory workers.

• Ironloop - Ironloop is a modern OT/ICS software platform that delivers configuration-level intelligence, giving manufacturers ground-truth security visibility and compliance confidence across air-gapped networks.

• T51 - T51 is an AI-native planning, scheduling, and optimization platform that integrates securely with your existing ERP, MES, and data stack.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Jiga, Upkeep, and Industry 4.0 Club.

Scaling AI Agents Is Less About Building More Agents and More About Designing the System Around Them

You went from programming systems on factory floors to P&L accountability at Honeywell, then led global business units at Rockwell. Now you’re at a hyperscaler selling into the same customers you used to serve. How does sitting on the Microsoft side change the way you think about what manufacturers actually need versus what they’ll buy?

What changed for me isn’t the definition of “need,” but the scale of what’s possible when you can connect data, AI, and workflows across the entire value chain. On factory floors, “need” is still the same: safer operations, better quality, more throughput, and lower cost. If a solution can’t move one of those needles, it’s noise. Now sitting on the Microsoft side, I’m even more passionate about practicality.

Manufacturers don’t buy AI to check a box; they invest in outcomes they can measure in months, not years. And it has to show up in the way people actually work: the maintenance workflow, the scheduler’s daily decisions, the quality engineer’s root-cause loop—not another dashboard. What excites me now is being able to bring that together on a single platform, connecting data, AI, and workflows across the entire value chain in a way that wasn’t possible in fragmented systems.

Microsoft has deployed over 25 AI agents inside its own logistics operations, with a target of more than 100 by year-end. What has that internal experiment taught you about scaling agents that you wouldn’t have learned from customer deployments alone?

The biggest lesson is that scaling agents is less about building more agents and more about designing the system around them. In real operations, speed matters, but the ability to coordinate actions across workflows matters more.

Internally, you see quickly where agents deliver value today: monitoring inventory signals, tracking supplier commitments, flagging issues early, and triggering actions, not just insights. That shift from recommendation to execution is critical. Especially in supply chains where conditions are constantly changing, supplier delays, demand shifts, and logistics constraints, agents help teams respond in real time, not just react after the fact.

You also learn the harder part: how agents work together, how they integrate into existing processes, and where to lean on human intervention. Scaling isn’t just about autonomy; it’s about orchestration. If people don’t have clarity on “what should happen next and why,” you won’t build trust or adoption at scale. That is where a unified platform becomes critical, so agents are not isolated, but operating on shared data, aligned to the same workflows, and able to drive coordinated action across planning, supply, and production.

You’ve set a six-month goal for delivering measurable AI impact. What is the most common reason you see industrial AI projects fail to hit that mark, and at what point should a leader kill a stalled pilot rather than pivot it?

The most common failure indicator is trying to do too much at once with no clear measurement tied to throughput, quality, safety, or working capital. AI doesn’t fail because the model isn’t clever; it fails because it’s not embedded into the daily operating rhythm. What we are really talking about is a shift in operating model, from periodic, human-driven decisions to continuous, AI-assisted decision loops embedded directly into those workflows.

If it lives only in a dashboard, it won’t move the needle. It has to show up inside maintenance, scheduling, or quality workflows where decisions get made every day. As for “kill vs. pivot”: I’d give any pilot a clear checkpoint at 6–8 weeks. If you can’t get reliable data, real user engagement, and a leading indicator that the outcome is achievable, you don’t have a pilot, you have an experiment. Pivot if the outcome is right, but the workflow, data source, or scope is wrong. Move on if you can’t tie it to a real operational constraint and an owner who will run it after launch.

You frame AI as a “workforce multiplier,” not a replacement. But sentiment around AI is extremely negative right now, especially here in the United States, and frontline workers hear “autonomous operations” and immediately think layoffs. How do you address that cultural resistance in practice and get them to buy in, particularly when they’re the ones providing the critical context that agents on the factory floor need to function?

You earn buy-in by starting with respect for the people doing the work. I’ve stood next to technicians and operators who don’t want another screen, they want clarity. So we position AI as reducing friction: fewer fire drills, fewer surprises, faster root cause, and more time spent on high-skill work that only humans can do. Practically, that means involving frontline experts early, designing the system to explain “what to do next and why,” and putting clear guardrails in place so people understand where autonomy ends and human judgment begins. Often, the hardest part isn’t the technology, it’s change management. If leaders want trust, they need to pair AI investment with upskilling and a consistent vision, and they need to measure success in operational terms that matter to the team. That only works if the tools show up where people already work, in their daily workflows, not as separate systems they have to learn.

