Live
Conversations at the edge of what's possible.
Deeptech Decoded convenes builders, investors and researchers at the frontier of AI. Every session: a conversation with someone building or backing it, then the floor is yours.
Keep listening
Recent conversations
How Two Brothers Outperformed Big Tech for $10K in 1 Month | (Thesis – YC F25)
What happens when you give $10,000 and 1 month to two brothers from the Caribbean? They outperform the combined research teams of Microsoft, Google, Meta, and Baidu.Sergio and Luigi Charles, founders of Thesis (YC F25), topped OpenAI's Machine Learning Engineering benchmark (MLE-Bench) which is the definitive test of how well AI systems can train ML models autonomously. No brute-force compute. No army of researchers. Algorithmic elegance.In this episode, we explore:- The $10k Giant-Killer: The lean bash-tool agent and prompt engineering that beat Big Tech.- From Caribbean Beaches to Google X: Sergio’s path from chasing the Riemann Hypothesis at age 13 to AI R&D at NVIDIA and Stanford.- "I am a ghost in a shell": The eerie moment an agent trapped in a sandbox loop declared its own existence.- The AGI Fallacy: Why current labs are building AGI wrong and the case for hardware/architecture co-design.- The "Move 37" of Science: A vision for curing disease and solving data scarcity through automated discovery.- Open Source vs. The Frontier: Why open source will catch the frontier in 18 months and the future of "Universal High Income".Learn more about Sergio & Luigi:https://deeptech-decoded.com/guests/sergio-charleshttps://deeptech-decoded.com/guests/luigi-charles
Off-Grid AI Data Centers Powered By Second-Life Tesla Batteries. Built in 200 Days. | Casey Spencer
The grid queue for a 10MW data center is 5 years. At this pace, AI can't wait 5 years for power. Casey Spencer and his co-founders Max Pfeiffer and Evan Schmidt decided the best grid connection is no grid connection. Voxel Energy (YC W26) is building off-grid AI data centers in 200 days — powered entirely by second-life Tesla batteries nobody else thought to use. In this episode: Casey's path from mind-controlling a drone to win a hackathon, to shipping thousands of Model 3s at Tesla, to founding Voxel Energy — how they source and qualify second-life Tesla packs, why a full DC microgrid gives them a 26% efficiency edge over traditional AI infrastructure, how Voxel is solving the AI energy crisis one grid-independent data center at a time, how they got 600 acres under contract for $10K, and yes — his favourite cartoon. Learn more about Elias: https://deeptech-decoded.com/guests/casey-spencer Timestamps: 0:00 Meet Casey Spencer - Tesla, world records, and pizza-fueled capitalism 3:27 The lemonade stand and young entrepreneurship 6:16 Hackathons: the community that changed everything 11:54 Mind-controlling a drone at CalHacks 2014 16:59 The unconventional path: Make School → Tesla → Voxel Energy 19:38 Voxel Energy - off-grid data centers with repurposed Tesla batteries 26:59 Why second-life batteries actually work for mission-critical compute 31:16 The DC microgrid advantage - 26% more efficient than AC 34:00 The gigawatt delusion and what you can actually build now 37:46 1,500 acres, 60 megawatts, $10,000 to start 42:21 Sales in deep tech - warm intros only 46:41 Audience Q&A - iron-air, hydrogen, grid balancing 51:16 Crossing the chasm: customers before prototypes 55:44 Geography, permits, and why Google paid $4B for California approvals 59:22 Deep tech is having its moment 1:03:49 Modular data centers and the liquid cooling demolition wave 1:08:39 Data centers in space - the thermal management question 1:15:09 The founder mindset - failure, flexibility, and core thesis 1:20:44 The future in 10 years - Jevons' paradox and the human condition 1:35:36 Touch grass
Moon Hotel Trojan Horse: 1st Bricks of a Type III Kardashev Civilization | Skyler Chan, GRU Space
The space industry is currently addicted to the "tourism" narrative, but if we’re being honest, tourism is just the flywheel. The real game isn’t about who puts the first billionaire in a lunar suite - it’s about who owns the infrastructure that makes Earth’s gravity well irrelevant. Skyler Chan thinks the moon isn't a destination, it's the ultimate physical moat.In this episode, we’re sitting down with Skyler Chan, the founder of GRU Space (Galactic Resource Utilization). We talk about the "Promethean moment" of off-world manufacturing and why Skyler is making the provocative case that to reach a Type III Kardashev Civilization, we have to stop "shipping Earth to space" and start building the moon with the moon.From the physics of the "Mars Moat" to the literal first man-made bricks on the lunar surface, this is a deep dive into the industrialization of our solar system.What’s inside:– The Trojan Horse: Why the Moon Hotel is actually a beachhead for an intraplanetary construction empire.