New AWS AI Hardware Company is Teaching AI to Think in Physics

With the boom in AI, all the attention of investors regarding computing hardware has been focused on ultra-advanced chips, GPUs, and memory.

However, to connect these components together in machines and consumer goods, from washing machines to cars to industrial robots, printed circuit boards (PCBs) are needed. They are the flat boards used to physically support and electrically connect electronic components.

Source: Quilter AI

Producing PCBs is not trivial, but it is a well-understood and scaled-up mature supply chain, with the actual manufacturing increasingly fully automated. However, designing new PCBs is all but simple and is actually one of the last manual steps in electronics development.

A specialist is drawing the plan by hand and placing hundreds or thousands of components and routing the copper traces that connect them across multiple layers. For a board of moderate complexity, layout takes four to eight weeks. Sophisticated systems like automotive electronics or computers take three months or longer.

This could be changing very quickly as a new startup called Quilter AI is now using AI to automate this process. It can turn a week- or month-long process into just a few days, saving a lot of work hours for the PCB designers.

Why Is PCB Design So Difficult?

PCBs are very complex 3D physical systems that need to balance multiple constraints for a good design:

• Space/mechanical limits: the components need not just to be arranged together, but also fit into a very tight space.
• Electrical noise: passing current creates potential power leakage, unintended signal transfer, and electromagnetic interference (EMI), requiring the anticipation of unwanted electrical effects.
• Thermal management:  High-power components should not create too many hot spots; a copper heat sink needs to be added to evacuate that heat, and airflows from fans must not be obstructed.
• Manufacturing Limits: constraints on how manufacturing chemicals and tools work mean design must avoid designs like copper traces meeting at acute angles or too thin solder masks, take into account feasible hole diameters, etc.

Solving many of these issues can make another one worse, meaning PCB design is a balancing act to meet all technical objectives while still fitting within the cost, power consumption, performance, and size requirements of a given PCB.

As a result, until now, most automated assistant tools for PCB designs were useful, but needed a heavy dose of human touch to detect eventual problems. In addition, sensitive circuits require specialized, non-standard layouts.

Source: Cadence

Quilter AI Overview

Quilter History

Quilter was founded by Sergiy Nesterenko after five years at SpaceX as a Sr. Radiation Effects Engineer, where he developed electronics for Falcon 9 and Falcon Heavy, and as a researcher in California before that.

“I triple majored in math, physics, and chemistry at Berkeley not because I wanted to specialize, but because I wanted to master the fundamentals. That makes everything else easier to learn.”

The company team is made up of engineers from SpaceX, Apple (AAPL ), NASA, Johns Hopkins APL, and MIT. It also hired Electronic Design Automation (EDA) builders from PCB automation companies Cadence and Synopsys.

The company raised $10M in its Series A in 2023, and another $25M in 2025, a year after the launch of the open Beta of its software. The company launched its Free Tier in August 2025, where the cost of the solution is on a per-use basis, departing from the often costly software subscription model dominating the industry.

“When you upload a board to Quilter, the number of unrouted pins at the time of upload is your design pin count. You pay for what Quilter needs to compile, nothing more.If you pre-route part of the board before uploading (RF nets, fan-outs, high-voltage sections), those routed pins don’t count. Quilter routes around your existing work and charges only for the pins it needs to handle.”

Quilter also offers no proprietary formats and no lock-in, helping the onboarding of PCB designers and integration into existing workflows and manufacturing tools.

Source: Quilter AI

Bringing Physics To AI

At large, AI has been mostly used so far for text and image generation, including computer code. But it often struggles with real-world physics, which is why a new focus of AI development has been physical AI, especially for robotics applications (follow the link for our full report on this topic).

Quilter has adopted an approach mixing reinforcement learning, machine learning, and neural networks to get their AI a better understanding of how real physics creates constraints in PCB design.

Source: Quilter AI

As a result, Quilter’s AI is not designed by imitating existing boards, without “understanding” why they were made this way and learning from humans. This approach, similar to most LLMs’ basic way of functioning, would often result in a serious problem when multiple constraints are in play.

“Quilter generates complete layouts using AI trained on real-world physics and manufacturing constraints, not human examples. This lets engineers explore design spaces closed off by human intuition, surfacing solutions that would have remained undiscovered without physics-first computation.”

