Build or buy a CAM engine?
At some point a software product that touches manufacturing runs into geometry. An ERP system needs to know how much stock a part will consume. An MES needs to predict cycle time. A quoting tool needs to price a part nobody has made yet. An automation platform needs to drive a machine that expects a toolpath, not a solid model.
The question that follows is always the same: do we build this, or buy it?
This is written for the people who have to answer that. It is not a sales argument — the section on when building wins is meant seriously, and some readers should reach the end and go and hire a geometry developer.
What “CAM capability” actually decomposes into
The usual mistake is to scope the algorithm and forget everything around it. A working capability has at least five layers, and the interesting one is rarely the one teams budget for.
Import. Your customers will send you whatever their CAD system produces. STEP and IGES are the neutral formats, but native files from SolidWorks, Inventor, CATIA, Solid Edge, NX and Pro/Engineer all turn up, and each has its own release history. Supporting “STEP” is not one thing — it is a decade of dialects and a body of edge cases.
Healing. This is the layer that surprises people. Real CAD data arrives with gaps between surfaces, duplicate faces, self-intersections, tolerance mismatches and features that were valid in the originating system and are not valid in yours. A geometry kernel that assumes clean input will fail on a large share of genuine customer files. Most of the engineering effort in a mature import path is here, not in the parser.
The kernel. Boolean operations, offsets, intersections, projections, tessellation. This is well-trodden ground and you can license one, which most people do. It is also the layer most likely to be blamed when the real problem is upstream in healing.
Strategy. Turning geometry into a manufacturing intention: where the cuts go, in what order, at what depth, with what approach and retract. This is where domain knowledge lives, and where a general-purpose kernel gives you nothing. Wire EDM taper and multi-pass work, for instance, has behaviour that a milling-shaped mind gets wrong the first three times.
Output. Toolpath in a form a specific machine and controller will accept, plus the things around it nobody counts as engineering — drawings, bills of materials, exports into whatever the customer’s ERP expects.
A team that scopes “we need to generate toolpaths” and budgets for the fourth layer has usually underestimated the first two by an order of magnitude.
The part that does not appear in the estimate
Version one is a project. Everything after it is a commitment.
CAD vendors ship new releases every year and file formats move with them. Operating systems and toolchains change underneath you. Customers find geometry that breaks assumptions you did not know you had made — and they find it in production, on a Friday, on a part they need quoted by Monday.
The honest way to think about this is not “what will it cost to build” but “what will it cost to own for ten years”. That figure includes at least one person who understands the geometry code well enough to fix it under pressure, and it includes what happens when that person leaves.
None of this is an argument against building. It is an argument against costing only the first version.
When building is the right answer
Genuinely, and not rarely:
- The geometry is your differentiator. If the thing that makes your product worth buying is the way it reasons about shapes, outsourcing it hollows out the product. Buy the kernel, build the strategy.
- The scope is narrow and stable. One material, one process, one family of parts, one machine. Narrow problems are tractable, and a general engine may be a poor fit for a specific job.
- You already have the people. A team with geometry experience will build something better suited than anything they license, because they can shape it around the rest of the product.
- Licensing terms will not work. Some markets, some jurisdictions and some customers make third-party runtime licensing genuinely awkward. That is a real constraint, not an excuse.
When licensing wins
- Geometry is a feature, not the product. If manufacturing intelligence is one capability among twenty, it should not consume a third of your engineering budget.
- You need breadth on day one. Supporting a dozen CAD formats properly is years of work that has already been done, several times, by other people.
- The domain knowledge is not in the building. Toolpath strategy is not something a strong generalist team picks up in a quarter. It is decades of accumulated cases, most of them learned the expensive way.
- Time to market matters more than control. Sometimes shipping in months against years is the whole argument.
Three shapes of buying
They are not equivalent, and the choice matters more than most teams expect.
An embedded engine licence. The computation runs inside your product, on your customer’s machine or your infrastructure. Best when you need the capability offline, in a controlled environment, or as an inseparable part of your own application. It is the tightest coupling and usually the highest commitment on both sides.
Bespoke development. Somebody builds the specific capability you need, to your specification, and it becomes part of your product. Best when the problem is genuinely unusual and no off-the-shelf component addresses it — which is more common in manufacturing than in most software markets, because processes vary by shop.
An API. The computation happens elsewhere and you call it. The lowest commitment and the fastest to try. The trade-offs are the obvious ones: a network dependency, and the question of whether customer geometry may leave their premises at all. For some customers that answer is a flat no, and it is worth asking before you design around it.
Can you sell it the way you sell your product?
This one gets asked too late, and it is expensive when it does.
Your product probably has a shape to it commercially — a standard edition and a professional one, a trial, modules people buy separately, a free tier that exists to get them started. That shape is not decoration. It is how your customers expect to buy, what your pricing page is built around, and what your sales team knows how to sell.
Now suppose the engine underneath it can only be licensed one way: everything, permanently, per seat. You have three options, and all of them are bad. Rebuild your commercial model around your supplier’s. Absorb the cost of the full engine in your entry tier and watch the margin go. Or ship capability in a free trial that you are paying full price to provide.
So ask early: at what granularity can this be licensed? Per feature, per module, per tier? Can it be time-limited for a trial? Can a capability be dark in one edition and lit in another, from the same build?
If a supplier has never had to answer that, they will tell you it is possible. The useful question is whether they have done it — in their own product, in front of paying customers.
We have. Nucleo ships as two tiers, XY and XYZ+, with individual capabilities gated per licence and a trial system alongside them. The weld automation suite is an add-on rather than an edition of its own, because that is how the market for it turned out to work. That machinery exists because we needed it ourselves, and it is the same machinery that decides what a licensed engine exposes inside somebody else’s product.
Which means the commercial packaging is a conversation rather than a constraint. If your product has three tiers, the engine can have three tiers.
Questions worth asking any vendor
Including us. If a supplier cannot answer these, that is information.
- What happens to my customers’ geometry? Where is it processed, what is retained, and can it be run entirely on-premises?
- Which CAD formats and which versions, specifically — and what happens when a customer sends a file from next year’s release?
- What does a bad file do? Fail loudly, fail silently, or produce a plausible wrong answer? The third is the dangerous one.
- Who fixes a geometry bug, how quickly, and is that in a contract or in an email?
- What is the exit? If the relationship ends, what happens to products already shipped to your customers?
- How long has the engine been in production, and against how many real parts? Geometry code is hardened by exposure, not by design.
- At what granularity can it be licensed — per feature, per tier, time-limited for a trial? And have you done that for someone before, or would mine be the first?
Where we fit
Camtek Software has been building this kind of software since 1983. PEPS — sold as OPTICAM Classic in North America — has issued over 40,000 licences worldwide, and OPTICAM runs inside SOLIDWORKS and Siemens NX, so embedding a computation engine inside somebody else’s application is not a new exercise for us. The Wire EDM toolpath engine was developed with the Camtek team in Germany, and Nucleo generates complete welding fixtures from imported solid models.
We work in all three shapes above: bespoke OEM development, B2B engine licensing, and a Web API currently in pilot with general availability due in 2027.
If you are weighing this up, the useful conversation is usually not about our software. It is about which of your problems is actually geometry, which is data plumbing, and which is domain knowledge you already have in the building. Get in touch and we will tell you honestly which parts we think you should build yourself.