Mar 6, 2026 · 1h 9m · neon-show
Where SMALL models will Win | Sudarshan kamath, Smallest ai
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Sudarshan Kamath, founder of Smallest AI, breaks down how his company built ultra-low-latency real-time voice models using specialized small-model architectures. He shares their rapid progression from a viral bedroom prototype in Bangalore to securing Silicon Valley backing and landing multimillion-dollar enterprise contracts.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
Siddhartha as informed peer 4.0 Guest teaching 5.4 Guest disagreement 1.7 Siddhartha pushing back 1.5
speaking balance: gold is Siddhartha, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 36:20 Rejecting application-level competitor categorization
Sudarshan flatly rejects the host's classification of voice application startups like Bolna as competitors, clarifying that Smallest AI powers them as infrastructure.
Hardest push from Siddhartha ▶ 42:00 Challenging the bespoke support cost of on-premise deploymentsSiddhartha directly pushes back against Sudarshan's preference for on-prem deployments, arguing it imposes heavy, non-scalable bespoke support overhead.
Biggest teaching moment ▶ 59:40 Explaining hardware memory-bandwidth bottlenecks in voice inferenceSudarshan delivers a technical breakdown of why NVIDIA GPUs bottleneck real-time voice due to physical distance between compute cores and memory.
Siddhartha holds their own ▶ 58:30 Highlighting OpenAI margin vulnerability and AI bubble risksSiddhartha challenges AI growth assumptions by highlighting OpenAI's unknown gross margins and the systemic risk if frontier capital drying up bursts the ecosystem.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Siddhartha as informed peer | Guest teaching | Guest disagreement | Siddhartha pushing back | Why |
|---|---|---|---|---|---|---|
| Trailer: The Multi-Trillion Dollar Vision of Smallest AI | 4 | 5 | 1 | 1 | The host opens warmly by expressing regret over missing out on investing in Smallest AI and sets up the narrative of breaking Indian ecosystem ceilings. Sudarshan delivers an extensive monologue recounting his background in physics, computer vision, self-driving cars with Arjun Jain, and the power of model distillation. | |
| Domain Inception, Legal Tech Experiments, and Going Full-Time | 3 | 4 | 1 | 1 | Siddhartha asks clarifying chronological questions about the founding timeline. Sudarshan explains his experiments with legal tech in India, early vector database adoption with Pinecone, context engineering, and the conviction to commit personal capital full-time. | |
| Discovering Real-Time Voice AI and the Viral LinkedIn Demo | 4 | 4 | 2 | 1 | The host tracks the fundraising timeline and seed VC metrics. Sudarshan humorously details the viral LinkedIn demo running off a single GPU and pitching VCs in Bangalore with a torn bag and shirt. | |
| Flipping to Delaware: Building a Global Deep-Tech Enterprise | 3 | 5 | 2 | 1 | The host inquires about banking and money transfers. Sudarshan articulates a decisive stance on flipping immediately to a Delaware C-corp to access global capital, talent, and more ambitious frontier AI problems. | |
| Unconventional Hiring and Moving to Silicon Valley | 3 | 4 | 1 | 1 | Siddhartha asks about round extensions and relocation. Sudarshan outlines his non-traditional hiring philosophy via Twitter/LinkedIn and the density of ambition in Silicon Valley compared to other global tech hubs. | |
| Building the Voice AI Infrastructure Portfolio and Pre-Revenue Seed Pitch | 4 | 6 | 2 | 2 | The host explores traction and initial validation metrics. Sudarshan details the full voice AI infrastructure stack needed to support sub-100ms TTS agents, including synthetic data generation and separating compute from infinite memory. | |
| Securing Seed Funding with Sierra Ventures and Enterprise Access | 4 | 4 | 1 | 1 | Siddhartha shares peer anecdotes about Sierra Ventures' dinner-closing style. Sudarshan describes securing their seed lead Tim Gullery and the immediate high-caliber enterprise introductions that followed. | |
| Closing the First Seven-Figure Multi-Year Telephony Contract | 4 | 5 | 2 | 2 | The host probes the contract timeline and conversion to a seven-figure deal. Sudarshan outlines navigating technical grilling by an enterprise CTO and moving from pilot to a multi-year enterprise contract. | |
| Preemptive Series A with Seligman Ventures and Passing on M&A | 4 | 5 | 1 | 1 | The host asks about acquisition interest and secondary sales. Sudarshan explains passing on a $150M+ acquisition offer to close a preemptive Series A with Seligman Ventures and deferring personal secondaries until sales repeatability is proven. | |
| Competitive Landscape: Differentiating Real-Time Voice from ElevenLabs and Cartesia | 4 | 6 | 3 | 2 | Siddhartha pushes on competitors like ElevenLabs and Bolna. Sudarshan firmly distinguishes applications from model providers and clarifies that ElevenLabs focuses on offline content generation while Smallest AI focuses on real-time conversational low-latency models. | |
