# AI Hardware Engineering: Robots, Satellites, Vehicles, and the End of Legacy CAD AI hardware engineering is moving beyond CAD assistance. Matthias Wagner explains why AI will design robots, satellites, and vehicles by the end of 2027. Written by [Matthias Wagner](https://flux.seobdtools.com/natarius) Published Nov 2026 - 11 min read ![AI Hardware Engineering: Robots, Satellites, Vehicles, and the End of Legacy CAD](https://flux.seobdtools.com/p/assets/blogs/ai-hardware-engineering-robots-satellites-vehicles/main.webp) AI hardware engineering is moving from answering questions to designing complete, manufacturable devices. I believe that by the end of 2027, AI will do the overwhelming majority of the design and engineering work for robots, satellites, vehicles, and industrial machines. This post is based on [my X article](https://x.com/MatthiasWagner/status/2120247606983225469). 2026 will be remembered as the year AI learned to engineer the physical world. Not just write code. Not just answer engineering questions. Actually design things that can be manufactured. And I believe we're much closer to AI taking on the world's most ambitious hardware engineering projects than almost anyone realizes. ## This is the future we started building in 2019 When we founded Flux in 2019, our vision wasn't to build another PCB design tool. We believed AI would eventually do the engineering itself. Electronics. Mechanical engineering. Industrial design. Firmware. Manufacturing. Hardware bring-up. Testing. Validation. Not as disconnected disciplines, each requiring a different application and specialist, but as parts of one integrated engineering process. At the time, the AI wasn't capable of doing much of this. We had to build toward a future that didn't exist yet. Then the models started catching up. ## Seven years of building toward this moment In April 2023, we launched Flux Copilot, bringing conversational AI into a hardware engineering environment around the same time Cursor was emerging for software development. By June, AI could modify electronic schematics. By December, it could understand images. In 2024, we introduced AI design reviews and AI-assisted PCB routing. In October 2025, we launched our autonomous AI hardware engineer. Instead of asking AI to help with individual operations, users could assign engineering work and let the agent plan, execute, inspect, and iterate. In April 2026, we introduced AI component placement, bringing autonomous PCB layout much closer to completion. By the middle of this summer, Flux could produce working, manufacturable electronics of moderate complexity from a single assignment. Then, in September, we expanded beyond electronics into mechanical CAD. Flux could design PCBs and their enclosures together, accounting for how the physical and electronic systems fit. Firmware was becoming part of the same integrated workflow. This wasn't simply a sequence of new features. It was the gradual removal of the boundaries between engineering disciplines. And the frontier models were improving underneath us the entire time. ## How are AI models getting better at hardware engineering? Early in 2026, GPT-5.4 demonstrated meaningful improvements in spatial reasoning. By summer, newer GPT models were becoming substantially more useful for mechanical CAD. September's Astra release represented another leap. Opus 5.5 pushed the frontier further. And now Fable 5.2 has arrived. The reported improvement on CAD-related benchmarks is extraordinary. More importantly, the qualitative change is becoming impossible to ignore: models are increasingly able to reason about geometry, manipulate engineering tools, recover from mistakes, and work through problems that previously required substantial human intervention. ![BenchCAD 2.1 results by reasoning effort and inference cost: OpenAI internal reaches 87.5, Claude Fable 5.2 reaches 83.5, and Claude Opus 5.5 reaches 65.3 out of 100. Full model results and cost comparison are shown in the table.](https://flux.seobdtools.com/p/assets/blogs/ai-hardware-engineering-robots-satellites-vehicles/benchcad-model-performance-dark.webp) BenchCAD 2.1 model performance with engineering tools, across reasoning effort levels and inference cost. Chart from my X article. The pattern should feel familiar to anyone who watched AI coding evolve. For years, language models could generate snippets of code. That was useful, but it wasn't autonomous software engineering. The breakthrough came when increasingly capable models were combined with coding-specific training, tools, execution environments, and feedback loops. Products like Claude Code made that shift tangible. Hardware engineering is now going through a similar transformation. ![Parametric CAD Bench V3 measured scores rise from 17.2% in May 2024 to 61.0% in