Training AI Models Without a Billion-Dollar Data Center | Steffen Cruz of Macrocosmos

What It Covers

Steffen Cruz, CTO of Macrocosmos, explains how his company uses BitTensor's blockchain infrastructure to train large language models through distributed compute nodes worldwide, targeting 5,000 nodes by mid-2025 and 70-billion-parameter models as a commercial milestone for cost-arbitrage AI training.

Key Questions Answered

• •Distributed Pretraining Economics: Centralized data centers lock training costs into upfront capital expenditure, but distributed training enables real-time energy cost arbitrage. Macrocosmos targets surplus energy pockets — such as off-peak Icelandic power — to reduce pretraining costs to roughly 10–20% of conventional data center rates, making 70-billion-parameter model training accessible to cash-constrained startups and academic institutions.
• •Model Parallelism at Scale: Rather than running full model copies on each node, Macrocosmos deploys small model "slivers" across distributed machines using pipeline parallelism. This allows frontier-scale models to be trained from consumer-grade hardware like Mac minis and prosumer GPUs, with an orchestration layer resembling Kubernetes routing data between nodes to simulate a unified supercomputer.
• •Blockchain as Coordination Layer, Not Compute Layer: The blockchain in BitTensor serves three specific functions — identity registry, synchronization clock, and transparent payout trigger — while all actual compute and training data remain entirely off-chain. Understanding this separation helps evaluate any blockchain-AI project: the chain handles trust and compensation, not processing or storage.
• •Consumer Hardware as Passive Income Infrastructure: Macrocosmos's Train at Home program lets owners of idle Mac minis, MacBooks, or consumer GPUs contribute compute during unused hours and earn IOTA token payouts proportional to hours contributed. With 2,500 macOS app downloads in the first two weeks, the supply-side network can scale without capital expenditure by monetizing already-purchased personal devices.
• •Two-Sided Market for Underutilized GPU Inventory: NeoCloud and hyperscaler providers typically rent out 90–95% of GPU inventory, leaving gaps of two or more hours between bookings. Macrocosmos targets these interruptible idle windows, offering providers better margins than spot inference pricing while giving demand-side users — researchers, startups, enterprises — a PyTorch-compatible interface requiring no additional workflow changes.