You talk about connecting every stage from design and simulation to production feedback, so manufacturers can test decisions before disruptions cascade. The “digital thread” is a vision the industry has been chasing for a decade. What is actually different now? Is it the AI layer, the data infrastructure, or something else that makes this achievable today when it wasn’t three years ago?

What we are describing is industrial intelligence, the ability to continuously connect data, AI, and workflows across the value chain and act on it in real time. What’s different now is the convergence of three things. First, the data foundation is finally catching up; more plants can bring together production, quality, maintenance, and supply signals in a usable way. Second, AI has advanced from “reporting what happened” to understanding patterns and cause-and-effect across silos, so you can see what is happening, not just what should be happening. Third, planning is moving closer to execution. Instead of static forecasts, teams can run frequent what-if scenarios using live production data, supplier updates, and demand shifts, so decisions get tested before disruptions cascade. So yes, the AI layer matters, but only because it turns the digital thread from a vision into something you can operationalize day to day. It is the combination of data platform, AI, and the applications people use every day that makes this practical, not just theoretical.

This newsletter is brought to you with support from our partners: Upkeep, Jiga, Industry 4.0 Club, Ironloop, and T51.

This newsletter is brought to you with support from our partners: Jiga, Upkeep, Industry 4.0 Club, Ironloop, and T51.

Content I Enjoyed Last Week 🗞️🔬 📚

Some Interesting Reads:

Capgemini surveyed 1,678 senior executives across 15 industries for its new 140-page Physical AI report, and the headline finding is that 79% of organizations are already engaging with the technology in some form. Below are the slides that caught my attention.

• A great case study from Bristol Myers Squibb, where AI now monitors bioreactor variables such as temperature, pH, and oxygen in real time, intervening mid-batch to reduce time to market and increase yield. [WSJ]

• 94% of manufacturing respondents reported an increase in freight spend in 2026, and 82% agree that maintaining their current procurement practices puts their organization at risk. [Xeneta]

• AI Adoption in Manufacturing with the COO at Ford [Google]

• CloudNC founder only now realizes AI is an enabler of value in manufacturing, and not where the value accrues. [Fortune]

• A great interview with Brad Porter, highlighting that the real value of Physical AI is converting labor into a digitally controllable operating layer consistently reprogrammable, akin to software. [WSJ]

  • argues something similar here in a post about the Physical AI application layer.

• NVIDIA and Corning announce a partnership that would result in 3,000 new jobs and 50% increase in production capacity [NVIDIA]

• Apple Manufacturing Academy is expanding its reach to help SMB manufacturers implement AI and smart manufacturing techniques [Apple]

• Oxford’s Environmental Change Institute found that 85% of industrial energy could be electrified by 2030, if the grid interconnect problem were addressed [Canary Media]

• An interesting conversation with Allentown, PA, Mayor Matthew Tuerk about the city's grand strategy for building back and sustaining its manufacturing base, implementing industrial policy at the local level, and how zoning has changed over the last decade. [Bloomberg]

• York Yang and Dyna Robotics highlight the deployment and integration hurdles that must be overcome to enable widespread robotics deployment. [X]

• Bosch’s Shopfloor Agent drives annual savings of around €850,000 for each Bosch plant by reducing unplanned downtime. [Bosch]

• Research with an amazing breakdown of the Humanoid Robotics ecosystem, well worth the read if you have the time.

  Citrini Research

  Humanoid Robots

  Yes, robots…

  Read more

  a year ago · 284 likes · 49 comments · Citrini

Products & Research:

Vinci proposes a new category called Continuous Physics Reasoning, enabled by a physics foundation model. The model should reason directly over physical structure rather than abstracted features, be deterministic, hit solver-grade correctness at manufacturing resolution, generalize to new geometries and materials out-of-the-box with no customer-specific training, operate as a single global model with no per-customer forks or forward-deployed engineering, and run efficiently across scales.

