Presented by
Live this Thursday, November 5 at 12:00 PM CT (1:00 PM ET): Walker Reynolds and Paessler. Details in the Sponsor Message below.
We Gained an Hour,but Technology Isn't Slowing Down.
Well, Halloween is behind us, the candy bowl is probably empty, and for those of us in the States, Daylight Saving Time ended Sunday. We got an extra hour of sleep, but judging by the pace of technology news, nobody told the tech industry to slow down.
Google just introduced an AI model designed to tackle complex engineering and cybersecurity problems. SAP is pushing AI agents deeper into enterprise operations. Meanwhile, Wall Street is starting to ask an uncomfortable question about all this artificial intelligence: When does the investment actually start paying off?
For manufacturers, these aren't just technology headlines. They're questions about architecture, security, investment, and how we put emerging technologies to work without creating more problems than we solve.
Here's what caught our attention this week:
Google's Gemini 4 Carbon: AI Is Moving Beyond Chatbots
Remember when AI was mostly about generating emails, summarizing documents, and answering questions? Those days are disappearing fast.
On October 28, Google announced Gemini 4 Carbon, its most capable AI model yet, designed to tackle complex engineering problems, write and debug code, and identify cybersecurity vulnerabilities. But the real story isn't just another AI model topping another benchmark. It's how much work these systems are beginning to perform independently.
The details:
Google says Gemini 4 Carbon can handle what it calls long-horizon tasks, meaning complex assignments requiring multiple steps, sustained reasoning, and the ability to work through problems without constant human intervention.
During internal testing, Google's teams observed the model working for days to identify and repair subtle software bugs across large codebases.
Its reported cybersecurity benchmark results are particularly interesting:
-
88.2% on SEC-bench Pro: Identifying vulnerabilities in real-world software.
-
71.0% on CWE-bench: Producing working code patches that address underlying vulnerabilities.
-
54.8% on ExploitGym: Testing whether software vulnerabilities can actually be exploited.
These are benchmark results, not guarantees of real-world performance. And because cybersecurity capabilities can be used defensively or offensively, Google is initially limiting access to selected cybersecurity defenders while conducting additional safety testing.
Why it matters for manufacturing:
Think about the software running in a modern manufacturing facility. You've got PLCs, SCADA systems, MES applications, ERP integrations, databases, APIs, and custom code holding everything together.
Now imagine an AI agent capable of reviewing an entire integration project, identifying potential vulnerabilities, tracing errors across multiple services, and proposing fixes.
That's a fundamentally different capability from asking a chatbot to explain a Python script.
But there's an important distinction: Being able to analyze industrial software doesn't mean an AI should automatically be allowed to modify production systems.
Here's how this could work on the factory floor:
An engineer is troubleshooting an MES integration that occasionally loses transactions between the production system and ERP.
Instead of manually reviewing thousands of lines of application code, logs, and API configurations, an AI agent could analyze the relevant software, identify a possible failure condition, and propose a correction.
The engineer would still need to validate that change in a test environment before anything touches production.
That's where the real opportunity lies: using AI to accelerate engineering work while keeping operational responsibility with the people who understand the process.
The bottom line:
AI is evolving from something that answers questions into something that can investigate problems and perform meaningful engineering work. For manufacturers, the opportunity is significant, but so is the need for proper governance, testing, and human oversight.
Read Google's Gemini 4 Carbon announcement
SAP Wants AI Agents Running Your Business. Is Your Data Ready?
SAP isn't just adding AI to its software anymore. It wants AI agents executing business processes across the enterprise.
At SAP TechEd on October 27, the company announced major expansions to Joule Studio, Joule Work, and its Business AI Platform. The goal? Move beyond AI assistants that answer questions toward agents that can reason, make decisions, and execute multistep workflows across business systems.
That's a significant shift, particularly for manufacturers running SAP alongside MES, SCADA, and other operational technology platforms.
The details:
SAP introduced several capabilities designed to make enterprise AI more autonomous:
-
Joule Studio: Developers and business users can build and extend AI agents using no-code and pro-code tools, with access to more than 70 AI models.
-
Joule Work: A common interface for coordinating AI agents and executing workflows across SAP and other enterprise applications.
-
Agent interoperability: Support for Model Context Protocol (MCP) and Agent2Agent (A2A), allowing agents to interact with tools, systems, and other agents.
-
SAP Knowledge Graph: Expanded capabilities for connecting business data and providing the relationships and context AI needs to reason accurately.
-
Governance and monitoring: New controls for managing agent access, tracking activity, monitoring costs, and enforcing enterprise policies.
SAP also introduced updates to its tabular AI models, including capabilities for anomaly detection, forecasting, and analyzing structured business data.