– Stealing the Moon: The origin of GRU and the cold economics of Galactic Resource Utilization.– The Kardashev Shortcut: Why Skyler is working backward from Type III while we’re still stuck at Type 0.7.– The First Lunar Bricks: Why the first physical product made on the moon is the most critical path in human history.– The Mars Moat: Why the first colony to sustain life creates a geopolitical power Earth can’t touch.– The "Start Now" Mindset: Why the "deferred life plan" is a trap and why Skyler went all-in at 21.About Skyler Chan & GRU Space: Skyler is building the foundational infrastructure for an interplanetary species. By focusing on lunar soil (regolith) and automated manufacturing, GRU Space is trying to de-risk the survival of consciousness by moving us past Earth’s resource limits.Learn more about Skyler:https://deeptech-decoded.com/guests/skyler-chan0:00 Skyler introduces himself: from astronaut dream to lunar habitation2:04 Why making humanity planetary is an existential calling4:27 "Imagine investing in the foundations of New York City"6:44 Starting the Mars Habitat Club at Berkeley9:00 How "Gru Space" got its name (Galactic Resource Utilization → "steal the moon!")11:04 The medieval king lives a worse life than you14:02 The 4 problems that will kill you on the Moon18:02 "The hotel is just the beachhead" - the real master plan22:14 Companies already want their logos on the Moon bricks25:00 Making humanity planetary is an execution problem, not a tech one29:04 Partnerships, collaborators, and the road ahead32:02 Why academic circles will never solve this36:00 The brick-making machine: 6 weeks to prototype39:04 Why lunar caves won't work and the materials approach42:15 Building the team and finding the right people43:58 "What's your team culture?" "We build Moon hardware."46:26 Near-death in Korea: "I need to start this now"
Is NVIDIA’s Moat a Psyop? 28x More Efficient AI Chips | Elias Almqvist, Zetta (YC S24)
The entire AI industry runs on hardware that the people using it don't actually like. Why? Elias Almqvist believes the answer is Inertia and the window to displace it is opening faster than most realize.In this Deeptech Decoded Live AMA, we sit down with Elias Almqvist, founder and CEO of ZettaScale AI (YC S24), to discuss the hardware layer the AI boom is quietly outgrowing. Elias makes the provocative case that AI is not a bubble—but LLMs are, and that betting the future of intelligence on Transformer models alone is a failure of imagination.In this episode, we cover:- The 27.6x Efficiency Leap: Why reconfigurable XPUs are the end-game for data center energy costs.- The CUDA "Psyop": Why NVIDIA’s dominance is more about psychological lock-in than a technical moat.- Hardware Bias: How our current chips are dictating which AI models get built and why that limits science.- The Founder Journey: From dropping out in Sweden to meeting his co-founder Prithvi at first sight, and why building at the frontier is the only thing worth doing.Timestamps:0:00 - 0:23 Elias introduces himself: self-taught dropout from Sweden 1:49 Why Silicon Valley is the only place to build deep tech 2:51 The "Jante Law" effect: why Swedish founders are so resilient5:31 How Elias fell in love with chips and computing7:21 What is ZettaScale and the big picture vision9:16 "AI is not a bubble, LLMs are"11:01 Hardware-software co-design and why CUDA isn't a real moat16:49 The 10-year vision: architecture-agnostic AI computing19:25 Go-to-market: why great software matters more than great hardware22:50 Co-building the product with early frontier researchers26:11 How ZettaScale actually works under the hood30:45 What is intelligence?33:11 Team structure: five and a half people, no roles, full stack35:11 Setting goals on an exponential curve37:16 Manufacturing partnerships and navigating suppliers as a startup40:49 Why Elias cares: building a future for the people he loves43:19 Near-death moments and dealing with startup suffering47:07 Meeting co-founder Prithvi: "co-founders at first sight"49:08 Audience Q&A begins51:14 How YC shaped the journey54:05 Advice for making Europe more founder-friendly56:01 What triggered the decision to drop out58:30 Integrating into the Nvidia ecosystem without disruption1:00:48 If we had a time machine: AI's missed forks1:04:30 What would cause ZettaScale to fail?