September 2026. Dashed projections reach 75.9% in August 2027 and 92.8% in August 2028. The chart separates measured results from projections.](https://flux.seobdtools.com/p/assets/blogs/ai-hardware-engineering-robots-satellites-vehicles/parametric-cad-performance-dark.webp) Measured Parametric CAD Bench V3 results and a trend projection. The source and projection assumptions are included in the chart. Except the problem is considerably harder. Software can often be executed, tested, and verified automatically. Hardware must obey the laws of physics, satisfy manufacturing constraints, and work in the real world. A convincing-looking 3D model is not necessarily manufacturable. A plausible schematic is not necessarily electrically correct. A PCB that passes basic design rules is not necessarily a working product. That's why general-purpose intelligence alone isn't enough. ## Why does the engineering environment matter as much as the model? At Flux, we've spent years building the environment AI needs to do real hardware engineering. Instead of asking a language model to generate scripts that might eventually produce a useful result, we give it direct access to specialized engineering tools. CAD and EDA kernels. Simulation. Design-rule checking. Analysis. Component information. Manufacturing workflows. The agent can create something, inspect the result, identify problems, make changes, and verify its work. And we're starting to see something even more remarkable: Increasingly, boards designed by Flux are coming back from manufacturing and working, even when experienced human engineers initially questioned whether they would. Some are also cheaper to manufacture than the designs engineers would have produced themselves. This points to something fundamental about AI engineering. Much like a compiler can optimize code in ways the average programmer wouldn't think to do manually, AI can explore design choices and optimizations that human engineers might overlook. It doesn't have to follow the same conventions, habits, or intuitions that shape human engineering decisions. It can explore unconventional layouts, component arrangements, and manufacturing tradeoffs, then use engineering tools to check whether those choices actually work. The result isn't just faster engineering...it's the possibility of better engineering! It can reason across disciplines rather than treating electronics, mechanical engineering, and firmware as unrelated files. This matters because improvements in model intelligence and improvements in the engineering environment compound. A better model can accomplish more with the same tools. Better tools allow the same model to accomplish more. And improvements to either side make the entire system more capable. We're also seeing the economics change rapidly. At Flux, the model cost of performing 1,000 design changes fell from approximately $1,100 in late August to $330 by late October 2026. That's a 70% reduction in roughly two months! We're not just getting better engineering capability. We're getting dramatically more of it per dollar. ## How fast is AI hardware engineering improving? The most important question isn't what AI can engineer today. It's how quickly that capability is improving. METR measures the complexity of tasks AI agents can complete autonomously, expressed as the time a human expert would need to complete the same work. Its research suggests that, for frontier models released since 2024, this task horizon has been doubling roughly every three months. ![METR AI agent task horizons at 50% success: historical measurements increase from 1.4 minutes to 17.4 hours. A dashed trend projection shows approximately 8 days in 2027 and 90 days in 2028, distinct from measured results.](https://flux.seobdtools.com/p/assets/blogs/ai-hardware-engineering-robots-satellites-vehicles/ai-autonomy-growth-dark.webp) METR AI agent task horizons and a trend projection. The source and projection assumptions are included in the chart. Think about what that means. We're moving from tasks measured in minutes toward tasks measured in hours and potentially days. If the trend continues, the scale of work AI can tackle could grow by orders of magnitude over the next two years. That distinction matters. And those same frontier models are now operating inside increasingly sophisticated hardware engineering environments. Which brings me to a prediction: ## Within a year, AI will be capable of designing even the most ambitious hardware projects Advanced robotics. Satellites. Vehicles. Industrial machines. Complex products that span mechanical engineering, electronics, firmware, and manufacturing. I believe that by the end of 2027, AI will be capable of performing the overwhelming majority of the design and engineering work involved in projects like