• Autodesk AI Lab released 1 million executable, interpretable CAD sequences synthesized entirely without real-world data [Autodesk]

• Julia Hub plans to demo Dyan 3.0, an AI-first environment to model, test, and validate industrial systems, on May 19 [JuliaHub]

  

• CoLab announced a partnership with Bombardier [CoLab]

• Nominal acquired FID Labs to connect the hardware data supply chain [Nominal]

• Emanate emerges from stealth, focusing on handling quoting, sales, and procurement across the industrial materials ecosystem [Fast Company]

• Genesis AI launches GENE-26.5, a foundational model with a dexterous robotic hand that mirrors the human hand exactly, after capturing proprioception, tactile, and action-training data via a data-collection glove.

Finance & Transactions 💵

Tessera Labs raised $60M from A16z and Foundation Capital
Agentic ERP • Series A • USA

Nova Intelligence raised $31M from Chemistry
Agentic ERP • Series A • USA

Spread AI raised $30M from DTCP Growth, IQT, and others
Industrial AI • Series B • Germany

Planned Downtime 🧑‍🔧

Robotic’s End Game

Breaking the Bottleneck is brought to you with support from our partners.

• Jiga - Is a platform that connects hardware teams with vetted manufacturers for custom parts. Instead of black-box instant quotes, you see exactly who’s making your parts and can talk to them directly.

• Upkeep - CMMS Software helping maintenance teams manage work orders, assets, and preventive tasks efficiently from any device

• Industry 4.0 Club - Leading cross-functional group accelerating the adoption of Industry 4.0, delivering better experiences to consumers, better profits to manufacturers, and better jobs to factory workers.

• Ironloop - Ironloop is a modern OT/ICS software platform that delivers configuration-level intelligence, giving manufacturers ground-truth security visibility and compliance confidence across air-gapped networks.

• T51 - T51 is an AI-native planning, scheduling, and optimization platform that integrates securely with your existing ERP, MES, and data stack.

💥 If you want to chat, feel free to reach out at aditya@machinafactory.org.

🏭 If you were forwarded this and found it interesting, please sign up!

This newsletter is brought to you with support from our partners: Jiga, Industry 4.0 Club, Ironloop, and T51.

Green Dashboards Don’t Mean Your Plant Is Actually Running Better🎙️

You were a process engineer at a membrane plant watching technicians walk back and forth to desktop terminals just to log work orders. What was the moment you decided to build the fix yourself rather than wait for a vendor to solve it?

I wouldn’t say there was one big dramatic moment where I just woke up and decided, “I’m going to go start a company.” It was honestly more the accumulation of seeing problem after problem in the physical world, and at the same time realizing that if you wanted to actually build something in that world, everything moved painfully slowly.

I went into engineering because I loved to build. One of the things that hit me pretty quickly working in a manufacturing plant was that in the physical world, even when you know exactly what should be better, making a change can take months or years because you are dealing with real equipment, real processes, real constraints, real safety considerations, and a lot of inertia. At the same time, I was seeing all these very obvious workflow problems around me, especially for technicians and frontline teams, where the software clearly was not built for how they actually worked. Watching someone walk back and forth to a desktop terminal just to log a work order sounds small, but it reveals how disconnected the software was from the real job.

What pulled me in was that I’ve always enjoyed tinkering and building. Once I started teaching myself to code, my perspective changed entirely because, for the first time, I had a way to see a problem during the day, build a solution that night, and potentially release it the next morning. That was completely different from what I had experienced in the plant, where the cycle time for change was so long. So it was not really one moment so much as a growing realization that there were all these problems in the physical world worth solving, and software was the fastest tool I had ever found to actually go solve them.

You talk about shifting maintenance KPIs from “work orders closed” to “uptime impact on margin.” What’s the most compelling example you’ve seen of a customer actually tying maintenance data to revenue? And for companies earlier in their data maturity, what are the first steps to getting on that path?

One of the examples that stuck with me the most was visiting a manufacturing customer who said something insightful: they did not actually care about PM compliance or closed work orders as the end goal. That may sound surprising coming from a maintenance organization, but their point was that those metrics alone can drive the wrong behavior if you are not careful.

