Bittensor's Compute Subnets Are Racing to Make GPUs Cheaper, Safer, and Less Idle

If there's a theme running through Bittensor's compute-focused subnets this month, it's that "cheap GPUs," "trustworthy GPUs," and "GPUs that would otherwise sit idle" are three separate problems — and three different subnets are each picking one to solve.

Chutes: pre-training without a cluster

Chutes — currently the network's largest subnet by market cap — used its own compute market to pre-train a 20-billion-parameter mixture-of-experts model for under $10 an hour, according to a July 20 post from the team. No dedicated cluster: eight rented single-L40S VMs spread across two continents, topped up with a handful of consumer RTX 4090s and a 5090, at roughly six seconds per training step. Pre-training is usually the workload people assume needs a real data center — Chutes' point was that its serverless, incentive-driven compute market did it anyway.

Targon: compute that doesn't have to trust its host

Targon, built by Manifold Labs, is solving a different piece of the same puzzle: if subnets are going to rent compute from strangers, the workloads running on it need to survive a host that can't be trusted. Targon's answer, built with Intel, is the Targon Virtual Machine — confidential, trusted execution environments (TEEs) that let developers run models on hardware they don't control without exposing the model itself to whoever owns the GPU underneath.

Macrocosmos / IOTA: nothing idle, nothing wasted

Macrocosmos, which runs Subnet 9 (IOTA), is going after a third angle — compute that would otherwise sit unused entirely. Its training architecture is built to stitch stranded GPU capacity into what the team calls "liquid training capacity," rather than assuming the answer is simply installing more hardware.

Put together: one subnet proving compute can be radically cheap, one proving it can be trusted without trusting the host, one proving it doesn't need to be idle in the first place. Different bets, same underlying resource.