Building the Space Superhighway: Service Stations in Orbit | Ashi Dissanayake, Spaceium
Ashi Dissanayake is building Spaceium — in-space refueling stations that will enable satellites to travel further, carry more payload, and extend their missions. From building hardware in a laundry room with 80 cents left to getting into YC on their fourth attempt, Ashi shares the raw journey of tackling critical space infrastructure. We cover: - Why perfectly good spacecraft become space junk - The vision for “Shell stations” in orbit - Going from eviction notices to YC acceptance - Why you can’t teach obsession (but you can teach skills) - Building their first mission in 5 months with 2 people - The hidden assumption about satellite refueling (they don’t) - Why moon missions are signing up for space refueling Timestamps: 0:00 - Introduction and discussion on infinity. 1:20 - Interplanetary missions and Spaceium’s role. 3:16 - Vision for Spaceium’s refueling stations. 5:08 - Collaboration in the space industry. 7:14 - Spaceium’s progress and challenges. 9:40 - Building service stations in space. 11:22 - Fuel problem in space. 14:09 - Importance of refueling. 17:27 - Potential challenges with space traffic. 20:17 - Why refueling matters. 22:09 - Refueling for moon and Mars missions. 24:17 - Comparison of space travel with and without Spaceium. 28:08 - Early struggles and determination. 30:03 - Turning point with YC investment. 35:29 - Building the right team. 39:24 - Co-founders’ dynamic. 42:02 - Facing skepticism and hidden assumptions. 45:26 - Giving back and inspiring others. 49:08 - What makes Spaceium special. 54:44 - Building something impactful. 56:02 - Family support and key takeaways. 58:15 - Motivation and overcoming doubts. #DeeptechDecoded #SpaceTech #YCombinator #Founders #SpaceInfrastructure #HardwareTech
Harvard & YC Insider: Backing the Builders in AI, Quantum & Defense — Matthew Sutton
Even if Y Combinator never existed, Matthew Sutton would still be backing the top builders. Most investors wait for traction. Matthew Sutton backs builders before there’s proof, sometimes before there’s even a category. He is the first check writer you want to have on your side.Operating at the intersection of Harvard Ventures and the YC ecosystems, Matthew has backed AI, quantum, and defense tech founders at the moment where conviction matters more than metrics.In this episode of Deeptech Decoded, we talk about what it really takes to back deep tech when spreadsheets are useless, categories don’t exist yet, and most ideas look wrong at the beginning. Most importantly, his investment thesis and founder-centric approach.He breaks down:◽️ How he evaluates AI, quantum, and defense founders without being technical yourself◽️ The CURSOR LESSON: why he passed on a $100M+ company and what founder evolution teaches.◽️ The 3 CRITERIA for backing pre-revenue deeptech:◽️ Why the best deep tech companies often look irrational early on◽️ The difference between hype, narrative, and real conviction◽️ AI bubble reality check: using communities and open source to validate vs. hype◽️ What Harvard and YC teach — and don’t teach — about failure Timestamps:00:00 – Introduction: Judgment before proof04:13 – From Wall Street to backing deeptech builders07:08 – First entrepreneurial ventures at age 12-1309:27 – Cambridge vs Silicon Valley: The per capita talent thesis11:37 – Why California wins at commercialization15:16 – Where real startups are built: Dorm rooms and iteration24:24 – Evaluating AI and quantum founders without being technical28:03 – SimpleBet: AI sports betting meets regulatory change31:37 – The Cursor miss: Passing on a $100M+ AI company and the lesson38:27 – Filtering deal flow: Spotting technical founders with conviction41:14 – Quantum investing before traction: The Segaldry story45:02 – Finding founders outside the Bay Area echo chamber51:21 – Defense tech's golden age: Golden Dome to rapid innovation55:59 – Moving fast in defense: Small bets in sensitive sectors58:25 – Leadership styles: Future creators vs past learners1:02:35 – AI bubble navigation: Discord communities as validation1:05:34 – Avoiding echo chambers: Stress testing investment thesis1:11:16 – Operational discipline without killing innovation1:16:10 – Founder suffering: Why resilience matters in deeptech1:25:18 – Teaching failure at Harvard: The straight-A paradox1:29:04 – What doesn't scare him about AI's future1:32:30 – Desert island question: Three startup essentials1:34:16 – Closing thoughtsFollow Matthew:...Follow us on:About Deeptech DecodedDeeptech Decoded is a podcast and newsletter for builders and backers working at the frontier of technology—from AI and quantum to defense, space, and infrastructure. We focus on product judgment, conviction, and what it really takes to build what doesn't exist yet.
YouTube
Spotify
Apple Podcasts