In other words, SAP is building infrastructure for AI agents to do more than retrieve information. It wants them participating directly in business operations.
Why it matters for manufacturing:
Here's where things get interesting.
Imagine an AI agent inside your ERP that identifies a material shortage, checks inventory, recommends a purchasing adjustment, and initiates a workflow to resolve the problem.
Sounds useful, right?
But what happens when the ERP doesn't have an accurate picture of what's happening on the plant floor?
Your MES may know what was actually produced. Your SCADA system knows what the equipment is doing. Your ERP knows what was planned and what transactions have been recorded.
Those aren't necessarily the same thing.
AI agents can only make reliable decisions when the information they're working with accurately represents the operation.
That's why data architecture, system ownership, and integration matter so much.
Here's how this could work on the factory floor:
A production line experiences an unexpected equipment failure. The MES records the downtime, the maintenance system creates a work order, and the ERP still expects the original production schedule to be completed.
An AI agent could potentially identify the conflicting information, assess the impact on material availability and customer orders, and recommend changes to the schedule.
But doing that correctly requires governed access to the relevant systems, accurate operational context, and clear rules about which system owns each piece of information.
A Unified Namespace can help distribute operational context, but it doesn't automatically resolve data ownership or grant an AI agent permission to make changes.
The bottom line:
SAP's announcement reinforces a critical reality: The next phase of enterprise AI isn't just about better models. It's about connecting those models to trusted data and controlled business processes.
If your systems can't agree on what's happening in production, adding autonomous AI agents won't magically fix the problem.
Read SAP's TechEd announcement
Sponsor Message
What Are We Missing Between IT and OT Monitoring?
Sponsored content with Paessler
Manufacturers spend considerable time connecting machines, applications, and data platforms. The infrastructure keeping those connections alive deserves the same attention.
When production data stops updating, the problem could be a controller, an unavailable OPC UA server, a failing switch, or a server running out of resources. Process visibility and infrastructure visibility help answer different parts of the same troubleshooting question.
Paessler PRTG brings monitoring across IT, OT, and IIoT infrastructure into a common environment. With support for OPC UA, Modbus, and MQTT, it helps teams monitor industrial devices and supporting network infrastructure, view their status in dashboards, and receive notifications when monitored conditions require attention.
For teams building a Unified Namespace, infrastructure health is useful operational context. Paessler’s PRTG OPC UA Server can expose monitoring data to industrial systems, providing a path for bringing that context into the broader architecture. PRTG complements SCADA, MES, and the UNS; it does not replace their roles.
On Thursday, November 5, Walker Reynolds is scheduled to join Eric Eckman, OT Channel Manager, US, and David Montoya, Head of Operational Technology, for a live conversation about where PRTG fits in modern industrial architectures and how manufacturers use infrastructure monitoring on the plant floor.
Explore Paessler’s industrial monitoring approach and talk to an expert
Wall Street Is Asking AI the Same Question Manufacturers Should: Where's the ROI?
For the past few years, the technology industry has been spending billions building the infrastructure needed to power artificial intelligence. More data centers, more GPUs, more computing capacity. The assumption? Build it, and the returns will follow.
Well, Wall Street is starting to lose patience.
On October 30, technology stocks took a beating after disappointing financial results from Amazon and Intel raised concerns about the enormous cost of AI infrastructure and how long it might take those investments to pay off.
The details:
Friday's market selloff was significant:
-
Nasdaq Composite: Dropped 2.9%, its largest single-day percentage decline since June.
-
Intel: Shares plunged 28.6% following disappointing earnings guidance, cash flow concerns, manufacturing challenges, and heavy capital spending.
-
Amazon: Reported significant cash outflows while continuing to invest aggressively in AI infrastructure.
-
Nvidia and AMD: Shares fell 4.2% and 10.9%, respectively, as concerns spread across the semiconductor industry.
The issue isn't necessarily that AI doesn't work. It's that the cost of building and operating AI infrastructure is enormous, and investors are increasingly questioning when that spending will translate into sustainable financial returns.
There's a big difference between demonstrating technological capability and proving economic value.
Why it matters for manufacturing:
Sound familiar?
Manufacturers have been dealing with this same problem for years. We've seen expensive digital transformation initiatives, massive software deployments, and ambitious AI projects launched with promises of greater efficiency and productivity.
But what happens when someone asks how much money the project actually saved?
Suddenly, the conversation gets uncomfortable.
Whether you're investing in AI, a new MES, predictive maintenance, or an enterprise data platform, the business case should come before the technology purchase.
Here's how this works on the factory floor:
Imagine a manufacturer considering a $250,000 AI-powered predictive maintenance initiative.
The vendor demonstrates impressive models, real-time dashboards, and automated failure predictions. Everyone's excited.
