Most laser powder bed fusion machines already photograph the powder bed after every layer. In most shops those images go into an archive and get looked at only after something has gone wrong. Icelandic startup Euler thinks they can do more. Its software-only monitoring platform reads the layer images a printer already captures, analyzes every layer in real time, and tries to predict defects before a build is lost. It needs no added sensors and no hardware retrofit. The company's pitch is short: "Euler uses the camera already in your printer."
The claim now has numbers behind it. A joint study with the Danish Technological Institute (DTI), reported by 3D Printing Industry, measured how much machine time goes to builds that are already doomed. The answer: more than a quarter of it.
The Study: 47 Days, 54 Builds, Four Machines
The DTI work ran for 47 days and covered 54 builds on four Nikon SLM systems. Six alloys were used, including 316L stainless steel and Ti6Al4V. That spread matters. A monitoring system tuned to one machine and one material can look good in a demo and still fail in a mixed production environment. Testing across several machines and several alloys is a better approximation of how a contract shop or a university lab actually runs.
The headline figure is that 27.8% of total print time went to builds that ultimately failed. Those were hours of laser time, inert gas, and powder handling that produced nothing usable. The more interesting number is the second one. According to the study, 77% of that wasted time could have been saved if operators had stopped each failing job at the point where Euler's engine predicted it would fail. Across the study period, that comes to roughly 10 days of printing.
Put another way, the problem isn't only that builds fail. Many failing builds keep running long after the outcome is decided, because nobody, human or software, has flagged them yet. A powder bed machine left running overnight will happily keep melting layers onto a part that warped hours earlier.
What the Software Actually Watches For
Euler describes its predictive engine as patent-pending and says it is designed to warn of likely failure modes before the job is lost. The failure types it cites are familiar to anyone who has run a powder bed system:
- Warpage: parts curling or lifting as residual stress builds up.
- Powder short feed: layers where the recoater doesn't deliver enough powder to cover the bed, leaving bare or thin patches.
- Streaking: lines dragged across the powder bed, which show up in the layer image as visible artifacts.
All three leave visual traces in a top-down layer photo, which is why a camera-only approach is plausible at all. The difference between Euler's framing and a simple anomaly flag is the word "predict." The company is not just pointing out that a layer looks bad. It is claiming it can tell when a pattern of layers means the job is heading toward failure, early enough that stopping is worth it.
The broader positioning is about consistency. On its site, Euler says its aim is to eliminate machine-to-machine and build-to-build variability, the long-running complaint that the same file on the same model of printer doesn't always give the same part. The platform is hardware-agnostic, and the company says it connects to an existing setup in under five minutes. That is a vendor claim, and anyone evaluating it should test it on their own machines.
Why the Economics Work
Metal powder bed fusion is expensive to run, and failed builds are expensive in ways that don't always show up on a quote. The 3D Printing Industry piece cites a Wohlers estimate of $75 per hour in operating cost for a $1 million machine. As a rough illustration, not a figure from the study, 10 days of recovered printing at that rate is about 240 machine-hours, or around $18,000 in operating cost for a single machine. The study's authors, using the same Wohlers figure, put the potential saving at over $115,000 per machine per year in operating costs alone, before wasted powder, personnel time or post-processing. Neither number counts the value of whatever the machine could have been printing instead.
That arithmetic explains the software-only design. Adding melt-pool sensors, thermal cameras, or acoustic monitoring to an installed fleet means capital expense, integration work, and sometimes requalification. A tool that runs on images the machine already produces avoids almost all of that. The hard part is entirely in the software.
Formnext and What's Coming
Euler will show the platform at Formnext 2026 at Messe Frankfurt, 17–20 November, at booth A32 in Hall 11.0. The company's event page lists three new additions: an Euler MCP, 3D heatmaps, and alerts through Slack and Microsoft Teams. The Euler MCP, based on the Model Context Protocol, lets engineers connect Claude or ChatGPT to their build data and ask questions such as which builds this week showed defects, and on which machines.
The alerts are the most practical of the three. A failure prediction only saves time if someone acts on it, and a notification in the team chat is far more likely to get attention at 2 a.m. than a dashboard nobody has open. 3D heatmaps suggest the per-layer analysis can be stacked into a volumetric view of where in the build problems are forming. That helps with root-cause work on orientation, supports, and scan strategy.
What It Means for Makers
Few hobbyists own a laser powder bed fusion machine, but plenty of FilamentFeed readers work near one: in university labs, service bureaus, makerspaces with industrial partners, or engineering jobs that send metal parts out. Three points are worth taking from this story.
First, the waste figure is a useful reality check. If roughly 28% of machine time across a controlled, multi-alloy study went to builds that failed, metal AM is still far from push-button. Anyone pricing metal parts, or wondering why lead times slip, should keep that in mind.
Second, the lesson carries over to polymer printing. Plenty of polymer printers have cameras, and the idea that existing imaging can do more than produce timelapses applies at every price point. The difference is what's at stake: a failed overnight PLA print costs a spool and some patience, while a failed titanium build costs days.
Third, treat the numbers as a strong first result, not a settled verdict. Fifty-four builds on one manufacturer's machines is a meaningful sample for an early study. The real test is whether the prediction holds up on other platforms, other geometries, and in shops where nobody from the vendor is watching. Euler's hardware-agnostic claim makes that the obvious next question, and Formnext is a good place to ask it.
