Lium Flips Chutes as Bittensor’s Top Subnet While Scaling High-Speed GPU Rentals

Lium, Bittensor’s Subnet 51, has published a benchmark that puts a concrete number on one of the network’s biggest infrastructure bets. It is showing what cheap, permissionless access to high-end GPUs can look like in practice.

On September 6, Lium reported 14,499 output tokens per second running Qwen3.6-35B-A3B-FP8 on a single rented NVIDIA B200.

The benchmark used vLLM 0.28.0, 256 concurrent streams, and 512 output tokens per stream. Lium calculated the GPU cost at $0.107 per million output tokens.

That is significantly below the $0.70 per million output tokens currently listed by OpenRouter for Qwen3.6-35B-A3B.

Lium then published the benchmark setup for teams to use and test themselves. The repository contains 36 runs across five GPU classes, including B200, H200, H100, RTX PRO 6000 Blackwell, and RTX 5090.

We just ran Qwen3.6 35B at 14,499 tokens per second.

on 1 lium GPU.

85% cheaper than Openrouter.

how you can do it too ⬇️ pic.twitter.com/pKJPOJeBCP

— lium.io (@lium_io) September 6, 2026

Where Lium Sits in the Field

The GPU market Lium is entering is already crowded, but the products are not identical. Against managed APIs like OpenRouter, Together, and Fireworks, it trades one-HTTP-call convenience for full control of quantization, engine version, and batching. Against raw GPU markets, the peer set looks like this:

Provider	Market	Main advantage

Lium SN51

Decentralized GPU rental

Specific GPUs, open serving stack, permissionless supply

Vast.ai

P2P GPU marketplace

Broad marketplace and aggressive pricing

RunPod

GPU cloud

Easier deployment and productized infrastructure

AWS

Hyperscale cloud

Enterprise reliability and ecosystem

Chutes SN64

Decentralized inference

Buy model inference directly, without managing GPUs

The Bittensor products in this space are complementary. Chutes (SN64) provides serverless inference endpoints, where users pay for tokens. Targon (SN4) provides confidential compute using TEEs for companies that cannot run sensitive workloads on public providers. Lium (SN51), meanwhile, provides the GPU itself. It sells the GPU-hours that produce those tokens, and other subnets can use that compute as an input.

Lium Is Building a Profitable Business

The benchmark is only part of the Lium story. The subnet has also been showing strong growth on the business side.

In its latest monthly review, Lium reported $964,000 billed to 1,187 renters, up 36% month over month. It added 1,908 new signups, retained around 80% of users, and recorded 7,697 rentals with a median rental time of 1.15 hours. Around 63% of rentals were initiated programmatically by agents.

Lium also announced a $1 million burn of SN51 alpha tokens, funded with roughly five weeks of product revenue.

The growth is now reflected in the subnet’s position within Bittensor. Lium has overtaken Chutes to become the #1 subnet by both price and market cap.

Why This Matters

Lium now has two numbers worth watching. The benchmark shows what its infrastructure can deliver, while its revenue, renter growth and token burn show that people are using the product.

The 14,499 TPS result is not a standalone story about GPU speed. It sits alongside a growing rental business, increasing agent usage, and a $1 million token burn funded by product revenue.

For Bittensor, that makes Lium one of the clearest examples of a subnet building both infrastructure and a business around it. Lium is a subnet to watch closely.