What they had seen over time was that when you tell teams the goal is to close more work orders or hit a certain compliance percentage, people get very good at optimizing for the check box, and you can end up in a culture where everyone is creating activity and green dashboards, but not necessarily creating better plant performance. In other words, you can make yourself feel productive without actually moving the business. What they did instead was tie everyone back to production outcomes, so maintenance, operations, and reliability were all fundamentally being measured against the same thing, which was keeping the plant running, driving throughput, and supporting the output of the business. Once you do that, the conversation changes because maintenance is no longer isolated, measuring success in a silo; it directly connects to what the business actually cares about: production, revenue, and margin.

To me, that is the right mental model. Maintenance metrics still matter, but they should be understood as drivers of business performance, not the end state themselves.

For companies that are earlier on this path, I actually think the first step is less technical than people think and much more organizational. Before you get into AI, dashboards, or complicated data models, you need shared goals. If maintenance is measured one way, operations is measured another way, and reliability is measured a third way, then by definition, you are going to create competing behavior. Maintenance reduces downtime now, reliability reduces downtime in the future by improving how assets perform over time, and operations is driving the day-to-day output of the plant, but all 3 functions really need to understand how they connect to the same business outcome. Once you have that, the technology becomes much more powerful because you are now using the data to drive a single scoreboard instead of 3 different ones.

You’ve said many plants are still running technology from the 1990s. What’s the single biggest non-technological barrier keeping them there, whether cultural, financial, or regulatory? And what go-to-market strategies have you found actually break through that resistance?

I think the biggest barrier is that the cost of disruption is much higher in the physical world. Because of this, people are a lot more cautious, and honestly, they should be.

In software, if you roll something out and it is a little messy, that is frustrating, but in a plant, a warehouse, or a facility environment, when something breaks, the consequences can be much bigger. You can affect production, create safety issues, or slow down a team that is already stretched thin. Consequently, many organizations stick with bad systems not because they love them, but because those systems are familiar, and the pain of change feels more dangerous than the pain of living with the status quo.

So yes, culture is part of it, but I think just calling it “cultural resistance” undersells what is really happening. The physical world moves more slowly because the downside of getting it wrong is real. That is why old systems survive way longer than they should.

From a go-to-market standpoint, the thing that works best is not leading with some giant abstract transformation message and trying to force it top down. What I have found works much better is finding the early adopters inside the business, the people who are naturally curious, ambitious, and frustrated enough with the way things work today that they actually want change. Every company has these people. They are usually the ones who are trying to do more, trying to improve things, and are constrained by old tools. If you help those people win, and you give them something that actually makes their life better and produces a visible result in the real environment, then they become your champions. That is when adoption starts to spread in a much more durable way, because now you are not selling hope, you are showing evidence.

UpKeep Studio lets local site managers build their own apps with a no-code AI builder. How do you prevent that from creating a new mess of fragmented, inconsistent tools across a multi-site operation? How do you balance flexibility at the factory floor with the need for enterprise-level structure?

That is honestly the whole point of Studio. We repeatedly heard from customers that every business and site operates differently, with unique workflows. Much of the software they receive forces them to contort their business around the tool instead of letting the tool adapt to the business.

Now with AI and modern tooling, it has become easier than ever to build apps or prototype workflows quickly, and that is exciting, but that is also exactly where the mess can start if you are not careful. The easy part now is building something fast. The hard part is everything that comes after: How do you deploy it into production? How do you ensure it scales? How do you manage permissions? How do you keep the data model clean? How do you prevent every site from spinning up random disconnected tools that break consistency across the organization?

That is why Studio matters. Yes, it gives teams the ability to build and move fast, but it does it on top of UpKeep’s architecture, infrastructure, data model, and permissions layer, which means you are not just creating random one-off apps floating around outside the system. You are building on top of a shared foundation and shared source of truth. In my mind, the balance is that local teams get the flexibility they need because they have unique workflows, while the enterprise still gains structure, consistency, governance, and scale.

That was the vision from the beginning. We did not want customers to have to choose between rigid software that does not fit their business and total chaos, where everyone builds their own disconnected tools. Studio is meant to give them flexibility without fragmentation.

As the U.S. pushes to reshore manufacturing, what does the maintenance technician’s role look like in a facility designed from the ground up to be highly automated and AI-integrated? For generative AI specifically, what guardrails should be built around “Autonomous Actions,” such as those in Nova, for the technology to be effective in safety-critical environments?

