Lium served 7,549 GPU rentals across 900 paying accounts in August 2026, and by 11 September 2026 the marketplace had 470 nodes online carrying 1,329 GPUs between them, real usage for a company that only took the Lium name in mid-2025. The five sections below cover how the marketplace works, what its latest numbers show, how providers, renters, and independent verifiers keep it honest, what protects a rented job once it is running, and what the growth signals for GPU rental more broadly.

Lium is a GPU marketplace built by Datura AI. It launched in late 2024 as Celium and became Lium in mid-2025. Instead of one company owning every card, independent providers around the world list their own hardware, and anyone can rent high-performance GPUs from that pool on demand. The catalog is wide: 21 different GPU models were rented in August 2026 alone, spanning everything from RTX 4090s up to H100, H200, B200, and B300 cards. Each provider sets its own hourly price, and billing runs per second at that rate, with no minimum, so a renter never pays for idle time between jobs. Early traction is one thing; sustained usage is another. Lium.io logged 7,549 rentals across 900 paying accounts in August 2026 alone. By 11 September 2026, the marketplace had 470 nodes online carrying 1,329 GPUs between them. Those are usage numbers, not traffic numbers: real renters paying for real jobs, spread across hundreds of independent providers rather than one company's fleet. Three roles make the marketplace work, each with a different job: - Providers list GPU hardware and get paid for idle hours, a verified, listed card earns even while nobody is renting it, plus 95% of a rental's revenue once it is rented. - Renters sign up, top up, paste an SSH key, pick a Docker template, and deploy; a pod is up in minutes. - Independent verifiers probe every machine over SSH before it is listed and keep re-checking it afterward, so a renter never has to take a listing's word for the hardware being what it claims. Hashrate's own rundown of renting and hosting GPUs already covers a handful of platforms working through some version of this same split, between the person who owns the hardware and the person who wants to use it. Verification before listing only covers the hardware, not the workload that runs on it afterward. Some Lium nodes close that second gap with a confidential-VM option built on Intel TDX, so the machine's owner cannot read what the renter is running on it. There is also a financial incentive to stay reliable. A provider who drops a live rental is penalized, and the renter is credited for the disruption. New hardware joins quickly too: a provider's node is typically listed about 20 minutes after it is added, per Lium's own documentation. The same trust problem is showing up at a much larger scale across the industry. Uptime Institute's own reporting on crypto mines turning into AI factories found operators like Applied Digital, Core Scientific, CoreWeave, Crusoe, and IREN converting old mining sites into AI infrastructure, largely on the strength of power access and industrial land they already controlled. That is the same idle-hardware economics that now pays independent GPU providers on marketplaces like Lium.io, just at a much bigger scale. Lium's numbers are still small next to the broader cloud GPU market. But a marketplace that went from a 2024 launch to nearly 1,330 GPUs and 900 paying accounts by September 2026, paying providers for idle time and 95% of every completed rental, is not a proof of concept anymore. Providers are getting paid reliably, too: every rental payout in the 30 days before 9 September 2026 landed within two days. Cloud vendors that keep treating independent GPU rental as a niche workaround are the ones most likely to get undercut by it once the model scales past its current size. By creating a direct, high-margin pipeline between underutilized hardware and price-sensitive developers, the platform is actively eroding the traditional cloud pricing moat. Smaller research labs and independent AI developers are finding that these decentralized networks offer comparable compute for a fraction of the cost, making the transition away from legacy providers increasingly painless. If these high-payout, low-friction settlement mechanics hold as supply scales, the shift from centralized data centers to distributed capacity will not just be an alternative; it will become the primary competitive baseline for affordable compute.