these. That doesn't mean humans disappear from the process. People will still define objectives, make tradeoffs, evaluate results, build prototypes, conduct physical tests, and take responsibility for safety and reliability. And AI-generated designs will still need rigorous verification before anyone should trust them in the physical world. But the role of the human engineer is going to change dramatically. Instead of manually operating a collection of specialized applications to produce a design, engineers will increasingly direct, supervise, and validate AI systems that do the work. The bottleneck will shift from how quickly humans can execute engineering tasks to how effectively we can specify goals, evaluate outcomes, and test physical reality. That is an enormous change! And it has profound implications for the software industry that supports engineering today. ## Why is traditional CAD software about to be disrupted? For decades, engineering software has been built around the assumption that a human operates the application. You learn the interface. You select commands. You manipulate geometry. You draw schematics. You route traces. You move between applications, export files, import them elsewhere, and manually reconcile the differences. The entire industry is organized around helping humans perform these operations. But what happens when the primary operator is an AI agent? We've already seen a version of this shift in software development. AI coding agents haven't made compilers, programming languages, or execution environments irrelevant. They've changed the interface through which software gets built. I expect the same thing to happen in hardware! The underlying CAD kernels, simulation engines, manufacturing knowledge, and verification systems remain enormously valuable. In many cases, they'll become even more valuable. Decades of engineering expertise are embedded in those systems. A mature CAD kernel knows how to construct geometry that behaves correctly. A sophisticated EDA tool understands electrical and manufacturing constraints that a language model cannot simply wish away. But the traditional GUI-centric applications built around those capabilities face a very different future. When AI can operate the engineering stack directly, why should every engineer need to spend years mastering a dozen different interfaces? Why should the primary unit of value be a software seat rather than an engineering outcome? Why should electronics, mechanical design, firmware, and manufacturing remain fragmented into separate workflows? I don't think they will. The value will increasingly shift from applications that help humans do engineering to systems that actually deliver engineering work. ## Hardware engineering is becoming a fundamentally different industry The implications extend far beyond CAD vendors. Today, building hardware is slow, expensive, and organizationally complicated. A product might require electrical engineers, mechanical engineers, firmware developers, industrial designers, manufacturing specialists, and project managers, all coordinating across different tools and representations of the same physical product. Every handoff introduces delay. Every iteration costs time and money. And every discipline has to reconcile its decisions with the others. AI changes that equation! Imagine being able to describe a product, have an AI engineer develop the electronics, mechanical structure, and firmware together, evaluate the tradeoffs, prepare manufacturing documentation, and iterate continuously as requirements change. Not as a one-off demo. As the normal way products get built! A small team could take on projects that previously required an entire engineering organization. Established companies could explore far more design alternatives, move faster, and bring products to market with dramatically less overhead. Entire categories of products that are uneconomical to develop today could become viable. This is the opportunity we've been working toward at Flux for seven years! ## We’re still at the beginning It's tempting to look at today's AI hardware engineering systems and focus on what they still can't do. The relevant comparison isn't between today's AI and a perfect human engineering organization. It's between what AI could do six months ago, what it can do today, and what happens if the underlying trajectory continues. We've spent seven years building for a world in which AI does the engineering. For most of that time, we were waiting for the models to become capable enough. Now they're arriving faster than almost anyone expected. The question is no longer whether AI will transform hardware engineering...it's how quickly the rest of the industry will adapt. The future of hardware engineering isn't better tools for engineers. It's