When AI designs the parts, it won’t design them the way we do.

Metal additive manufacturing has been poorly explained for fifteen years, and this is partly our fault, the industry's. We've spent that time defending part-by-part technology against a rival that almost always wins: a part designed for milling is better manufactured on a milling machine.

It's not a machine problem. It's a problem with who designed the part.

The world is full of parts designed to tear material.

Today's designers learned to design by thinking about how the work would be done. This gives rise to a whole vocabulary of shapes: flat surfaces for fastening, straight holes because the drill bit is straight, draft angles because the part has to be removed from the mold, minimum radii because the milling cutter has a specific diameter, and uniform thicknesses because the material cools evenly.

That vocabulary is so ingrained that it's practically invisible. And when one of those parts arrives at an LPBF machine, the comparison is stark and fair: it takes longer and costs more. Of course. You're asking the additive manufacturing machine to produce a part designed for something else.

That's why the commercial argument for additive manufacturing has always been defensive: short runs, spare parts without blueprints, prototypes, impossible geometries. Niches. Real ones, but niches nonetheless.

What changes when the designer optimizes by loads

An optimizer doesn't have that vocabulary. It doesn't know what a draft angle is, nor does it care. You give it available space, anchor points, loads, and safety factors, and it returns the material distribution that supports that with the minimum mass.

What emerges doesn't resemble a catalog item. It resembles a bone, a root, or a spider web: nerves that follow the path of the load, voids where no material was needed, sections that continuously change shape, lattice infill where the part needs rigidity but not mass.

And that geometry has a property that changes everything:

• It cannot be molded. There is no possible demolding direction.
• It cannot be machined. There is no way for the tool to enter.
• It cannot be cast without cores, which are, in turn, impossible to manufacture.

In other words: the part that is impossible for all other processes is exactly the one that an LPBF machine manufactures effortlessly, because the complexity of the shape doesn't matter to it. What it struggles with is the volume of material, and the optimizer has just cut it in half.

At that point, additive treatment ceases to be the expensive alternative. It becomes the only one.

What will really multiply this is not the tool: it's who can use it

Topological optimization isn't new. It's been in simulation software for over a decade, but its use has been limited for a practical reason: it required a specialist. Someone who could formulate the load case, interpret the result, and redraw it as a manufacturable part. That profile is scarce and expensive, so the technique remained confined to aerospace and competition.

What's changing isn't so much the power of the optimizer as the level of expertise required to use it. When simply describing the problem in words is enough to obtain a reasonable design, the technique moves from R&D departments into regular engineering offices.

And that's where the scale changes. It's not that the parts being optimized today are being designed better: it's that parts that no one is currently considering optimizing are going to be optimized, because it wasn't worth paying a specialist for a part of which two hundred are made per year.

Each of those parts is a new application of technology that is not currently on any list.

What's not going to happen, and it's worth saying

This story is usually told without the fine print, so here it is.

Optimizing doesn't always pay off. Removing 30% of the mass from a part is a triumph in an airplane, where every kilogram is paid for in fuel for twenty years. Not so in a bracket bolted to a wall. If the weight savings aren't worth money, the optimized part is simply a more expensive part.

Organic doesn't mean manufacturable. Much of what an optimizer returns can't be printed as is: overhangs that would require supports impossible to remove, thicknesses below what the process can handle, internal cavities from which there's no way to remove the powder. The designer still needs to understand the process; what changes is that now they need to understand this process and not the milling process.

Trapped powder is the real limit. A lattice with closed cells is a part full of powder that will never escape. That's not an aesthetic rule: it's the first constraint that must be imposed on the optimizer.

And certification is lagging behind. An organic part is more difficult to inspect and more difficult to defend before an auditor than a prismatic part with traditional blueprints. That will be resolved, but today it's a real obstacle.

None of these four things invalidates the argument. What they do is highlight where the work lies: not in waiting for the tool to arrive, but in knowing how to design for this process when it arrives.

What does this mean if you're buying a machine now?

The parts you'll be manufacturing in five years' time will probably be very different from the ones you're manufacturing today. And from this arise two practical criteria that determine which machine is best:

That you can parameterize it. If the machine comes with fixed parameter sets, when a new geometry appears that requires a different sweep strategy, you won't be able to develop it. You'll have to request it. Our software is ours and the parameterization is open is precisely for this reason.

Make testing inexpensive. Most new applications don't appear in a meeting: they appear because someone tried them. If each trial requires filling the entire vat with expensive powder, almost no one tries. That's what the SamyFlex is for, and it's the reason we have an open quote calculator: to lower the cost of asking "what if…?"

And the other half of the story

This is about the AI before the machine: the one that decides what shape the part has. The AI inside the machine—the one that looks at the powder bed layer by layer and decides if the build is going well—is a different story, with its own list of what's working and what isn't yet.

It's in what can AI really do in additive manufacturing, with six questions to ask any manufacturer who tells you that their machine uses artificial intelligence.

And if what you have is a specific part and you want to know if it fits, bring it to us: there's no need to wait for that.