AI firms rent strangers' supercomputers

A July 8 post by X user @0xndra pushed a niche infrastructure practice into public view: AI firms and researchers are renting compute from machines they do not own, including large GPU clusters described online as decentralized or third-party supercomputers. The post argued that rented capacity creates a verification problem if buyers cannot prove what hardware they actually received or whether a job really ran as claimed. Other services already market that arrangement directly, offering on-demand access to outside-owned GPUs for training and inference. Epoch AI researchers Konstantin Pilz, Robi Rahman, James Sanders and Lennart Heim gave the clearest reason the practice exists. In an April 2025 paper and companion publication, they said leading AI supercomputers have doubled in performance every nine months, while hardware cost and power needs have doubled every year. Their dataset said the leading system in March 2025, xAI’s Colossus, used 200,000 AI chips, cost about $7 billion in hardware and required about 300 megawatts of power. (clore.ai) ### Why would an AI company rent someone else’s machines? AI training already depends on systems that most startups and labs cannot afford to build. Epoch AI said several companies deployed AI supercomputers more than ten times larger than 2019-era systems by 2024, and that industry’s share of AI supercomputer capacity rose to about 80% by 2025. That leaves smaller groups looking for access rather than ownership. (arxiv.org) CLORE.AI, one marketplace advertising that model, says users can rent servers for AI training, machine learning inference and rendering from a decentralized GPU cloud. Cerebras, a more established AI hardware company, says customers can either build on-premise supercomputers with its systems or use pay-as-you-go cloud access. ### What exactly is the concern in the July 8 post? The July 8 discussion centered on cheating and verification. (epoch.ai) If a compute broker says a training run used a certain number of GPUs, a buyer may still need evidence that the hardware existed, that the code ran on it, and that the output was not fabricated or altered. That concern is not limited to social-media speculation. 0G Compute says one problem in decentralized compute is that there is “no way to verify” whether models ran correctly or whether results were tampered with, and it advertises cryptographic proof for execution. (clore.ai) BitSage and other vendors make similar claims about verified hardware and cryptographic proof of execution. ### What does “proof of compute” mean here? Proof-of-compute is a broad label for systems that try to verify that a computational job was actually performed. Some projects describe that proof as cryptographic, while others rely on hardware attestation or trusted execution environments rather than a single standard. A Yahoo Finance release from 0G Labs in April said trusted execution environments can generate attestations showing that specific code ran on specific data and produced a specific result. (compute.0g.ai) That is close to the mechanism @0xndra was calling for, though the post did not point to one formal industry standard. ### Is this already a real market, or still a crypto pitch? Several live services indicate it is already a market, even if the size is unclear. CLORE.AI advertises more than a token concept and says users can rent servers directly; Hivenet describes “AI rent” as on-demand access to GPUs, TPUs and HPC clusters; and Cerebras offers paid cloud access to its training systems. The open question is verification. As of July 9, the public evidence around this week’s X discussion shows a real supply model for rented compute and a parallel push from protocol builders to prove that rented work was actually done. (finance.yahoo.com) The next public marker is likely to come from the same venues: @0xndra’s July 8 post on X and product pages from marketplaces that say they can verify execution on rented hardware. (compute.0g.ai) (clore.ai)