But before signing the contract, the manufacturer should be asking:
-
How much does unplanned downtime actually cost us?
-
Which equipment failures are we trying to prevent?
-
How much downtime can this system realistically eliminate?
-
What will integration, infrastructure, maintenance, and ongoing operation cost?
-
How will we measure whether the investment worked?
If the project costs $250,000 and saves $50,000 annually after ongoing expenses, that's a five-year simple payback period. If it saves $250,000 annually, the economics look very different.
The technology didn't change. The business case did.
And that $250,000 example is hypothetical. Your actual numbers should come from your production data, maintenance history, and operating costs.
The bottom line:
Wall Street is learning something manufacturers should already understand: Technology spending isn't the same thing as business value.
AI can be transformative, but transformation without measurable results is just an expensive experiment.
From the Floor: A Week in the Community
While the tech industry was busy making headlines about AI, our Industry 4.0 Community was digging into the practical challenges of connecting systems, governing data, and getting new technologies working on the plant floor.
This week, 8,425 community members had plenty of technical ground to cover, with conversations ranging from Unified Namespace architecture and ERP integrations to new industrial software releases. Friday, October 30, was the busiest day, with 29 messages exchanged.
Building a UNS Is One Thing. Governing It Is Another.
Over in #unified-namespace, members explored what an end-to-end governed Unified Namespace actually looks like. The discussion included MQTT topic design best practices and shared resources for structuring industrial data.
It's an important distinction. Getting data into an MQTT broker is relatively straightforward. Building an architecture where that data is organized, contextualized, and governed across an enterprise? That's a different challenge.
When the Plant Floor Meets the ERP
In #💬-general, conversations touched on Tulip-to-SAP integration and electronic Device History Records (eDHR), highlighting the challenges of connecting operational workflows with enterprise systems.
Meanwhile, a question in #enterprise-resource-planning raised an interesting issue: Is charging by API transaction becoming standard practice for cloud-based ERP platforms?
That's worth thinking about if you're designing integrations that could generate thousands or even millions of API calls. An architecture that works technically still needs to make financial sense.
New Tools Worth Checking Out
A few industrial technology updates caught the community's attention:
-
Proxus 2.0: A redesigned interface, Siemens Industrial Edge support, and a built-in AI assistant for plant data.
-
i3X Explorer v1.1.0-0: A rewritten version of the industrial data exploration tool.
-
PracticeOS: More than 200 Model Context Protocol (MCP) tools added for automating workflows across marketing and technology platforms.
Members also shared comparisons of Modbus TCP, OPC UA, MQTT, and Sparkplug B in #🔗-content-links, along with resources on OPC UA companion specifications in #opc-ua.
AI, Second Brains, and Smarter Workflows
AI conversations weren't limited to manufacturing systems.
In #💬-general, members discussed building a personal knowledge management system using Obsidian and Andrej Karpathy's LLM Wiki approach.
The conversation explored how AI can help organize, retrieve, and work with information instead of simply generating answers.
And in #🔗-content-links, members shared examples of using Claude for video editing and content production.
A Couple of Questions Still on the Floor
The community also has a few technical questions looking for additional perspectives.
One member is looking for recommendations on hardware and equipment for water pump monitoring and control. Another is exploring API transaction pricing in cloud-based ERP systems.
If you've implemented either of these in a real industrial environment, your experience could help someone else avoid an expensive mistake.
We also welcomed new members, including doki3 (John), who works on IT/OT data solutions in pharmaceutical manufacturing, and Roslyn_sourcix, a strategic advisor joining from Scotland.
Whether you're architecting a UNS, integrating an ERP, experimenting with AI, or just trying to get a stubborn piece of equipment connected, there's probably someone in the community working through something similar.
Not part of the conversation yet? Join the Industry 4.0 Community Discord and get involved.
Byte-Sized Brilliance
AI Got 280 Times Cheaper. So Why Is Everyone Spending More?
Here's a number worth thinking about: 280.
According to Stanford University's 2025 AI Index Report, the cost of running an AI model capable of GPT-3.5-level performance dropped more than 280-fold between November 2022 and October 2024. What once cost $20 per million tokens fell to just seven cents for comparable benchmark performance.
That's a staggering reduction. But here's the irony: While the cost of using AI has plummeted, technology companies are pouring billions into building the infrastructure to support increasingly powerful models.
For manufacturers, there's a lesson buried in those numbers. Just because the technology is getting cheaper doesn't mean the implementation is getting simpler. Integrating AI with production systems still requires reliable data, infrastructure, security, and people who understand the process.
The real question isn't how cheaply you can run an AI model. It's how much value you can create with it.
Source: Stanford University — 2025 AI Index Report
Responses