Business models for sdl

I believe there are, fundamentally, 4 distinct business model approaches for SDL's:

Building Infrastructure

I haven't seen a lot of companies focusing on delivering system integration solutions for laboratories. Many are limited to specific movement tools (robotic arms, liquid handling), some are technique-specific (a High-throughput system for XRD, for example) but none offer the possibility of creating a whole system.

This is one of the things I learned at Orange Quantum Systems: you can become a system integrator expert. You don't need to redevelop everything from the ground up, you can still work with what customers prefer and build an expanding expertise.

Selling Results

If you have the SDL infrastructure in place: i.e. a collection of tools that work together and the algorithms to run them, you can sell results. For example, a company may approach you with a request to explore a given range of material compositions and to optimize for a metric. For example, they may be interested in exploring alloys with different stoichiometries and to measure corrosion resistance for each one.

The customer owns the data, while the SDL keeps the IP on the data generation.

Selling Discovery

This approach is trickier but could be very valuable. Instead of selling results on a given exploration space, you leverage the hardware and the software to come up with new formulations.

The challenge is that there's no specific metric to make the process "fair": I can't guarantee I will discover something new (I would do it for myself), but the effort itself must be compensated.

The customer pays for the "time" it takes to run the discovery cycle. The company keeps the learnings (model optimisation).

The IP Discussion

The last two business models have a very subtle discussion on IP and who owns what. The result-driven approach, for example, warrants a discussion regarding the training of new algorithms. It would seem fair that the person who pays for results is the sole owner. Therefore no new training is possible and if a second customer comes with exactly the same request, the output would be exactly the same (no better).

But this is not how things work in practice. There is always some learnings (even if not algorithmically). Operation of the instruments, brands that work better, influence of other parameters such as room temperature, etc.

The discovery path, on the other hand, relies heavily on what algorithm is available. First to produce hypothesis, then to validate them experimentally. The algorithm should be designed to improve over time, which means it will leverage past experiences.

In the "Selling Discovery" approach, it is likely that the "best candidate" is the only result presented. While the entire exploration context is what the company keeps as trade secret. A second discovery campaign can reach better and more accurate results.

Selling Data

(See: modeling data as a scarce resource, which is a bit broader than what is intended in this discussion).

In the perspective of more and more companies trying to get their hands on data to train and refine their models, there's a high likelihood that data itself is a commodity that can be sold. Curated (and expanding) datasets including as much metadata as possible.

I foresee this to be a topic that could be financed by large consortia (publicly funded), and which could be made accessible to companies via licensing.

Data, however, is not an asset class that has been identified clearly by investors and other companies. It is also unclear how compartmentalisation would happen, if data generated in one campaigned can be leveraged to create better "discovery" paths, or whether "results" could be used to increment the pool of available data.