Shodh AI’s LUCAN: What Physical AI Means for Founders Outside Software
For three years, nearly every conversation about AI in business has been about words. Chatbots, copilots, writing assistants and coding agents took the spotlight because they worked on the things software companies already do: reading, writing and reasoning over text. The companies that live on screens moved first, and they captured most of the early value.
A New Delhi company is betting the next wave will happen somewhere far less glamorous. It will happen inside reactors, production lines and chemical plants, where a wrong answer does not produce a typo. It produces a failed batch worth thousands of dollars.
That company is Shodh AI, and its new model is called LUCAN. If the approach works, the founders who benefit most will not be the ones building another chat app. They will be the ones running companies that make physical things.
What Shodh AI Actually Built
Shodh describes LUCAN as a physical AI foundation model built to reason across every scale of the physical world, starting with how molecules behave and moving up through materials, process conditions and industrial equipment to full manufacturing.
The pitch is simple to explain. You tell the model three things: what you want to produce, what inputs you are starting with, and what plant you have. The model then tries to work out a viable route to making it work at scale.
The company was founded by Dr. Arastu Sharma, who did research at the University of Cambridge and Microsoft Research before working in India’s DRDO defence AI ecosystem. Shodh was later selected under the Government of India’s IndiaAI Mission to build foundation model capabilities for science. That support funded the first generation of the model, and the government is now backing a larger scale up of its training and scientific computing capacity.
The Problem LUCAN Is Trying to Solve
Anyone who has worked in chemicals, pharmaceuticals or materials knows the gap between the lab and the plant. A reaction that works perfectly in a flask can fail completely in a reactor thousands of times larger.
The reason is physics. As production grows, mixing behaves differently. Heat moves through the system differently. Fluids flow differently. The equipment itself introduces constraints that never existed on a lab bench. According to Shodh’s announcement, engineers compensate with experiments, pilot plants and accumulated intuition, and moving a new process from discovery to manufacturing can take years.
For a founder, those years are the real cost. Every month spent in scale up is a month of burn with no product revenue. Every pilot plant run is capital that could have gone into sales, hiring or distribution. In industries like specialty chemicals and biomanufacturing, the scale up phase is often where startups run out of money, not the discovery phase.
Shodh’s goal is to move part of that trial and error loop off the plant floor and into computation. Instead of running ten physical experiments to find the right conditions, a team might run the model first, narrow the options, and run two.
The Early Numbers, and How to Read Them
Shodh says the approach is already being tested on real industrial problems. Two figures stand out.
First, in a specialty chemicals project, the company says its modelling identified process changes linked to 17% more product from the same raw material. In a margin sensitive industry, a gain of that size on existing inputs is significant.
Second, Analytics India Magazine reported that LUCAN reduced one factory optimization task from roughly 500 hours to 56.
Both numbers deserve a careful read. They come from the company and from early deployments, and neither has been independently verified. Founders evaluating any claim like this should ask the same questions:
- What was the baseline? A 500 hour process run by a small team is very different from one run by an experienced process engineering group.
- How many projects? One strong result is a case study. Ten consistent results are evidence.
- What still needed humans? A model that narrows options still depends on engineers to validate the result on real equipment.
- Did the result hold in production? A modelled improvement matters only once it survives the plant.
None of this makes the numbers wrong. It just means they are a starting signal, not proof.
Why This Matters Even If You Do Not Run a Factory
Most FounderFeat readers do not operate chemical plants. Physical AI still matters to three groups of founders.
Hardware and deep tech founders
If your company makes batteries, materials, food ingredients, medical devices or anything that needs a production process, the cost of iteration is your biggest enemy. Tools that shorten scale up change your funding math. A company that needs 18 months less runway before revenue needs a smaller seed round and gives up less equity.
SaaS founders selling into manufacturing
Physical AI models need data about equipment, processes and outcomes. Most manufacturers store that data badly, across spreadsheets, old control systems and the heads of senior engineers. That creates an opening for software companies that collect, clean and structure industrial data. When the models arrive, the companies that own the data pipes will be positioned to sell into every plant that wants to use them.
Investors and operators
For years, investors preferred software because it scales without factories. If physical AI lowers the cost and risk of scaling physical products, that preference could soften, and capital could flow back into companies that make real things. Operators who understand both worlds will be valuable.
From Language AI to Physical AI
The first major wave of foundation models learned language. Shodh frames the next frontier as understanding the physical world, with models that reason about molecules, materials, fluids, reactions and machines together.
This is a much harder problem than language, for three reasons.
The data is scarce. Language models trained on the public internet. There is no internet of reactor data. Industrial data is private, expensive to produce and usually locked inside companies.
Mistakes are expensive. When a chatbot gets something wrong, you rephrase the prompt. When a process model gets something wrong, you lose a batch, damage equipment or create a safety risk.
Physics does not negotiate. A language model can produce an answer that sounds right. A process either works at scale or it does not. That makes validation slower but also makes success easier to measure.
Those same barriers are why a working physical AI model would be hard to copy. Whoever builds reliable models and gathers proprietary industrial data first could hold a lasting advantage.
What Founders Should Watch Next
LUCAN is early. Here is what will show whether it becomes infrastructure or stays a promising announcement:
- Independent validation. Results published with partners, or reviewed by outside engineers, would carry far more weight than company figures.
- Access model. Whether Shodh offers LUCAN through an API, through partnerships or only as a service will decide who can actually use it.
- Pricing. A tool priced for large chemical companies does little for a ten person materials startup.
- Data requirements. If the model needs years of detailed plant data to work well, smaller companies will struggle to benefit.
- Competition. A category this valuable will attract larger players. How Shodh defends its position will matter as much as the model itself.
The FounderFeat Take
LUCAN is unproven at scale, and the published results come from the company itself. But the direction is right. The AI conversation has been dominated by tools that make knowledge workers faster, while the industries that make the physical world run have seen far less benefit.
If physical AI works, it will not look like a chatbot. It will look like fewer failed batches, shorter scale up timelines and smaller funding rounds for companies that build real products. For founders outside software, that is the AI story worth following.