I think the technician becomes more important, not less. I believe much of the mainstream conversation around AI misses this point because it is too software-centric and not grounded enough in what work in the physical world actually looks like.

AI is not going to climb a ladder, it is not going to turn a wrench, it is not going to walk a site and notice something feels off in the way an experienced technician does. In safety-critical environments, there is still an enormous amount of judgment, physical presence, situational awareness, and real-world accountability required, and I think that remains true for a long time. What AI can do, though, is make that technician dramatically more effective by helping prioritize, helping triage, surfacing patterns earlier, automating the administrative layer of the job, and helping the team make better decisions faster. So instead of replacing the technician, I think what it really does is elevate the role and increase the leverage of the best people.

That is why I actually think this is one of the most exciting moments for people in the industry. The winners will be the technicians, supervisors, operators, and reliability leaders who learn how to use AI and software to augment themselves and move faster, not those who assume this technology is just hype and ignore it.

On the guardrails side, especially in safety-critical environments, I think the wrong way to think about autonomous actions is as unchecked automation. The right way to think about it is in terms of layered trust. There are some things the system should absolutely be able to suggest, some things it should be able to prepare, some things it should be able to execute with approval, and some things that should always remain gated by a human. You need clear permissions, clear audit trails, and a very clear understanding of what the system did, why it did it, and what data it used to reach that conclusion. In these environments, explainability matters, accountability matters, and human oversight matters. So I think AI becomes incredibly powerful here, but the near-term model is not “remove the human,” it is “make the human far more effective while keeping the right safety and trust boundaries in place.”

To contact Ryan, reach out to him on LinkedIn here. He’s always open to chatting and sharing valuable insights.

This newsletter is brought to you with support from our partners: Jiga, Upkeep, Industry 4.0 Club, Ironloop, and T51.

Content I Enjoyed Last Week 🗞️🔬 📚

Some Interesting Reads:

with an interesting analysis of the WSJ article stating that we were in a “stealth manufacturing boom.” The uptick is the result of the AI CAPEX boom, with labor data supporting the clear bifurcation in job growth between the advanced manufacturing (semiconductors, electrical equipment, and power gear) and the traditional manufacturing ecosystem.

Noahpinion

No, America is not in a "stealth manufacturing boom"

Greg Ip of the WSJ is one of my favorite economics writers, and you should always read what he writes. But in a recent post about manufacturing, I think he gets the main narrative wrong. Greg writes that America is in the middle of a “manufacturing revival”, which his headline writer calls a “stealth manufacturing boom…

Read more

4 months ago · 287 likes · 26 comments · Noah Smith

• A BCG analysis of the applications of AI across different companies in the Aerospace and Defense industry. A quick snapshot of some of those initiatives is below.

  

• A case study with GE Appliances and Gemini embedding Gemini Enterprise into its Brilliant Factory manufacturing data platform. [Google]

• Semantic context architectures drive a 42% TCO reduction versus stovepipe architectures in manufacturing [Kearney]

• Bain’s 2026 industrial automation analysis argues that value in the sector is migrating from the traditional control layer toward AI-enabled software and data platforms at the top, and intelligence-embedded smart field devices at the bottom. [Bain]

• KPMG released its Industrial Manufacturing Global Tech Report, with 76% of respondents citing unreliable data as a top AI risk, and 48% of executives plan significant increases in cybersecurity investment. [KPMG]

• Bessemer’s Robotics and Physical AI predictions for 2026. [BVP]

• An interview with the CEO of Neura Robotics on the Sim-to-real gap and the European opportunity for Physical AI. [McKinsey]

• A great breakdown of what’s needed to drive effective AI adoption in manufacturing: a structured digital model of the value stream as context, a closed-loop plan-versus-actual feedback mechanism anchored to that value stream, and human-controlled digital actions. [Threaded]

• The significance of the workforce availability problem on reshoring, tariffs, tax policy, and deregulation. [Dynamo Ventures]

• Ford’s joint venture with Geely is politically fraught but strategically interesting. Rather than competing head-on with China’s manufacturing machine, it creates an opportunity to learn from it directly. [WSJ]

• A pretty neat breakdown of Proto Town and the ecosystem of industrial tech startups that are using it as a test bed for building technologies [Bloomberg]

Products & Research:

The architecture behind Google’s 8th-generation TPU with two purpose-built systems, 8t for pre-training and 8i for inference and reasoning. 8t focuses on 47 petabits/sec of higher-bandwidth and 10x faster storage access, with an accelerator to reduce the bottleneck in the matrix multiplication units (see the ASIC block diagram below).