| Architectural Deep Dive: State Space Models vs. Transformers and JEPA | 5 | 7 | 2 | 2 | The host asks for a layman explanation of architectural approaches. Sudarshan breaks down quadratic vs. subquadratic computation in SSMs and why Smallest AI favors predictive world-model architectures inspired by Yann LeCun's JEPA. | |
| Virtual Private Cloud Deployments and Scaling to $2M ARR | 5 | 6 | 2 | 2 | Siddhartha challenges the operational burden of bespoke on-prem deployments. Sudarshan clarifies that deliveries are containerized in client VPCs, offering 100% software margins and eliminating the GPU autoscaling forecasting headache. | |
| Horizontal Platform Positioning and System Integrator Partnerships | 4 | 5 | 1 | 1 | Siddhartha explores enterprise application builds and partner enablement. Sudarshan explains their horizontal platform strategy and why systems integrators like Accenture and Capgemini are better suited for verticalized enterprise implementations. | |
| Core Research Frontiers: Asynchronous Thinking and Infinite Memory | 4 | 7 | 2 | 1 | The host asks what the doubled research team will focus on. Sudarshan gives an educational breakdown of asynchronous thinking in voice interactions, continuous learning, and separating compute intelligence from infinite memory layers. | |
| The Divergence of LLM Context Scaling and Conversational Intelligence | 5 | 6 | 2 | 2 | Siddhartha presses on why major frontier labs like OpenAI and Anthropic wouldn't capture the entire voice market. Sudarshan uses an airplane-versus-bird analogy to explain that giant context-scaling LLMs represent a different branch of intelligence than agile, low-latency conversational models. | |
| Mathematical Playbook for Scaling from $1M to $100M ARR | 4 | 5 | 1 | 2 | The host asks about capital and timeline requirements to hit $100M ARR. Sudarshan breaks down a sales capacity mathematical model (reps doing $1M-$30M ARR combined with cloud and SI channel partnerships). | |
| Bottom-Up Developer Adoption and Horizontal Voice Use Cases | 4 | 5 | 2 | 2 | The host questions whether enterprise AI adoption is getting clogged by top-down vendor motions. Sudarshan argues that bottom-up developer self-serve adoption creates organic internal pull, citing Cursor as an example. | |
| Evaluating AI Bubble Risks and Hardware Bottlenecks in Voice Inference | 5 | 7 | 3 | 3 | Siddhartha presses on AI bubble risks and OpenAI's unverified gross margins. Sudarshan argues that voice modernizes existing utility markets rather than inventing speculative ones, and explains the physical memory bottleneck in standard NVIDIA GPUs for real-time voice inference. | |
| The Palantir-Style Sales Strategy and Deep Custom Kernel Engineering | 4 | 6 | 1 | 1 | The host asks how large accounts were cracked in detail. Sudarshan shares his Palantir-style account targeting playbook, winning over skeptical public-company CTOs through rapid two-day feature turnarounds, and writing custom NVIDIA kernels. |
Statements from this episode (41)
Kamath: Byju's acquired Toppr for $150 million
“We got acquired by juice for like one, fifty million dollars back then.”
Kamath: Distilled AI models match large model accuracy at a fraction of size
“When you use that pre-training model as a teacher for this small model, and we could go into what teacher means, but it would basically Act at similar accuracies as the larger model at 100 or one 10th of the size.”
Kamath: Legal tech startups must target the US market to survive
“If you have to build for the legal market, you have to build from the U S basically, or for the U S at least, right? Every, everywhere else is the market is just too tiny for you to get any money out of that.”
Kamath: Founder bootstrapped Smallest AI with $150K of his own money
“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.”
Kamath: Smallest AI built sub-200ms real-time TTS before ElevenLabs
“11 Labs was completely into offline dubbing and things like that. They did not have a real time model, right? So we were one of the first people to build a real time text-to-speech that ran under 200 milliseconds.”
Kamath: Smallest AI raised initial pre-seed checks at a $3M valuation
“Upsparks was like, not much like a hundred K if I am not wrong. Better was another two 50 or so. The total round was three 50 K and then couple of angels of 400 or something like that. This is at like a three million post money sort of around, right?”
Kamath: Smallest AI reduced 3one4 Capital's $1M check to prevent dilution
“We asked them to reduce the check size. We said, we can't take one at six. It's like too much dilution for us. So we, I think we ended up at like some 500 or six or something.”
Kamath: Indian language AI is an operational dataset problem, not scientific
“Indian languages for me was like just an operational problem to solve getting data sets, et cetera. But it was not a scientific enough problem that helps us answer what's like the next level of AI and that how can small models be the future, et cetera.”
