Siemens and NVIDIA’s Erlangen Blueprint: 85% Less Ramp-Up Engineering

In short: On 19 October 2026, Siemens and NVIDIA showed what the industrial AI partnership they announced at CES in January actually does inside a running plant. At Siemens’ electronics factory in Erlangen, Germany, a fully virtual production ramp-up cut engineering effort on new PCB assemblies by up to 85% and time to market by up to 40%, while agent-based production planning is credited with up to 20% lower cost. Three days later the two companies extended the same agentic approach to 3D chip design, where it evaluates up to 50 times more design options. For a maintenance or purchasing team, the useful question is not whether the demonstration is impressive. It is what an AI-driven plant needs from the automation hardware underneath it, and what that does to the spare-parts list you are holding.
What Siemens and NVIDIA actually announced
The 19 October event was not a conventional product launch. Siemens opened its Erlangen electronics plant, where the two companies have been building and testing the first applications of the “Industrial AI Operating System” they announced at CES 2026. Siemens says the first solutions went from development to real production in under ten months, which matters: these are not lab demonstrations dropped into a keynote slide.
The plant makes power electronics and automation products for machine tools, battery production and process and discrete manufacturing. The PCB assemblies at the centre of the announcement are the boards that go inside drives and frequency converters, which is why the workflow being described is inspection preparation rather than, say, final assembly.
| What was shown | Figure | What the figure actually covers | Basis |
|---|---|---|---|
| Fully virtual production ramp-up | Up to 85% less engineering effort | Preparing automated optical inspection for a new PCB assembly by training the vision system on images rendered from the digital twin, instead of waiting for a physical sample | Siemens, 19 Oct 2026 |
| Time to market | Up to 40% shorter | The same new-product introduction workflow for PCB assemblies | Siemens, 19 Oct 2026 |
| Agent-based production planning | Up to 20% cost saving | Agents built on NVIDIA Factory Operations Blueprint, integrated into Siemens Intelligence Center X for root-cause analysis and coordination of plant operations | Siemens and NVIDIA, Oct 2026 |
| Agentic EDA for 3D ICs | 50x more design options; up to 95% less engineering effort | Electrical, power and thermal co-design across seven Siemens EDA tools and nine process stages; tool setup from days to under five minutes | Siemens, 22 Oct 2026 |
| AI factory reference designs | Not quantified | Validated starting points for planning and deploying AI factories, combining NVIDIA’s compute architecture with Siemens electrification and simulation | Siemens, 22 Oct 2026 |
| Simulation throughput | Simcenter up to 10x faster; PhysicsAI up to 1,000x faster prediction | Vendor benchmarks on GPU-accelerated solvers and surrogate models | Siemens, 22 Oct 2026 |
Read the 85% before you quote it
The figure is real, and it is narrower than it sounds. It describes one workflow: getting a new board’s inspection ready. The old sequence was sequential by necessity. Engineers waited for a physical sample, photographed and labelled it, then configured the vision system by hand. That took weeks, and it could not start until the sample existed. The new sequence inverts it. Images are generated from the product digital twin, the inspection model is trained before the first physical board exists, and the sample is then used to confirm the model rather than to create it.
That is a genuine change in the order of operations, and it is exactly the kind of gain that lands in new-product introduction. Nothing in the announcement says the plant runs 85% cheaper, or that a mature line produces 85% more. The same caution applies to the 40%: it is compressed calendar time for a specific class of product launch, not a 40% cut in total manufacturing cost.
“Up to” is carrying weight as well. These are best-case figures for a specific product family at a specific site, published by the vendor, without an independent audit. Erlangen is also an unusually favourable test bed. Siemens owns the product lifecycle management, the manufacturing execution layer, the automation and the simulation tools, so no integration boundary has to be negotiated. The plant has been instrumented for more than a decade and already runs more than 100 AI algorithms across production, logistics, inspection and maintenance. Most readers are starting from a plant where none of that is true.
