Training AI Models Without a Billion-Dollar Data Center | Steffen Cruz of Macrocosmos
Training a frontier AI model today requires hundreds of thousands of GPUs, months of compute time, and a budget that only a handful of companies on earth can afford. Steffen Cruz, co-founder and CTO of Macrocosmos, thinks that model is about to break, and he's spending his time building what comes next. His project IOTA, operating within the BitTensor blockchain ecosystem, uses distributed training to split large language models across thousands of devices located around the world, coordinated by blockchain, and powered by surplus cheap energy wherever it exists. After nine months of research, the system can reproduce baseline benchmark performance using what Cruz calls "wonky vegetables" - unreliable, churning, globally distributed compute - and turn it into something indistinguishable from centralized training if you use the right approach.
The conversation with Craig Smith covers the mechanics of how this actually works, why the blockchain's role is far narrower and more practical than most people assume, and why the Mac mini stockpiling trend creates an unexpected supply of distributed compute that can earn passive income when idle. Cruz's target: a 70 billion parameter model by mid-2025, trained at 10-20% of what it would cost through a hyperscaler, and aimed squarely at the legal firms, hospitals, and cash-strapped startups that have been waiting to train their own sovereign models but couldn't afford the price tag.
Subscribe to Eye on A.I. for weekly conversations with the people building and deploying the future of AI.
#340 Steffen Cruz: Training AI Without Data Centres
What if you could train a frontier AI model without building a single data centre?
In this episode of Eye on AI, Craig Smith sits down with Steffen Cruz, co-founder and CTO of Macrocosmos, to explore a radical alternative to the way AI models are built today. Instead of billion-dollar GPU warehouses, Steffen is training large language models using idle compute from devices distributed around the world, coordinated through the Bittensor blockchain.
Steffen breaks down why the centralised data centre model is heading toward a wall. Projects like Stargate and Colossus cost tens of billions of dollars, and as appetite for larger models grows, the economics simply stop making sense. He explains how distributed training flips this on its head, tapping into surplus energy, underutilised GPUs, and even consumer devices like Mac Minis to train models at a fraction of the cost.
We also get into IOTA, Macrocosmos's flagship technology, an orchestration layer that takes compute nodes scattered across the globe and makes them act like a single supercomputer. No single device runs the full model. Instead, each one carries a small slice, a technique called model parallelism, and together they can train frontier-scale models that would otherwise be out of reach for startups, researchers, and enterprises.
Finally, Steffen shares what he's building toward: 70 billion parameter models trained at 10 to 20 percent of centralised costs, a two-sided marketplace for compute, and a future where anyone with a spare GPU or Mac Mini can earn passive income while contributing to the democratisation of AI.
Subscribe for more conversations with the people building the future of AI and emerging technology.
Stay Updated:
Craig Smith on X: https://x.com/craigss
Eye on A.I. on X: https://x.com/EyeOn_AI
Timestamp:
(00:00) Introduction: The Problem With Blockchain AI Projects
(06:39) Meet Steffen Cruz: From Subatomic Physics to Decentralised AI
(09:16) What Is a Bittensor? The Blockchain Built for AI
(11:53) How the Blockchain Actually Works: Registry, Clock, and Rewards
(15:08) Why Data Centres Are Hitting a Wall
(22:01) Distributed Training vs Federated Learning: What's the Difference?
(27:47) Train at Home: Turning Your Mac Mini Into a Passive Income Machine
(32:49) IOTA Explained: Building a Global Supercomputer From Spare Parts
(39:43) How the Network Scales: From 256 Nodes to Limitless Compute
(44:39) The Road Ahead: 70B Parameter Models and the Future of Affordable A