The views and opinions expressed in the interviews published on Made in CA are those of the interviewees and do not reflect the official policy or position of Made in CA.

The information provided through these interviews is for informational purposes only and does not constitute an endorsement or recommendation of any products, services, or individuals featured. We strongly encourage readers to consult with appropriate professionals or authorities in the relevant fields for accurate information and advice.

Jay Dawani

At Lemurian, we believe that AI will best serve humanity when it is available to everyone to use and benefit from. But the cost of AI systems has ballooned to the hundreds of millions, making it only accessible to the largest companies. To address this divide, we rethought the software stack from pytorch down to the existing hardware. Then, we created a more efficient way of representing numbers called PAL (parallel adaptive logarithms). It is around the needs of the software and the number format that we began designing our high-performance accelerator for AI training and inference. We call our accelerator an SPU (Spatial Processing Unit), and it is designed for higher utilization and, therefore, throughput while consuming a lot less power relative to its legacy counterparts.

Our mission fundamentally is to democratize AI with a scalable and affordable solution to the industry’s unprecedented growth and solve the hardware limitations currently stunting its progress.

Tell us about yourself?

I got introduced to programming when I was 12, and then one of my father’s friends introduced me to Norbert Wiener and Marvin Minsky’s works when I was 15, which got me into AI. And I have been fascinated by it ever since. I went to school to study applied mathematics because there wasn’t a program for AI then.

I got lucky that I was getting into the industry just as deep learning was starting to catch on, and I decided to focus on that. In the past 10 years, I’ve been fortunate to have worked on a diverse set of problems ranging from AI for material discovery, design optimization, and multi-physics simulation; autonomous cars and robots; space robots; quantum computing; web3 gaming at startups, research labs, and publicly traded companies. In 2018, I co-founded Lemurian Labs to build a foundation model for general-purpose autonomy and a platform for its continued deployment and domain adaptation. Some of the problems we encountered on that journey led to us pivoting to solve the AI compute crisis, which is holding back a lot of the high-potential applications of AI in various industries.

I have also served as an advisor to the NASA Frontier Development Lab and SiaClassic and authored ‘Mathematics for Deep Learning’ to lower the entry barrier to the field.

If you could go back in time a year or two, what piece of advice would you give yourself?

I could probably write a book on the things I have learned along the way that I wish I could give to my younger self as a cheat code, like in Back To The Future. The biggest advice would be: “No matter the effort, you cannot change the world in a day. Change takes time, so don’t worry about trying to get everything done immediately. Don’t be afraid of taking time off and taking care of yourself. And trust your instincts.”

What problem does your business solve?

The compute and energy requirements for training and deploying state-of-the-art AI has been growing at an unprecedented rate. It’s gotten to the point where only a few companies can afford to play in this sandbox. Given the amount of compute required to serve these models to end users and the performance available in accelerators, we can easily see a future not too distant where 20 per cent of the world’s total energy consumption is allocated to data centers, a 10 times increase from today.

To address this inequity, we’ve created a solution that can accelerate AI faster at 20 times throughput and one-tenth the cost of legacy GPUs. With this, we’re dramatically reducing the environmental impact of AI while rethinking the economics of AI at scale. Ultimately, we’re making it so that any company, regardless of market cap, can get into building their own AI models and create value for their customers.

What is the inspiration behind your business?

I’ve spent the majority of my career in artificial intelligence and robotics and have always gravitated toward solving significant complex problems in the industry.

In 2018, I co-founded Lemurian Labs with Vassil Dimitrov. We started with a focus on autonomous robotics, trying to build a foundational model to make it easier for robotics companies to transition from a mechanical-first mindset to an AI-first one. We soon realized that the amount of compute we would require to realize this vision was unattainable given the cost of compute.

This is a problem that still exists today. Despite the advances in the industry, the buy-in to become a contender in the generative AI space is 10,000 GPUs. The demand for larger models is already stressing our current hardware infrastructure to the point of fracture, and there’s a multi-year waiting list at some of the cloud providers for high-end GPUs. As a result, AI development requires exorbitant costs and excessive power consumption, resulting in environmental concerns. It’s entirely unsustainable.

Solving this requires fundamental changes to the way we think about hardware, and before you earn the right to do hardware, you really have to do software right. That’s exactly what we’re doing at Lemurian Labs and why we’ve taken a software-first approach. We can give significant throughput gains at much lower energy consumption while making it easier for developers to do their work.

What is your magic sauce?

There are three things: our software stack, our number format called PAL, and our SPU, which is based on a distributed dataflow architecture.

Our software stack is designed around the needs of AI developers to make it easier for them to get performance out of hardware and gain productivity. Our number format is around the needs of AI models, and it consumes less energy than equivalent floating point types while running faster. Our SPU was designed around the needs of the former two to give the highest possible throughput across a diverse set of workloads while consuming less energy than legacy accelerators.

Combined, they can provide a 20-times throughput benefit while reducing the total cost of ownership by nearly an order of magnitude.

What is the plan for the next 5 years? What do you want to achieve?

We’re thrilled to share Lemurian Labs has recently secured $9 million in seed funding, led by Oval Park Capital, with significant participation from investors Good Growth Capital, Raptor Group, and Streamlined Ventures, among others.

We’ve got a spec for our architecture, which we have developed in a simulation right now, and are focusing very heavily on our software stack, which can target heterogeneous clusters. We will be productizing our software stack ahead of having our accelerators available for customers.

We’re looking forward to continuing to grow our team and advancing our product to make it easier for everyone to build and deploy their own AI models.

What is the biggest challenge you’ve faced so far?

In 2018, I co-founded Lemurian Labs with Vassil Dimitrov. We started with a focus on autonomous robotics, trying to build a foundational model to make it easier for robotics companies to transition from a mechanical-first mindset to an AI-first one. We soon realized that the amount of compute we would require to realize this vision was unattainable given the cost of compute.

How can people get involved?

We’re in the midst of a growth phase and are always looking for exceptional talent to join our team. If you’re passionate about challenging the status quo and want to work on extremely challenging problems, we invite you to visit our website to explore the exciting job openings we have available.

To learn more about Lemurian Labs and our exciting job opportunities, visit our website at and follow us on LinkedIn @Lemurian Labs.