The 20% is softer still. It is a projected saving from agents that identify root causes and coordinate operations, not a booked result across a plant network. Treat it as a direction of travel with an indicative magnitude, not a number to put in a capital request without your own baseline.
Why Erlangen is both the right showcase and the wrong template
The plant is a fair place to prove this particular idea. It has been running since 1971, employs around 1,000 people, and produces a high mix of power electronics and automation products. High-mix electronics is where virtual ramp-up pays best, because every new board forces a new inspection setup, and every inspection setup is a queue in front of the physical sample. Remove that queue and the whole introduction sequence moves left.
Now the uncomfortable part. A continuous-process plant, such as cement, pulp or refining, has almost no new-product churn. The 85% barely applies there; the value sits in energy, yield and throughput instead. A job shop with thousands of customer-specific variants has the churn but rarely the digital twin, the CAD discipline or the metrology to render trustworthy synthetic images. And a plant that already struggles to keep its historian populated will not get an inspection model out of a twin that does not match the line.
There is also a showcase effect to keep in view. Siemens says the site receives close to 3,000 customer visits a year. A plant that is visited that often is, by design, presentable. That does not make the numbers false. It does mean the line you are looking at has been curated, and that the effort of making an ordinary line presentable is not included in the 85%.
What is actually running, and what it reads from your plant
Strip away the branding and the stack has a consistent shape. Each layer consumes data that originates in the automation and instrumentation layer, which is the part a maintenance team already owns.
| Layer | What Siemens and NVIDIA put there | What it needs from the plant |
|---|---|---|
| Virtual ramp-up | Process Preparation X and Teamcenter feeding NVIDIA Cosmos world models that generate training images | An accurate product twin, CAD and BOM data, camera and lighting specifications |
| Production planning | Agents on NVIDIA Factory Operations Blueprint inside Siemens Intelligence Center X | MES orders, machine states, cycle times, quality events, maintenance history |
| Engineering simulation | Simcenter on CUDA-X libraries, PhysicsAI surrogate models | Validated CAE models and as-built geometry |
| Chip design | Agentic EDA orchestrating seven tools across nine stages | Process design kits, thermal boundary conditions, package models |
| AI factory planning | Reference designs with Siemens electrification and digital twin | Electrical load, cooling capacity, floor space, grid constraints |
| Energy | Plant digital twin used to test demand and grid interaction before physical change | Metering, load profiles, tariff structure |
The pattern is consistent: the AI does not replace the control layer, it interrogates it. Every agent above needs tags, timestamps and context from PLCs, I/O modules, drives, cameras and meters that already exist. What changes is the economic consequence of that layer being wrong. A missing signal used to be an annoyance on a trend page. In an agentic system it becomes a wrong root cause, a bad plan, or a model that quietly learns the wrong thing.
Can a brownfield plant copy any of this?
Yes, in stages, and the stages are unglamorous. Five prerequisites decide whether a retrofit produces a working system or an expensive dashboard.
- Tag-level data access, not screen scraping. If the only way to read a machine is a screenshot of an HMI, the data is already too late and too lossy to train anything on. OPC UA at the controller or a gateway is the floor, not the ceiling.
- Time-aligned history. Agents reason across events. If the drive fault, the quality reject and the operator intervention are stamped by three different clocks, correlation becomes guesswork.
- An asset register that matches the line. Nameplate data, firmware revision, installed options and the spare-parts bill of materials need to be true before a model is trained on them. This is usually the largest hidden cost of the project.
- Segmentation between IT and OT. An agent that reads plant data is a new attack surface with a legitimate reason to reach into the control network. Network architecture has to be decided before the first connector, not after the pilot.
- A process stable enough to be worth optimising. AI compresses preparation and coordination work. It does not fix a line with an unstable takt time, chronic minor stops or an unresolved mechanical fault. Automating a chaotic process produces faster chaos.
The plants that fail usually fail at step three. They buy sensing and software, then discover that the asset register says one drive variant and the cabinet contains another, that three firmware levels coexist on the same machine type, and that nobody can say which spare was fitted during the last repair. The model inherits all of it.
