Building with metal 3D printing is a well-established technique, but it's by and large too clumsy, expensive, or imprecise to compete with other traditional methods on scale. Funded to the tune of $14 million from Nvidia and Boeing, Freeform aims to do that by building a new metal additive printing process that it says changes the game—and yes, there's an AI angle to this, too.
Co-founders Erik Palitsch (CEO) and TJ Ronacher (president) both worked at SpaceX, where they were principal architect and lead analyst, respectively of the Merlin engines and other programs. While there they could see the potential of making parts with 3D printing using metal, but they also got to personally understand the shortcomings of the method.
"We saw the potential of metal printing-it can change basically any industry that produces metal goods. But its adoption has been slow and its success at best very marginal," said Palitsch. "Why is it not practical to use at scale? Fundamentally, because of three things: crappy and inconsistent quality; speed-commercial printers are very slow; and cost-the price for these printers is astronomical."
They concluded that if they could operationalize the process in order to have a printing service rather than selling the printer, they would break the whole thing wide open. So they teamed up with Tasso Lappas, former CTO of Velo3D to start Freeform.
The first mistake that companies were making was mimicking the likes of CNC machines, mainly because that's what they're used to in traditional manufacturing. With a CNC machine, you sell the machine and the software, and it works with whatever shape and process you use. But metal additive is different, Palitsch said.
"The way these things work today is they're 'open loop' — they're basically playing back a file," he explained. "They needed to be smarter than that, because the process by which you melt metal powder with a laser is extremely complicated, and in a way infinitely variable."
Selling people a machine and saying "become an expert to make it work, good luck" isn't a recipe for success.
"But when you decide you're not going to build and package a printer into a box, when you have the freedom to build an automated factory from clean sheet, there's a lot you can do," Palitsch said.
Their solution will be printing as a service through a closed-looping process in a custom-made machine that it monitors on a microsecond scale, adjust most all the factors that will give you the kind of print you expected to happen in a place like SpaceX.
SpaceX already has plenty of tech advances but what is more immediately to the point are the feedback loop and the AI managing it.
"We have high-speed computer vision feedback on our system running at the microsecond scale, and all that data is being processed on state of the art FPGAs and GPUs. We had to build this whole stack ourselves out of stuff that's only become available in the last few years," added Palitsch.
An important note is that the closed-loop system means it has real-time monitoring, thus avoiding some quality issues that may develop, but very efficiently allowing speed printing of complex geometries. And by making it operate as a service, they keep the business model simple.
Implementing that part of the system was a need for the second tech breakthrough: a fast machine-learning model expert enough to virtually perform that monitoring.
"Erik and TJ lived this and reached the same conclusions, that this industry required a level of compute and sensors that no one had ever deployed before," Lappas said.
"To properly understand how to control the process we needed datasets working at timescales that no one had. So we started building a state of the art telemetry system, a platform that would collect curated, controlled, almost self-labeled datasets."
This allowed them to bootstrap a model to produce more data for a better model, and so on.
But then they confronted the need for velocity.
"There's a lot we have in common with generative models, and a lot we don't. But one thing that's absolutely different is the latency," Lappas said. "Our inference needs to happen in microseconds so that we can close the loop on these processes." With no off-the-shelf solution available for the data or the compute, they had to build the GPU/FPGA "AI on steroids" combo from scratch.
A consequential side effect: Freeform is "building the largest metal additive dataset in the world — that's why companies like Boeing are coming to us," said Palitsch. "We have this fundamental, core data collection and processing ability no one else has."
Add that to the fundamental benefits of printing-based manufacturing, like the agility and versatility of factories, and it makes a pretty compelling business case.
A total of $14 million went from Boeing's AE Ventures and Nvidia to Freeform, although neither company would provide a breakout for that amount. Each brings benefits: Nvidia provides access to H100s and other compute hardware, and Boeing will guide it through the supplier qualification process and likely purchase a quantity of parts. (Freeform will also join Nvidia's Inception startup program.)
Palitsch said they have customers in the aerospace, automotive, industrial, and energy sectors, "the whole nine." They declined to put any on the record, but did mention they're making everything from rocket engine components to exhaust parts for Formula 1 cars. They plan to use the money to scale up, build out their next generation of (much faster) printers, and hire up to around 55 people total over the next year.
A source of pride and a big complement: It's taken some time to get off the drawing board and into reality, but Palitch cited the slow, methodical, technical approach they took to their problems as part of the reason for their success.
"It was a slow transition," said Palitch. "But I look back at it… with six people, we built, from scratch, the fastest laser melting platform on the planet, and the hardware and software for it. We did things people said you couldn't do."