02 / 04 GIX · Jun 2026
A spot market for GPU inference.
GPU Inference Exchange: a decentralised spot-market exchange where owners of idle GPU capacity post asks, workloads post bids, and the market clears at the crossing. Second place at the Encode Hackathon.
- Hackathon: [Encode listing url]
01 — Problem
Idle GPUs on one side, queued jobs on the other.
Inference capacity is bought on long contracts and sits idle for much of the day, while small workloads queue or overpay for on-demand instances. A spot market lets both sides post a price and clears where they meet.
[Why decentralised: who holds the book, how settlement works, what the hackathon brief asked for.]
02 — Method
Post, match, clear, settle.
Order book. Asks carry a price per GPU-hour and a quantity; bids the same. The book is sorted, bids descending and asks ascending, and cleared at the largest quantity where a bid still meets an ask.
[Matching and settlement: how a matched job is dispatched to the GPU, how completion is verified, how payment moves.]
[What was built in the hackathon window and what was mocked.]
Fig. III bEight jobs to six GPUs: a greedy first-fit pass, then the optimal assignment by the Hungarian method. Both costs measured.
03 — Results
What it did on the day.
[One paragraph on the demo and the judges' feedback.]
04 — Stack
What it is made of.
System design6 layers, 11 flows. Arrows are the data path; dashed ones are side channels. Hover a layer.