I’ve been using large language models to design a few circuit boards recently, and the more I use them, the more I feel that something important is starting to change.
Until recently, when we talked about AI-assisted hardware design, we mostly meant asking AI to search for information, recommend components, explain a datasheet, or write some firmware. But once it came time to draw the schematic and lay out the PCB, an engineer still had to open KiCad, Altium, or another EDA tool and do the work manually.
That boundary is beginning to move.
Large models can now understand KiCad project files, modify schematics, place components, route PCBs, run DRC, and sometimes correct their own mistakes after checking the result.
They are nowhere near reliable enough to design every complex product on their own. But the direction is becoming clear.
And I don’t think this will only affect the EDA industry.
It could reshape a much larger part of the electronics business, from design tools and component databases to manufacturing, test, supply chain, and even the way hardware products are sold.
The first thing to lose value may be the skill of operating an EDA tool
Building an electronic product has traditionally had a high barrier to entry.
You need to understand analog and digital electronics. You need to choose components, read datasheets, draw schematics, and then spend a long time learning how to use Altium, Cadence, KiCad, or some other EDA package.
A lot of people with good product ideas never get past that point.
Something similar has already happened in software.
A few years ago, if you couldn’t program, you simply couldn’t build software. Today a product manager, designer, student, or hobbyist can use tools such as Claude or Codex to create a surprisingly capable application.
Hardware may be heading in the same direction.
Imagine a mechanical engineer, industrial designer, or maker saying:
I want to build a small robot that uses a camera to recognize and sort objects. It should use an ESP32, drive a few motors, include a microphone, and run from USB-C.
The model could choose the major components, find reference designs, build the schematic, lay out the PCB, generate the BOM, and prepare the manufacturing files.
At that point, “knowing how to use Altium” is no longer such a high barrier.
That does not mean EDA tools disappear.
In fact, I think many EDA tools may start to look more like compilers in software development.
Programmers can use AI to write code, but that does not make compilers, tests, or debuggers less important. If anything, the more code gets generated, the more important verification becomes.
PCB design is similar.
A model can generate a board, but someone—or something—still needs to determine whether the design is electrically correct, whether the power system is stable, whether the signal integrity is acceptable, whether it will overheat, and whether it can pass EMC.
Those questions depend on the physics engines, simulation tools, design rules, and verification systems behind the EDA interface.
So I expect engineers to spend less time manually clicking through complicated menus, while AI agents increasingly call these tools in the background.
The value of the UI may decline.
The value of the solver, router, constraint engine, SI/PI analysis, thermal simulation, and physical verification may increase.
The large companies are already moving
It is interesting to look at what the major engineering software companies have been buying.
Synopsys spent roughly $35 billion to acquire Ansys.
Siemens spent around $10 billion to acquire Altair.
Cadence has continued expanding its simulation portfolio and is introducing agent-based systems for PCB and advanced-package design.
Renesas is even more interesting.
A semiconductor company best known for MCUs, analog, power, and connectivity products spent A$9.1 billion acquiring Altium.
These companies clearly do not believe EDA is becoming irrelevant.
They are expanding EDA in two directions at the same time:
toward AI, and toward the physical world.
Traditional EDA mainly answers:
How do we create this electronic design?
The next generation of engineering platforms may need to answer:
How do we turn an idea into a verified product that can actually be manufactured?
That is a much bigger problem.
And it means the boundaries between EDA, simulation, PLM, component data, manufacturing, and supply chain will become less clear.
Component databases may become much more valuable
Today, an engineer searching for a component usually visits DigiKey, Mouser, LCSC, Arrow, or another distributor and manually filters the results.
In the future, the engineer may not do the search at all.
An AI agent could simply ask for:
- 5V to 12V input
- 3.3V / 2A output
- efficiency above 90%
- cost below $1
- production lifetime of at least five years
- at least tens of thousands of units currently available
Then it could select the part itself.
When that happens, the appearance of a component supplier’s website becomes less important than the quality of the data behind it.