Kamath: Smallest AI raised $900K extension, totaling $1.8M
“We raised another 900 K extension at a higher valuation. And we had like 1.8 of which most of it was in the bank.”
Kamath: Smallest AI cut TTS latency to sub-100ms by seed round
“So we had a text to speech model which basically you could use to give natural realistic voice to the voice agent in under a hundred milliseconds. So initially it was at 200 milliseconds. It came down to a hundred milliseconds by the time our seed round was th…”
Kamath: Smallest AI is working with ServiceNow
“For example service now, for example, Right. We are working with them.”
Kamath: Smallest AI signed a seven-figure US telecommunications enterprise deal
“It is converted to a seven figure multi-year multimillion deal. But we will announce it formally sometime like in April or so once we sort of clear a few things, but yeah, it is that like all of those things are in paper right now.”
Kamath: Only ~20 Indian companies can offer seven-figure enterprise deals
“There are probably 20 companies in India who can give us that kind of a deal size and they would spend an incredible amount of time and effort.”
Kamath: Smallest AI closed Series A round with Seligman Ventures
“We closed the series A, we could call it a pre-series A, but like we closed with the Seligman Ventures.”
Kamath: Smallest AI received an acquisition offer over $150M
“We got in between like an acquisition offer for like one 50 plus million dollars.”
Kamath: AWS is an active reseller of Smallest AI
“AWS is an active reseller of smallest by the way.”
Kamath: ElevenLabs is focused on content creation and dubbing, not real-time
“Like for example, 11 has a lot of focus on content creation and like dubbing. And if you want to build like an Instagram video or like a AI movie 11 would be a great player. We don't focus on content at all. We are focused on the real time space.”
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: Cartesia relies on state space models, Smallest AI does not
“But they are heavily dependent on state space models and we don't train state space models.”
Kamath: State space models are not a big leap over transformers
“Our strong belief is like state space models are not like that big a leap over transformers.”
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: Smallest AI is building a predictive architecture for voice
“We are trying to build like a predictive view of voice and like conversations that we can have so that you can guess what the answer should be rather than How LLMs typically operate.”
Kamath: Smallest AI achieves 100% margins via customer-managed VPC deployments
“We recommend it because for us it's a hundred percent margins, right? If that happens.”
Kamath: AI vendors overprice models to offset GPU auto-scaling costs
“And the biggest issue is today people end up pricing the models extremely expensively because they have to manage the scale up of GP capacity.”
Kamath: AI research hires take six months to show ROI
“Research gives you ROI in like six months. So you can't like hire someone right now and then expect, if you have goals for 2026, like in September, you have to get those people right now.”
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 raised pre-seed funding from OpenAI personnel
“So some of our investors in like our pre-seed words from OpenAI as I spoke to them, they are building a different kind of intelligence.”
Kamath: Scaling LLMs has diverged from building human-like AI
“Very similarly, right now in AI, we are doing LLMs, which is a different form of intelligence than human intelligence. So they have something and they are playing the sort of let's make LLMs bigger and bigger and smarter. And that is a market as well. But To b…”
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: Indian financial firm runs 100K+ daily debt calls on Smallest AI
“There is a financial services company based in India, which is doing large scale debt collections using us, like, you know, every day we are like at a 100,000 calls plus using that.”
Kamath: US telephony firm records internal meetings with Smallest AI
“There is a telephony company that has all of their like meetings internal meetings, getting like recorded in the US using our voice models.”
Kamath: Most coding agent startups will die or face fire sales
“There are a lot of companies that are just trying to build these like very specific coding agents, which apparently will be like better than that's not going to happen. Right. So I feel like there are like, Few winners in that space. And like the others are ju…”
Kamath: Voice AI expands established markets instead of inventing new ones
“The voice space is a existing market that is getting disrupted. You're not inventing a market. You already have a dictation in mobile apps, which sucks. You already have that desktop note, which was really bad. Now with generative AI, you have much better dict…”
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…”
Kamath: Voice AI inference currently costs 1,000x more than text
“Today, the cost of voice is like a thousand times more than the cost of text. We want to like bring it at par basically.”
Kamath: Smallest AI targets top 50 accounts via Palantir-style sales
“We are literally creating the like list of top 50 dream customers that we have and going aggressively through like a top down motion, a very targeted motion of how could we, you know, get them through a Palantir style approach.”
Kamath: Smallest AI's Pulse speech-to-text operates at 64ms latency
“So we have this amazing voice model that converts voice to text called pulse. It's a speech to text model runs in real time 64 milliseconds latency lot of other features like emotion detection, gender detection, bunch of other things, right?”