The 8i architecture cuts the chip-to-chip network diameter from 16 hops to 7, adds 3x more on-chip SRAM, and includes an acceleration engine that reduces on-chip latency by 5x (see the ASIC block diagram below). [Google]

Hannover Messe Highlights:

• Sight Machine launches Agent Crews to simulate possibilities, determine optimal settings, and make recommendations. [Sight Machine]

• Litmus launches Data Catalog to drive visibility, lineage, and governance when managing industrial data. [Litmus]

• Siemens released the Eigen Engineering Agent to help with common engineering tasks such as PLC coding, HMI visualization, and device configuration. [Siemens]

• Bobcat uses Arbortext within PTC Windchill to create and manage BOM‑linked technical publications and to help optimize the service supply chain. [PTC]

• Rockwell demonstrated a new AI Factory System design tool with the Emulate3D digital twin, Copilot in VSCode as an AI‑assisted engineering interface, and FactoryTalk Design Studio. [Rockwell]

• NVIDIA demonstrates the future of AI-driven manufacturing at Hannover Messe 2026 alongside a trove of partners [NVIDIA]

• NVIDIA Omniverse libraries and the open source NVIDIA Isaac Sim robotics simulation framework are now available on Google Cloud Marketplace [NVIDIA]

• ABB revealed Genix Industrial AI on Microsoft Azure, which acts as a real-time industrial co-pilot on the factory floor by analyzing streaming data from equipment and sensors. [Microsoft]

• SAP revealed a Production Master Data Agent and Production Planning and Operations Agent [SAP]

• Autodesk Fusion MCP and Fusion Data MCP launch to connect tools, help with understanding, and manage designs in context. [Autodesk]

Blogs and Threads

on the core bottleneck in automated engineering simulation: the cognitive and setup costs of building parametric automation logic that connects the intermediate steps, and the obsolescence of that logic once pipelines change.

Generative Engineering

Removing the bottlenecks in Engineering AI

Engineering systems are getting more complex, and at the same time the pressure to ship them faster is increasing. Teams that can explore more of the design space and iterate faster have a decisive advantage…

Read more

4 months ago · 4 likes · 1 comment · Laurence and Generative Engineering

thoughts’ on Mythos in Manufacturing

Breaking the Bottleneck

Manufacturing Mythos

Breaking the Bottleneck is a weekly newsletter and interview series covering manufacturing technology and physical AI. Want to chat? Reach out at aditya@machinafactory.org or connect with me on Linke…

Read more

4 months ago · 18 likes · 3 comments · Manufacturing Tech

Finance & Transactions 💵

Project Prometheus is raising $10B at a $38B from JP Morgan and Blackrock
Physical AI / Robotics • Capital • USA

Humble raised $24M from Eclipse
Physical AI / Autonomous Freight • Seed • USA

Rillian raised $17.5M from 8VC
Cybersecurity • Seed+ • USA

C-Infinity raised $16M from Canaan
CAD/Manufacturing • Series A • USA

Antioch raised $8.5M from A* and Category Ventures
Physical AI / Robotics • Seed • USA

Crewline raised $7.1M from Initialized Capital and Nebular
Physical AI / Autonomous Construction • Seed • USA

Planned Downtime 🧑‍🔧

Breaking the Bottleneck is brought to you with support from our partners.

• Jiga - Is a platform that connects hardware teams with vetted manufacturers for custom parts. Instead of black-box instant quotes, you see exactly who’s making your parts and can talk to them directly.

• Upkeep - CMMS Software helping maintenance teams manage work orders, assets, and preventive tasks efficiently from any device

• Industry 4.0 Club - Leading cross-functional group accelerating the adoption of Industry 4.0, delivering better experiences to consumers, better profits to manufacturers, and better jobs to factory workers.

• Ironloop - Ironloop is a modern OT/ICS software platform that delivers configuration-level intelligence, giving manufacturers ground-truth security visibility and compliance confidence across air-gapped networks.

• T51 - T51 is an AI-native planning, scheduling, and optimization platform that integrates securely with your existing ERP, MES, and data stack.