AI engineers that build everything. And that future is arriving now 🚀 Read more: [Why AI hardware engineering needs more than computer use](https://flux.seobdtools.com/p/blog/ai-computer-use-altium-pcb-routing) and [how to adopt Flux as a team](https://flux.seobdtools.com/p/blog/how-to-adopt-flux-as-a-team). [Start Free Trial](https://flux.seobdtools.com/signup) ## FAQs about AI hardware engineering ### What is AI hardware engineering? AI hardware engineering uses AI to do engineering work across electronics, mechanical design, firmware, and manufacturing as one integrated process. The AI creates designs, inspects results, makes changes, and verifies its work with engineering tools. ### When will AI design robots, satellites, and vehicles? My prediction is that by the end of 2027, AI will be capable of performing the overwhelming majority of the design and engineering work for these projects. Engineers will define objectives, make tradeoffs, evaluate results, and test physical prototypes. ### Why do engineering tools matter as much as the AI model? CAD and EDA kernels, simulation, design-rule checking, component information, and manufacturing workflows let AI inspect and verify what it creates. Better models and better tools compound: each makes the whole system more capable. ### How quickly are AI capabilities and costs changing? METR research suggests that the task horizon for frontier models released since 2024 has been doubling roughly every three months. At Flux, the model cost of 1,000 design changes fell from approximately $1,100 in late August to $330 in late October 2026, a 70% reduction. ### Will AI replace traditional CAD software? I expect value to shift from GUI-centric applications that help humans do engineering to systems that deliver engineering work. CAD kernels, simulation engines, manufacturing knowledge, and verification systems remain valuable parts of that future. ## About the author [Matthias Wagner](https://flux.seobdtools.com/natarius) Founder & CEO at Flux Matthias is the Founder and CEO of Flux, a hardware design platform that's revolutionizing how teams create and iterate on circuits. Find him on Flux [@natarius](https://flux.seobdtools.com/natarius). Let’s start building Build your big idea together with Flux. All you need to do is describe what you want to make, or what problem you want to solve. Flux will work with you to make it a reality. ![Close-up photo of a PCB made with Flux, with an ESP32 module, buttons, and a microSD slot.](https://flux.seobdtools.com/p/assets/blogs/miscellaneous/sidebar-cta-pcb.webp) [Try it Free](https://flux.seobdtools.com/signup) ## Related Content [ ![Isometric teal line drawing of a sensor interface board with a tall vented housing, cable ports, and a mounting plate.](https://flux.seobdtools.com/p/assets/blogs/flux-fable-5-2-preview/v2/main.webp) ### Flux + Fable 5.2 (Preview) 3 Min Read](https://flux.seobdtools.com/p/blog/flux-fable-5-2-preview)[ ![Isometric teal line drawing of a two-board test instrument with a wide chassis, mounting rails, and an elevated vented cover.](https://flux.seobdtools.com/p/assets/blogs/flux-gemini-4-argon/v2/main.webp) ### Flux + Gemini 4 Argon 3 Min Read](https://flux.seobdtools.com/p/blog/flux-gemini-4-argon)[ ![Isometric teal line drawing of an industrial inspection camera with a lens, stacked circuit boards, a heat sink, and an exploded protective housing.](https://flux.seobdtools.com/p/assets/blogs/flux-gemini-4-carbon/v2/main.webp) ### Flux + Gemini 4 Carbon (Preview) 4 Min Read](https://flux.seobdtools.com/p/blog/flux-gemini-4-carbon)[ ![Isometric teal line drawing of an industrial controller board with terminal blocks, a mounting base, and an elevated enclosure lid.](https://flux.seobdtools.com/p/assets/blogs/flux-gpt-6-1-sol/v2/main.webp) ### Flux + GPT 6.1 SOL 3 Min Read](https://flux.seobdtools.com/p/blog/flux-gpt-6-1-sol)[ ![Isometric teal line drawing of a data acquisition board above a vented enclosure with four front connectors.](https://flux.seobdtools.com/p/assets/blogs/flux-gpt-6-astra/v2/main.webp) ### Flux + GPT 6 Astra 3 Min Read](https://flux.seobdtools.com/p/blog/flux-gpt-6-astra)[ ![Isometric teal line drawing of an embedded control board with a heat sink above a flanged enclosure with circular connectors.](https://flux.seobdtools.com/p/assets/blogs/flux-opus-5-5/v2/main.webp) ### Flux + Opus 5.5 3 Min Read](https://flux.seobdtools.com/p/blog/flux-opus-5-5) See All (6) ## Let’s start building. Let’s build your idea together. All you need to do is describe what you want to make, or what problem you want to solve. Flux will work with you to make it a reality. [Try it Free](https://flux.seobdtools.com/signup) ![Close-up of a Flux-designed printed circuit board](https://flux.seobdtools.com/p/assets/homepage-v2/mission-visual.webp)