What this does to spare-parts strategy
The announcement is a software story, but it lands on hardware budgets. Five effects are worth planning for.
- Downtime becomes more expensive per minute. An agent-coordinated, highly automated line carries fewer people watching it. When it stops, the loss is throughput, not idle labour, and the recovery often depends on a person who has to be called. That argues for holding the small number of parts that genuinely stop the line on site, rather than trusting a distributor’s lead time.
- Faster introduction means faster churn. If ramp-up effort falls by up to 85% and time to market by up to 40%, the economics of launching more variants, more often, improve. More variants means more changeovers and more setup wear on feeders, grippers, sensors and drive parameters, and a shorter useful life for electronics tied to a specific product generation.
- The retrofit money goes around the controller, not into it. In most brownfield AI projects the spend is on sensors, I/O, edge computers, network hardware and panel modifications around an existing controller. Replacing a working S7-400 or a working drive is rarely the first move. That surrounding layer is what a spare-parts supplier actually ships.
- Obsolescence turns into a data problem. Agents reason over asset records. If your critical-spares list is a spreadsheet last updated during the previous migration, the system inherits the gap and will plan around a part that no longer exists.
- Vendor coupling moves up the stack. The control layer stays comparatively open, through OPC UA and PROFINET. The AI layer is licensed, and it will be renewed on the vendor’s cycle. Keep that boundary explicit in contracts so that a software renewal never becomes the pretext for a hardware refresh.
Which Siemens parts sit in the retrofit path
Autronexa’s Siemens inventory is short, and we would rather say so than pad a table. None of these parts runs an agent. They are the kind of catalogue hardware that ends up around a data or retrofit project: a drive that stays in service while the line around it is instrumented, an I/O card that adds signals to an existing rack, a control unit that replaces a failed one on a machine the plant cannot afford to stop.
| Part number | What it is | Where it fits in a retrofit |
|---|---|---|
| 6SE7016-0TP50-Z | SIMOVERT MASTERDRIVES 6SE70 series inverter | Legacy drive on a line being instrumented; the drive stays, the measurement around it is new |
| 6ES7422-1BL00-0AA0 | SIMATIC S7-400 SM 422 digital output module, 32 channels, 24 V DC | Adding outputs to an existing S7-400 rack instead of replacing the controller |
| 6SL3243-0BB30-1PA3 | SINAMICS G120 control unit | Restoring a G120 drive that the line depends on, without changing the power section |
| 6ES7193-4CD20-0AA0 | SIMATIC ET 200S terminal module | Field-level termination when a retrofit adds points to an existing ET 200S station |
If you are new to the range, the Siemens spare parts buying guide explains how the families are organised and how ordering works, and the Siemens brand page shows what we currently hold. Variants and cross-references are confirmed on request rather than assumed from a photograph.
How to sequence a retrofit without stopping the line
The failure mode in most AI retrofits is scope. A plant tries to instrument everything, discovers in month seven that the data is not trusted, and writes off the programme. A sequence that survives contact with production looks like this.
- Measure before you instrument. Two to four weeks of OEE by station, recorded manually if necessary, so that any later improvement has a baseline that predates the project.
- Pick one constraint, not the plant. A single bottleneck station, a single inspection step, or a single changeover is enough to prove the loop.
- Model the asset, not the brochure. Walk the cabinet, record nameplate and firmware, reconcile the spare-parts bill of materials. This is the step that determines whether the project is credible.
- Add sensing where the constraint is. Usually a handful of I/O points, one or two additional sensors, and a gateway, not a controller replacement.
- Prove that the loop closes. The output of the model has to change a decision that somebody actually makes, on a timescale short enough to observe. If nothing changes, the pilot has not proved anything.
- Freeze the critical-spares list against the new design. A retrofit changes the failure profile of a line. The list that was right before the project is not automatically right after it.
What Siemens and NVIDIA have not disclosed
- How the 85% and 40% were measured. No baseline, sample size or product mix is published, and it is not stated whether the comparison is against the same team’s previous introduction or an industry average.
