Skip to main content
top of page
Speciale Invest Logo

Investing in Arborum Labs – World Model for Complex Networks

Writer: Sunil Cavale
Sunil Cavale
6 hours ago
5 min read

Speciale Invest has co-led the pre-seed round in Arborum Labs, a company building world models for critical networks. Its models learn how cause and effect play out inside power, telecom and other networks, so that AI can help understand them and, in time, run them.



Almost everything we do depends on a handful of huge networks: the power grid, telecom, logistics. They're among the most complicated things people have ever built, with millions of parts that change all the time, and a failure in one place often turns up as a problem somewhere else.


Most of them are still run the way they were twenty years ago, with alarm screens, rulebooks and a few senior engineers who know the network's quirks by heart. That's getting harder to sustain. We're building infrastructure faster than at any time in decades: new highways, rail corridors and freight networks, grids stretched to carry more power, data-centre demand that could double by 2030, and more than five lakh 5G base stations put up in India in just a few years. Each addition makes networks that were already difficult to run a little more so.


Operators are making big calls about where to build, how to run what they have and which investments will pay back, often with a fair amount of guesswork. That's the problem Arborum Labs is going after, and why we're co-leading its pre-seed round. The company is building world models for critical networks, starting with power and telecom: AI that learns how a network behaves, including cause and effect, so it can explain what's going on and gradually start acting on it.


Twenty years of “almost there”

Telecom has talked about autonomous networks for two decades, and power companies about self-healing grids. Both have stayed a few years away for most of that time, largely because the tools haven't been up to it. Rules only catch the failures someone thought of in advance, and pattern-matching AI only catches the ones it has seen before. Networks keep coming up with new edge cases. When dropped calls spike in a city, the cause could be weather, a software patch, a cut fibre or a single tower that's lost power and working out which usually takes a person that has context or experience or both. Until a system can explain its reasoning, operators are understandably reluctant to let it make the call.


Learning how networks behave

Language models (LLMs) learned from words. We think the next big step in AI comes from learning how real systems behave. A model trained on a thousand outages can tell you another is probably coming, but it can't tell you why, or what would happen if you changed something.


Arborum Labs sees autonomy as a question of understanding. For a network to make its own decisions, it needs a model of cause and effect that can predict what a change will do, work out what went wrong and weigh up the options before anyone acts.


People sometimes compare this to a digital twin, and the difference matters. Digital twins are usually built by hand from engineering specs, so they describe how a network was meant to behave and need rebuilding as it changes. Arborum Labs' model learns from the network's own data and keeps updating as the network evolves. Most monitoring tools today tell operators that something is wrong; this one aims to tell them why, and what they could do about it.


Yann LeCun has been making a similar argument for years. He says language models learn from text, which is a thin record of reality, and so struggle to reason, plan or predict consequences. He felt strongly enough about it to leave Meta and start AMI Labs, which raised a $1.03 billion seed round in early 2026 to build world models.


In September 2026, AMD agreed to buy Fei-Fei Li's World Labs for about $8.2 billion. Most of that work, though, is about 3D scenes, robots and video. Very few teams are applying world models to power grids and telecom networks, and that's the space Arborum Labs has chosen.


Already in the field

Arborum Labs is further along than we'd expect at this stage. In telecom, it works with one operator and two equipment makers, and reports about 90% accuracy and improving on hard root-cause cases and more than 100 complex workflows automated. In power, back-tests suggest roughly ₹25 crore a year in savings for one large consumer and around ₹60 crore from better battery planning. The same core model sits under both, which gives us confidence it can stretch to other kinds of networks.


Why this is hard to replicate

It's fair to ask why equipment makers and operators wouldn't build this themselves. Part of the answer is people: teams that combine serious research in complex systems with years spent inside live networks are rare. The models also get better with each network they learn, so every new deployment is quicker than the last. Most real networks use equipment from several vendors, too, and any single vendor would struggle to offer something neutral that works across all of it. On top of that, infrastructure operators tend to stick with systems once they've come to trust them.


The founders

Himanshu Tyagi and Pooja Rajdev have spent their careers modelling complex systems, from the brain and pandemics to 5G networks and banks. 


Himanshu is a professor at IISc, Bangalore, where he works on information theory and machine learning. He previously co-founded Sentient and is an INSA Young Scientist.


Pooja has a PhD from Purdue. She has worked on seizure prediction and performance modelling at a US Army lab, built products at Optum and most recently led GenAI at BCG India.


In our conversations, two things came through clearly. The first is that this team can build frontier AI and keep pace with global peers. Research in this area moves every few months, and they adapt to it quickly, helped by close ties with researchers, companies and industry stakeholders around the world that mean they often know about a development well before it becomes common knowledge. The second is how firmly their thinking is rooted in customers and the problems they deal with. The people we spoke to who had worked with them were consistently positive about both.


Having both the research background and that field experience is unusual, and it's a big part of why we backed them.


What's next

Over the next decade, infrastructure will be built faster than ever, and we think AI that understands cause and effect will play a growing part in running it. Arborum Labs is building that in India, where networks are growing at a pace few countries can match, with customers abroad in mind too.


This round will help the team build the first versions of its world model for power and telecom and put them live with customers in India and abroad.


At Speciale Invest, we are the first institutional partners for science-led founders working on problems that will matter for a long time, and Arborum Labs fits that well. It's also the kind of AI company we're keenest to back: one building core technology for sectors AI has barely reached so far. And we think it matters that it's being built here. The systems that help run India's grids, telecom networks and logistics should be developed in India, trained on Indian networks and owned locally.


We're thrilled to back Himanshu, Pooja and the Arborum Labs team from day one. We're also grateful to Ideaspring Capital and the strategic angels who joined the round.


If you run a power or telecom network, or you'd like to help build a new kind of foundation model, write to hello@arborum.ai.





 
 
bottom of page