AI Summary
→ WHAT IT COVERS This 99-minute episode of This Week in Startups features AstroForge CEO Matt Gialich detailing plans to mine platinum group metals from near-Earth asteroids using $10.4M spacecraft, Templar's Sam Dare explaining how his BitTensor subnet trained a 72-billion-parameter AI model for $2–3M through decentralized compute, and a demo of OpenNotes, an open-source real-time meeting intelligence tool. → KEY INSIGHTS - **Asteroid Mining Economics:** AstroForge's Deep Space 2 spacecraft costs $10.4M all-in including launch on a Falcon 9, yet can return up to 1,000kg of platinum group metals worth approximately $105M per mission. That 10:1 return ratio means the business model breaks even if even one in ten missions succeeds. The key cost lever is dramatically reducing data transmission rates — communicating at just 400 bits per second at the asteroid keeps hardware and operational costs minimal. - **Platinum Group Metal Market Scale:** Platinum group metals represent the second-largest commodity market cap in the world after gold, and are embedded in LCD screens, OLED displays, cancer therapies, catalytic converters, and iridium crucibles used to manufacture silicon chips. The US holds no domestic reserves, with Russia and South Africa controlling global supply. AstroForge calculates it would take approximately 100 successful mining missions before returned supply would meaningfully suppress market prices. - **Near-Earth Asteroid Abundance:** The number of cataloged near-Earth asteroids has grown from roughly seven at the turn of the century to over 600,000 today. AstroForge targets metallic asteroids approximately 200 meters in diameter, composed of roughly 70% iron-nickel, which enables magnetic docking rather than physical landing. The target asteroid for Deep Space 2 sits approximately 10 million miles from Earth, making round-trip missions far shorter than asteroid belt concepts. - **Decentralized AI Training Cost Breakthrough:** Templar's BitTensor Subnet 3 trained a 72-billion-parameter model — comparable to Llama 2 in capability — for approximately $2–3M by distributing gradient computation across competing miners worldwide. Miners earn subnet tokens by reducing training loss the most each epoch, creating a self-organizing competitive incentive. The model is still roughly 300 times below frontier performance, but the team scaled from 1.2 billion to 72 billion parameters within nine months. - **Decentralized Compute Architecture Shift:** Templar's next training run will deploy a new algorithm called Heterogeneous Sparse Logo, which removes the requirement that every node hold the full model in memory. This allows training across mixed hardware — consumer GPUs, older A100s, and data center B200s simultaneously — targeting any machine with an internet connection and a free GPU. This mirrors the SETI@home distributed computing model and could dramatically lower the cost floor for AI pretraining globally. - **Open-Source AI Tooling Speed:** OpenNotes creator Yazan Ali Rahim built a fully functional Mac meeting intelligence app in Swift — a language he had never written — in approximately one day using Claude Code. The app transcribes locally via Parakeet, surfaces real-time insights from a connected knowledge base, and fact-checks speaker claims mid-conversation. He then deployed an AI auto-maintainer bot that independently triages GitHub issues, replicates bugs, opens pull requests, and merges fixes without human involvement. - **Streaming IP Economics at Scale:** HBO Max's Harry Potter reboot targets one book per season at a reported $100M per episode. Across seven seasons of eight episodes each, total production cost reaches approximately $5.6B. The math works if the franchise converts just 4% of HBO Max's existing 125M subscribers into retained viewers, or captures a fraction of the estimated 100M people who purchased all seven books — each representing roughly $1,000 in lifetime subscription value over a six-year run. → NOTABLE MOMENT During the AstroForge segment, Matt Gialich reframed what looks like a communication failure on the Odin mission. He explained that communication is always the symptom, never the cause — the actual problem was a power failure preventing radio activation. The solar panels failed to deploy, leaving the craft unable to generate sufficient power, yet navigation systems were still validated throughout the mission. 💼 SPONSORS [{"name": "Deel", "url": "https://deel.com/twist"}, {"name": "Circle", "url": "https://circle.so/twist"}, {"name": "Whisperflow", "url": "https://wisprflow.ai/twist"}, {"name": "Plaud", "url": "https://plaud.ai/twist"}] 🏷️ Asteroid Mining, Platinum Group Metals, Decentralized AI Training, BitTensor, Open-Source Software, Space Startups, Streaming Economics