- Which lines are running it in series production. The announcement describes production use but does not say how many boards have shipped under the new workflow.
- What the stack costs. Licences, GPU capacity, integration effort and training are not quantified. A saving without a cost basis is not yet a business case.
- Whether the 20% is measured or modelled. The wording points to an estimate.
- The labour effect. How inspection preparation and planning roles change is not addressed.
- How it performs on old hardware. There is no published result from a plant running 15 to 20 year old controllers with a thin historian, which describes most of the installed base.
Bottom line
The 19 October announcement is a credible but narrow proof point. AI can compress the preparation work that happens before a line starts, and agents can absorb some of the coordination work once it is running. It is not evidence that factories run themselves, and the headline numbers are vendor figures from a carefully instrumented showcase site.
What is not in doubt is the direction. These systems consume data produced by the automation layer, so the condition of that layer, and the availability of its spare parts, becomes more commercially important rather than less. The plants that benefit first will be the ones that already know what is in their cabinets. The ones that benefit last will be the ones still deciding whether the asset register is worth the effort. If your transition plan includes a controller generation change, the CompactLogix 5390 launch analysis covers the same question from the other side of the market.
Frequently asked questions
What exactly did Siemens and NVIDIA announce on 19 October 2026?
They presented AI applications running in production at Siemens’ electronics factory in Erlangen, Germany: a fully virtual production ramp-up for new PCB assemblies that cuts engineering effort by up to 85% and time to market by up to 40%, plus agent-based production planning credited with up to 20% lower cost. The plant is being positioned as the blueprint for AI-driven autonomous manufacturing that the two companies announced at CES 2026.
Is the 85% a reduction in total factory engineering cost?
No. It refers to the preparation work for automated optical inspection of a new PCB assembly, where the vision system is now trained on images generated from the digital twin instead of waiting for a physical sample. It is a best-case figure for a specific workflow at one site, published by the vendor.
Does the Erlangen blueprint require replacing existing PLCs and drives?
Not on the evidence published. The AI layers described sit above the automation layer and read data from it. The hardware that has to change in a brownfield project is usually the sensing, I/O, network and edge computing around the existing controller, not the controller itself.
What is NVIDIA Factory Operations Blueprint?
It is a blueprint, available in early access on GitHub, for building AI agents that coordinate factory operations. Siemens is integrating agents built on it into its Intelligence Center X system for root-cause analysis and operational coordination, which is where the up to 20% cost saving is claimed.
What does the chip design announcement have to do with the factory?
It is the same partnership, announced three days later on 22 October 2026. Siemens extended its agentic EDA capability to 3D IC electro-thermal co-design, orchestrating seven EDA tools across nine process stages, cutting tool setup from days to under five minutes and evaluating up to 50 times more design options than a manual workflow.
How should a maintenance team respond to this announcement?
Treat it as a prompt to fix the data foundation rather than a reason to buy new hardware. Reconcile the asset register and the critical-spares list, confirm that the line produces time-aligned tag data, and pick one constraint to instrument. The parts most likely to be needed are I/O, control units, drives and field-level termination around equipment that stays in service.
Sources
- Siemens press release, 19 October 2026: Siemens partners with NVIDIA to advance AI-driven, autonomous factories
- Siemens press release, 22 October 2026: Siemens and NVIDIA are advancing Industrial AI
- Siemens press release, CES 2026: industrial AI operating system announcement
- Siemens and NVIDIA partnership overview
- Siemens Process Preparation blog, 21 October 2026: Process Preparation X and Digital Twin Composer at NVIDIA GTC Berlin
Autronexa is an independent supplier of industrial automation spare parts and is not affiliated with Siemens AG or NVIDIA Corporation. Performance figures quoted above are the companies’ own published claims for specific workflows at a specific site; they are not independently verified and should not be read as guarantees for other plants. Part numbers and descriptions are given for identification and cross-reference only.
Need a Siemens part confirmed against your cabinet rather than against a catalogue page? Send the nameplate details with your request or contact us directly, and we will check availability and compatible variants.
