What Should a Company Really Optimize with a Self-Developed MPCVD System?

A few days ago, I published a piece in the knowledge community called “Self-Developed MPCVD Equipment: You Save on the Purchase Price, but You Are Betting on Yield.” I ran through a simple calculation: for one type of equipment configuration I have encountered, developing the tool in-house might save RMB 50,000–100,000 per unit. That is only about 2–5% of total manufacturing cost. If stable output falls by two, three or four percentage points, the saving is largely gone.

After writing it, I thought about Diamond Foundry. It has also developed its own reactors for years, and its website says the technology has reached a tenth generation. That led to a broader question: if the procurement price difference affects only a few percentage points, should a company that invests people, time and capital in equipment development not receive something more valuable than those few points?

Three things self-development can optimize

In several lab-grown-diamond projects I have encountered, reducing purchase cost is the most direct and easiest reason to quantify. But self-development can create three layers of value.

The first is cost: procurement, maintenance, spare parts and expansion expenses. For a company already in scaled manufacturing, this layer is real.

The second is control: ownership of equipment interfaces, core parameters, maintenance capability and data. Process changes are no longer fully dependent on an external equipment supplier.

The third is learning: shortening the time between a problem appearing and the next verification cycle. When a material defect appears, how quickly can the team determine whether an equipment variable caused it? When a new process window is found, can the equipment be modified quickly? When a customer asks for something new, can the reactor, process control and recipe change together?

These three layers are not parallel choices. Without data and modification rights, there is no shorter feedback loop. Control is a precondition for learning.

The real variable is therefore not simply “buy or build.” It is how open the equipment is to the materials team and how quickly the team can turn results into changes. Purchased equipment can also reach the second and third layers if the supplier provides sufficient data and responds quickly to process modifications.

The Diamond Foundry example

My earlier reading of Diamond Foundry was that its core self-developed capability was to shorten the feedback path from equipment performance to material output, allowing the reactor to be repeatedly redesigned around the material result.

The company website says that its founding team brought a manufacturing lesson into diamond: instead of pursuing a one-time “10x breakthrough,” it values rapid, iterative and compounding improvement. It says the team began in 2012 by designing a new plasma reactor from Maxwell’s equations and extensive physics simulation, built three reactor generations in 2013, and spent eight months developing its first complete reactor in 2014, nearly exhausting its funds.

This suggests that rapid iteration was part of the company’s method from the beginning, at least within its own account, rather than a story added only after success.

The website currently says that the reactor technology has reached a tenth generation, with each generation improving quality, yield and productivity while continuing to adjust plasma geometry. “Tenth generation” by itself only shows that the equipment has remained an R&D object for a long time. It does not prove that the feedback loop is short: outsiders cannot see the interval between generations or the specific problem that drove each change.

It does at least indicate a continued effort to adapt the equipment to material-production capability.

R&D must change; production must remain stable

Learning-oriented self-development immediately creates a conflict: how can multi-tool consistency be maintained if the equipment keeps changing?

If dozens of production tools change every day with the R&D team, comparing batches becomes increasingly difficult. A more reasonable arrangement is continuous iteration on the R&D side and periodic freezing on the production side. New chambers, sensors and control logic are first validated on development tools. Once their effect is proven, they are fixed as a version and copied into production.

In short: change is allowed between generations; stability is required within a generation.

Semiconductor manufacturing has a related practice known as Copy Exactly, associated with Intel: process and equipment configurations validated on an R&D line are copied into the production line. Data is what prevents iteration from destroying consistency. Every generation must record what changed, which state it targeted and what result followed.

The logic for self-developed MPCVD is similar. A new generation can keep changing, but once it enters production it must become a stable version.

How to tell whether a self-developed system is learning

I now use one question: how many of the problems exposed by the material and production sides have actually entered the next equipment and process design?

Did a certain defect lead to a chamber or susceptor change? Did a batch-consistency problem lead to an upgrade in sensors or process control? Did difficulty restoring the post-maintenance state lead to a new monitoring or calibration method? When the fleet expanded from a few tools to dozens, were multi-tool differences written into the standards for the next generation?

The reverse is also true: an equipment platform remaining unchanged for years does not necessarily mean that it has stopped learning. Once equipment reaches mature production, stability is exactly what should be expected. The question is whether the company can change quickly when change is needed and copy consistently when freezing is needed.

Continuous iteration requires organizational and financial capacity

Reaching ten equipment generations is not only a technical matter. Each generation requires engineers, prototypes, tests, failure costs and time, together with new process validation and production data.

My hypothesis is that truly valuable equipment capability often begins to take shape only after the second or third generation. If a company’s cash flow is sufficient only to build the first generation, self-development can stop at “we managed to make it work.”

Diamond Foundry’s tenth-generation equipment also required more than technical persistence. The company says that its initial funds were nearly exhausted while it was developing the third-generation machine. It later continued through financing and cash flow from producing lab-grown diamonds. In 2021, it completed a USD 200 million Series C financing round.

Building prototypes, validating them and expanding production require a long financial window.

There is also an open question about Diamond Foundry’s starting point: perhaps the reactor performance and product it wanted—large-area single crystal aimed at semiconductor-grade applications—were not available as an off-the-shelf tool. I do not yet have enough evidence to confirm that hypothesis. If it is true, self-development began as a technical-route requirement, with learning speed emerging as an additional benefit.

Returning to the original question

Self-development is not the goal. The goal is to build manufacturing capability that is more valuable than external procurement.

Before starting, a company can ask three questions:

1. Am I optimizing cost, control or learning speed?
2. Can equipment problems and material results form a sufficiently short feedback path?
3. Once the R&D design is mature, can I freeze it and reproduce it across dozens or even hundreds of tools?

Equipment can be purchased or built, but it cannot be separated from the material result.