Instead, the AI needs to understand the physics of electrical current, electromagnetism, and the physical space the components use, and design the PCB accordingly.

Each subcomponent agent of Quilter AI places components, routes traces, evaluates the physics, and learns which choices produce better outcomes. Millions of iterations refine placement and routing strategies, balancing constraints without human bias, producing unconventional routing methods and placement strategies.

“We’re not trying to match humans. We’re trying to surpass them by avoiding their constraints entirely.”

This means that dozens of layouts are generated simultaneously, each ranked for manufacturability and constraint coverage. Design teams can explore with this tool 100× more design variants without delay or compromise. Users can also quickly test how small or dense a board can realistically be by uploading multiple variations in parallel.

“Quilter is creating the first autonomous PCB design engine. It is not an autorouter, a co-pilot, or an LLM. It is a physics-first AI system that learns from natural law itself, not from human shortcuts.”

Empowering Human Decision

Quilter AI can complete simple board designs in as little as 15 minutes. But for more complex designs with thousands of pins, the system will run overnight.

For each submission, Quilter explores multiple placement and routing candidates across different stack-ups in parallel, giving engineers a range of options to evaluate. They can then download the results into their native ECAD tool, review and refine, and resubmit if needed.

Quilter operates its commercial platform “app.quilter.ai“, out of AWS’s US West region, with plans to expand to US East and eventually Europe. This gives the company a powerful and simple way to scale up its operation by relying on Amazon (AMZN ) cloud computing capacity. It is also possible to run a self-hosted AWS environment so that sensitive data never leaves a company’s infrastructure.

The fact that QUilter does not train on its clients or any other existing board data is also a bonus in a sector where IP protection is extremely important, like aerospace or defense.

The workflow transforms engineers from manual trace-drawers into orchestrators who can run multiple board variants in parallel, turning a quarterly design cycle into a weekly loop of experiments and learning.

“A world with Quilter is a world where boards are as plentiful and iterative as software builds, powering a new paradigm we call Hardware-Rich Development™. Quilter gives top PCB designers the superpower to turn weeks into days. It’s a complete paradigm shift. When you iterate faster, you can out-innovate your competitors.”

The process offers advantages for all professionals involved with PCBs, with faster turnouts, new designs, and easy integration with existing CAD tooling.

Source: Quilter AI

Recently, Quilter performed its most ambitious project yet, Project Speedrun, designing an 843-component Linux computer (two boards, 5,141 pins, high-speed DDR4, eMMC, PCIe, CSI/DSI, GigE). It would normally take 400–450 hours of manual layout to achieve this result. Quilter reduced that to 38.5 hours of human labour, with the rest done autonomously.

“Quilter took care of the repetitive design work while the engineer stayed in control. Automation handled placement, routing, and physics checks, freeing him to focus on firmware prep, documentation, and constraint refinement.”

Investing In Physical AI

NVIDIA

From its origin as a GPU hardware maker for video gaming and other graphic rendering tasks, NVIDIA (NVDA ) has evolved into a massive AI hardware company, giving its stock the world’s largest market capitalization.

NVIDIA realized AI’s potential early, long before anybody, out of specialized researchers, cared about neural networks. This was, at the time, a risky move into an unproven, barely existing sector, or as Jensen Huang put it:

“We’re investing in zero-billion dollar markets.”

In 2016 & 2017, NVIDIA released the Pascal and Volta architectures, respectively, the first GPU-based AI accelerator, while Volta introduced the Tensor Cores, which accelerated deep learning tasks by up to 12 times by 2024.

Investors have been somewhat worried that NVIDIA could soon run out of new markets to justify its high valuation multiples. At CES 2026 (Consumer Electronics Show), NVIDIA announced a new focus on physical AI.

To do so, NVIDIA released Cosmos (ATOM ), a platform to accelerate the development of physical AI for autonomous vehicles (AVs), robots, and video analytics AI agents; Isaac GR00T N1.6, a vision-language-action model built specifically for humanoid robots; and OSMO, an “orchestrator” software, purpose-built for physical AI.

The physical deployment of AI in robots, self-driving cars, and other autonomous systems will provide NVIDIA with many new markets to sell its hardware to. And it seems that AI will also be an enabler in designing physical hardware as well, further enhancing the potential of growing demand for AI computing capacity.

Latest NVIDIA (NVDA) Stock News and Developments