The agent needs accurate information about:
- electrical parameters
- pin definitions
- package dimensions
- symbols and footprints
- 3D models
- SPICE and IBIS models
- reference designs
- alternates
- lifecycle status
- inventory
- pricing
- application limitations
In the past, this information was mainly prepared for engineers to read.
In the future, much of it will also need to be prepared for machines to understand.
DigiKey, Mouser, Arrow, Avnet, LCSC, and other distributors may eventually compete not only for engineers, but for the agents working on behalf of those engineers.
The supplier with the best structured data, the most useful API, and the most accurate real-time inventory may have a better chance of being selected automatically.
The same applies to semiconductor companies.
TI, ADI, ST, Renesas, Infineon, and others may increasingly have to think about a new question:
How do we make it easier for an AI system to design our devices into a product?
That is one reason I find the Renesas acquisition of Altium so interesting.
It suggests a semiconductor company can move from simply selling chips to helping customers design complete systems around those chips.
PLM may become more important, not less
PLM has traditionally been something that becomes important after the design has already been created.
It manages files, BOMs, versions, changes, approvals, and product history.
But once AI agents start doing more engineering work, PLM may become even more important.
The biggest problem for an engineering agent is often not drawing the schematic.
It is understanding the context of the project.
Why did the previous revision use this component?
Why was it replaced?
What did the customer ask for?
Which component caused problems during the last prototype?
What changed between revision 17 and revision 18?
Which parts have already been approved by the company?
What waveform should we expect at TP7?
Most of that information is not inside a schematic file.
It is scattered across PLM systems, ERP systems, test reports, purchasing records, emails, and the memories of engineers.
That is why I increasingly think of EDA as the tool used by the agent, while PLM may become the long-term memory of the agent.
We have been running into exactly this problem while adding online KiCad viewing, PCB inspection, iBOM, BOM processing, and various agents to ezPLM.
It does not make much sense to ask a large model to reread an entire KiCad project every time someone asks a question.
A better approach is to convert the schematic, PCB, components, nets, BOM, test points, and other design information into a common hardware design graph.
What the AI really needs to know is something like:
U7 pin 3 connects to R23, which connects to the ADC input. That net is routed for 18 mm on the PCB with no vias, and a high-speed clock passes 3 mm away.
It does not need thousands of UUIDs, font settings, drawing coordinates, and other details that exist only to operate the CAD tool.
I think this kind of structured hardware data may become an important part of the infrastructure for AI-assisted electronics.
PCB manufacturers may be some of the biggest winners
A common reaction is:
If AI can design PCBs using KiCad, what happens to companies that built an ecosystem around their own EDA tools?
I actually think manufacturing platforms such as JLCPCB could benefit from the change.
Their real advantage is no longer only the EDA tool.
They have connected:
EDA, components, PCB fabrication, assembly, CNC machining, and 3D printing.
JLC’s business has grown far beyond being a simple PCB fabrication service.
If AI lowers the barrier to hardware design and increases the number of people capable of creating PCBs by ten times, that could mean more prototypes, more component orders, more SMT assembly, more CNC parts, and more 3D-printed enclosures.
The part that may change is the lock-in created by the design tool.
Today, if someone spends years learning a particular EDA platform, there is a natural tendency to stay inside the surrounding manufacturing ecosystem.
Tomorrow, an AI agent may generate a KiCad project and automatically ask several suppliers:
How much will this board cost?
When can you deliver it?
Can you source every component?
What is the expected yield?
At that point, the real competitive advantages of a manufacturer become:
price, lead time, quality, component availability, DFM capability, logistics, and APIs.
The manufacturer that is easiest for an agent to work with may have a better chance of winning the order.
The faster AI generates designs, the more valuable testing becomes
This is one of the changes I think is still underestimated.
In the past, an engineer might spend two weeks creating one PCB revision.
In the future, AI could generate ten—or even one hundred—candidate designs in the same period.
Then the main question becomes:
Which one actually works?
