The Perfect Simulation Company

Over the past few months we have been continuing to dive into different forms of digital twins and simulation technology. We wrote about agentic entity simulation and have continued to explore the new forms of sim tech that have been emerging.

This week we want to take some time to recap what we have learned from dozens of operators and highlight the key areas that we think will make or break a company building in this space.

Why Now?

Over the past few months we have spoken with dozens of companies that are building in or around simulation technology, and they very broadly fall into two buckets (however in practice they overlap meaningfully):

1. Situation simulation / decision co-pilot: These take large amounts of information that were previously either inaccessible or difficult to synthesize and provide potential action plans.
   1. Examples: Stock trading, business strategy, policy decisions
2. Physics / Process simulations: These embed the laws of physics or other rules/processes into the AI, allowing it to solve different problems with enormous numbers of variables by rapidly iterating through viable opportunities.
   1. Examples: Wavelength physics, factory planning, chip verification

We believe that the recent proliferation was catalyzed by AI models that allow these companies to access and process large amounts of data that can be used in simulations (specifically the situation simulation bucket), and also rapidly iterate through variables (primarily the physics and process buckets).

For each of these buckets we are going to walk through the initial data that is used, how the products validate outcomes, and how they improve over time.

Data-In

As with any data model, the data that is going into these simulations is going to dictate the quality of the answers that are coming out.

For situation simulation the amount of data that could reasonably be used to impact a decision is infinite. For example, a decision co-pilot for a Fortune 500 company with fierce competitors and a robust supply chain could benefit from an enormous variety of data from competitor website pricing to weather conditions for their core suppliers. The goal here is not to be comprehensive, but to utilize all of the data that is in front of the decision maker and synthesize it in a way that is digestible and actionable.

In physics-based/process-based simulations, the data is primarily the set of rules and conditions that you are required to follow in order to have a viable output. That could be a list of regulatory demands, best practices for how to build a factory, or even very specific laws of physics that will be used to solve complex problems.

Data Compounding & Validation

Validation has consistently come up as the single most important part of a simulation business. If the end user can not trust the outputs explicitly then they are largely considered useless. We have seen dozens of different approaches to this, from user-level validation to industry-specific solvers, to mathematical approaches like Lean.

For both buckets of sim tech, finding a way to utilize the validation step to create new data and better systems is what will create a moat and a defensible business.

For situation simulation, a simple version of compounding data could be the continuous capture of data from internal systems such as a Customer Relationship Management (CRM) system. However, the more valuable data is the feedback from the user on which scenarios or proposed action plan is correct and why. The problem is that “what is correct” is not easy to identify or explain. Complex decisions tend to have more than one right answer and for experts who rely on years of experience, it is not always easy to explain why it is correct. Even if you were able to perfectly provide feedback to the system, each situation is different, so the data and learnings from one problem may not transfer to the next, making it challenging to capture enough data to dramatically improve the predictive capabilities of the system.

For physics-based or process-based simulations the underlying initial data tends to not change much as the laws of physics have been established for quite some time. However, what can improve is the way in which these rules or laws are strung together, allowing for more valuable or faster iterations.

As a general rule of thumb, when you have the ability to monitor a closed-loop system, that is where the best data can be utilized. For example, if I am presenting a decision to senior management and they can directly provide input on its accuracy, that is a closed loop. If I provide a chip-based simulation platform to a leading chip designer and they do not tell me how the chip performs, it becomes extremely difficult to improve over time because you are not getting the validation data. Equally importantly is how fast the loop closes. Senior management may not be able to validate a simulation’s strategy was effective until results come in months or years down the road, but a chip is validated on fabrication.

Adoption Timing

In the long run we believe that, in the same way that everyone is adopting AI, we will see the adoption of different forms of simulation technology. We also expect that adoption will be staggered. While large incumbents are likely the most equipped to have the data and feedback loops for these tools, they also tend to be the slowest to move. We think that there is a unique period of time where startups can utilize these tools to rapidly accelerate into an industry, either reducing lead times, improving operational efficiency, accelerating research, or just making better decisions.

Because we believe that these will be used so broadly, we also think that it is important for any startup in this space to focus on a specific niche. There will be room in the market for general physics-based tools or general decision-making frameworks, but we believe that this will be the focus of some of the large AI incumbents as well. A tool that is uniquely qualified to understand hydrodynamics, for example, will likely grow unencumbered.

We are excited about any company that is innovating with simulations and digital twins to accelerate their growth and we would love to talk.

Takeaway: Simulation startups split into two buckets: situation simulators that synthesize messy data into decisions and physics/process simulators that iterate within fixed rule sets. Both have seen rapid adoption alongside the proliferation of AI. Validation has become the most important part of the simulation process. Users will not act on outputs they can not trust, and the winners will turn validation into a rapid feedback loop that compounds. Since we believe that general-purpose tools will be a major focus of AI incumbents, we think hyper-verticalization is the best path forward for an early stage startup.