Kamath: Founder bootstrapped Smallest AI with $150K of his own money
Sudarshan Kamath · Where SMALL models will Win | Sudarshan kamath, Smallest ai · Mar 6, 2026 · at 8:52
Sudarshan Kamath, founder of Smallest AI, explains how he funded the early compute and model training for his startup.
0:00 / 0:07exact quote · 7.0s
▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark
“I put the first 101 50 K into smallest of my own. And we basically got like some GPUs and we started like training models.”
quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →
More from Sudarshan Kamath
Kamath: Cartesia is the only other player focused on real-time voice
“Right now, the only other player who's focused on real time is Cartesia.”
Kamath: JEPA architectures are closer to AGI than state space models
“In our opinion, like those models are actually closer to how we can build AGI than like state space models, which is just a, Play on compute capacity.”
Kamath: Future AI will decouple infinite memory from small-model compute
“The way we see the future is you will separate memory from intelligence. So the memory will be captured by infinite layers and the intelligence will be captured by finite layers of, you know, intelligence and that small model that has those finite layers that …”
Kamath: Voice AI will diverge and is definitely not winner-takes-all
“I think the markets will diverge at some point of time. Like so for example, 11 is sort of doubling down on the content space but the real time space, like mimicking how humans do conversations is, it's like a very different problem than mimicking how movies a…”
Kamath: Smallest AI has enough capital to reach $100M ARR
“Having said that, I still think the capital we have right now is enough to build a hundred million ARR company.”
Kamath: Nvidia GPUs suffer severe memory bottlenecks during voice inference
“And media chips are generally very bad for voice inference. Like if you have 40 gigabytes of RAM in a NVIDIA GPU, you can generally, if you want to do real time, like a hundred milliseconds, you can use maybe two or four gigabytes of it for our models. And bey…”