As design generation becomes cheaper, verification becomes more valuable.
This is one reason simulation companies have become attractive acquisition targets, and why the major EDA vendors continue adding more physics to their platforms.
The same thing will happen in the real world.
Once the PCB arrives, somebody still needs to answer questions such as:
Is the power ripple acceptable?
Why does USB enumeration occasionally fail?
Why is there noise in the camera image?
Does the measured waveform match the simulation?
Eventually, AI will need to connect the design in the computer with the physical board sitting on the bench.
That means oscilloscopes, logic analyzers, multimeters, cameras, and automated test equipment may become the eyes and hands of the engineering agent.
The AI designs a PCB, sends it to manufacturing, creates a test plan, measures the finished board, compares the result with the expected behavior, finds the problem, and then modifies the next revision.
That closes the loop:
Design → Manufacture → Measure → Analyze → Modify
When that loop becomes practical, hardware development changes much more fundamentally than it does when AI simply learns to draw a schematic.
Modules probably will not disappear
If more people without deep electronics backgrounds begin building hardware, what will they use first?
Probably modules.
ESP32 modules, camera modules, motor drivers, power modules, and sensor modules are much easier for an agent to combine than designing everything from transistors, resistors, and capacitors.
So I actually expect the module market to grow in the short term.
But the definition of a good module may change.
Today, a module often comes with a PDF datasheet and a few examples.
In the future, a really useful module may need to provide:
- schematic symbol
- footprint
- 3D model
- reference schematic
- firmware
- driver
- simulation model
- test procedure
- operating limits
- known-good measurements
Ideally, it should also describe its behavior in a form an agent can understand.
At that point, a module begins to look less like a traditional breakout board and more like an npm package in software.
The same may happen to reusable circuits.
A proven USB-C PD power stage, a 10 MHz analog front end, or a BLDC motor driver could become a kind of hardware IP block that an agent can search for, instantiate, parameterize, and verify.
That could create an entirely new market for reusable, verified hardware IP.
Eventually, the bottleneck may become finding customers
When hardware gets easier to build, the difficult question may shift from:
Can I make this?
to:
Who will buy it?
The software industry is already there.
Using AI to build another app is becoming easy. Getting people to use it is still hard.
Hardware may follow the same path.
We may see many more small teams, makers, and even individuals creating their own products.
That could make platforms such as Tindie, Crowd Supply, and other specialized marketplaces more valuable.
If the number of new hardware products grows dramatically, scarcity moves further down the chain.
Manufacturing capacity matters.
Quality matters.
Trust matters.
Reviews matter.
Community matters.
Distribution matters.
Getting discovered matters.
That is why I do not think the most valuable company in the future electronics industry will necessarily be the company with the best PCB editor.
The bigger opportunity may belong to whoever can connect the entire path:
Idea → Design → Components → Simulation → Manufacturing → Test → Sell
Large models can reduce the barrier at the front of that chain.
But somebody still has to turn the idea into a reliable physical product and then get that product into the hands of a customer.
So I do not think AI will make the electronics industry smaller.
Quite the opposite.
The software industry became enormous not because software became harder to build, but because it became easier.
If millions more people become capable of turning their ideas into electronic products, the markets for PCBs, components, manufacturing, testing, modules, reusable IP, and sales platforms may all become much larger.
What loses value is one particular skill:
knowing how to operate a specific tool.
What gains value is very different:
verified design knowledge, trustworthy data, physical simulation, manufacturing capability, test data, and access to customers.
If this transformation really happens, the largest electronics companies of the future may not fit neatly into categories such as “EDA company,” “PCB manufacturer,” or “component distributor.”
They may look more like a combination of:
GitHub + npm + AWS + Shopify — but for hardware.
Whoever can connect an idea to a real product that people can actually buy may have the biggest opportunity of all.
source https://blog.tindie.com/2026/09/when-large-models-start-designing-pcbs-what-happens-to-the-electronics-industry/
No comments:
Post a Comment