Loading page…
An $8,500 GPU rig rented out to strangers for AI workloads at $0.50/hour — $12 a day, ~$365 a month, and a payback period of two to four years paid in a volatile token
~$84/week · ~$365/month before electricity · ~70% utilisation over the first 4-5 days · $8,000-8,500 hardware cost · 2-year payback at 100% uptime, ~4 years at 50%
The honest arithmetic, from the person who built it: "at an $8,000 investment, if we were rented out at $12 a day, 24/7, 100% uptime — 708 days. So around a 4-year ROI on this hardware" at realistic utilisation. That's the whole story.
Fewer bars = easier, cheaper, or faster for an AI-assisted solo builder. Editorial judgments based on the case details.
The operator builds GPU rigs and rents them out for AI and crypto workloads. This video documents an $8,500 dual RTX 5090 build and — unusually for the genre — walks through the actual returns without inflating them. The build: an EPYC 7B12 server-grade motherboard with eight PCIe slots, 128GB of DDR4 RAM (eight 16GB sticks at 2133MHz), an Arctic Freezer 4U CPU cooler, a 1TB NVMe SSD, two RTX 5090s (a Zotac and an ASUS TUF), a 1600W Cooler Master Platinum PSU paired with an older 1300W Seasonic for headroom. His RAM rule of thumb: twice the combined VRAM, so 64GB of VRAM across two 5090s means 128GB of system RAM. The build itself went badly in the way real builds do. The CPU cooler's mounting screws wouldn't seat properly — he suspects the unit had been used and repacked, and the supplied thermal paste was old and hard. He improvised. The motherboard wouldn't post with the PSU's 12V-2x6 cable, so he ran four separate PCIe cables instead. He's candid about all of it, which is part of why the numbers are believable. Once running: HiveOS as the Linux base layer, OctaSpace as the rental marketplace. He lists the rig at $0.50 an hour, which produces about $12 USD a day — roughly $17 AUD, which he notes feels better on the exchange rate. And then he does the arithmetic honestly, which is rare enough to be the reason this entry exists at all. At $12 a day with perfect 24/7 uptime, an $8,000 rig takes 708 days to pay back — just under two years. At the 50% utilisation he considers more realistic across a full year, "around a 4-year ROI on this hardware." His GPUs carry a three-year warranty, which is why he bought new.
There is no outbound here. OctaSpace supplies the demand and the operator competes inside the listing on two variables: price and stability score.
The quick-reference card — acquisition channels, replication playbook, risks & traps, and our verdict — is members-only.
Join the Unicorn Club — $5/moCancel anytime, in one clickPaste into Claude Code / Codex and get a working version of this product end-to-end.
Two production-grade prompts per idea — a full build spec and an SEO growth plan — plus 10 tailored build specs a week. Members only.
Join the Unicorn Club — $5/moCancel anytime, in one clickEvery claim traces back to a listed source — open any of them and check the number yourself.
Data credibility: 🗣 Founder-Reported by the person who built and operates the rig, on his own channel. The hardware build is shown on camera component by component, and the HiveOS and OctaSpace dashboards are shown live including a screenshot of 124W draw and just under 3kWh consumed by 11am. The revenue figures — $0.50/hour, ~$12 USD/day, ~$84/week, ~$365/month — are his own, computed on camera, and the utilisation estimate of ~70% comes from only four or five days of operation during which he also lost roughly 12 hours to BIOS issues, so the sample is very short. The ROI arithmetic is his own and he presents it conservatively (708 days at 100% uptime, ~4 years at 50%), which is the opposite of how this genre usually behaves. Three material things are absent from his accounting: hardware depreciation and residual value, the effect of a declining marketplace rate as GPU supply increases, and any cost for his own time. His margin depends entirely on paying nothing for electricity due to solar and battery, which he states plainly — for anyone on retail power his own estimate of $4-5/day would consume a third to half of gross revenue. Revenue is received in the OctaSpace token, so the dollar figures depend on token price at the moment of sale. He links the tools he uses in the description.
The unit sold is an hour of access to the machine, and the money starts the day the rig is listed — there is no sales cycle to wait through.
Two things compound here, and the compute is not one of them. Inside the marketplace the only loop that runs in the operator's favour is uptime: crashes lower the stability score, a lower score forces a lower hourly rate, and a lower rate stretches the payback.
For this business the quiet stretch is the payback period, and he puts a number on it rather than a feeling: 708 days rented 24/7, and "around a 4-year ROI on this hardware" at the 50% utilisation he considers realistic. That is longer than the three-year manufacturer warranty on the GPUs that make up 80% of the cost.
You are selling compute time on hardware you own, through a marketplace that matches renters to rigs. OctaSpace is a decentralised compute marketplace. Renters browse by GPU model, filter for what they need, and rent a machine by the hour. Your rig appears in that listing alongside everyone else's, with a price you set and a stability score the platform assigns based on whether your machine has crashed. HiveOS sits underneath as the operating layer — a Linux distribution built for GPU rigs, giving remote monitoring of temperatures, power draw and utilisation, plus the ability to undervolt and tune from a dashboard. He uses it to cap boost clocks for efficiency after finding the motherboard BIOS wouldn't let him do it cleanly. What renters actually run on it varies and he can't see much of it. During the video the rig is running an image called "Anything LLM" — someone else's local language model workload — drawing only 124 watts, essentially idling on GPU while holding VRAM. An earlier rental was crypto mining, which pulled 800-900 watts and can reach 1,500. Another was something called "comfy worker" that he couldn't identify. He raises the obvious concern himself and dismisses it for his own case: "I don't exactly know what they're doing. So if you're someone who is a bit security... stresses about things like that, then this probably isn't for you. But for me personally, I'm not really too worried about it." Our read: that's a real consideration and worth sitting with. You are renting root-level access to hardware in your home to anonymous strangers, and you do not control or audit the workload. The financial risk is quantified in this video; that one isn't.
Hourly rental income, paid in the OctaSpace token. His rate is $0.50 an hour. Competitors on the marketplace list as low as $0.40, but he points out those tend to have lower stability scores — machines that have crashed rent for less. At $0.50 with the utilisation he's seeing, the income line is: about $12 USD a day, $84 a week, and roughly $365 a month before electricity. In Australian dollars that's about $547 a month, which he acknowledges flatters the figure. Utilisation is the number that decides everything. Over the first four or five days online he was rented roughly 70% of the time, and he believes 70-80% is sustainable now that he's resolved his BIOS issues — but he also lost about twelve hours to those problems in that window, so the sample is short and the estimate is his own. Electricity is the variable that makes or breaks this for most people. The current LLM rental draws 124 watts; the rig had used just under 3kWh by 11am, which he prices at roughly 30 cents at 10 cents per kWh. Under crypto mining load it draws 800-1,500 watts. His own honest range for a rig like this: "you could expect it to use anywhere from $4 to $5 per day." On $12 a day of revenue, that is a third to nearly half of gross income. He pays nothing, because he has solar and a battery. "I don't pay for any electricity, fortunately. I've got solar and I've got a battery, so for me, this is all profit." That single fact changes the business from marginal to viable, and it is not a fact most readers can replicate. And the payment is in tokens, not dollars. His strategy: "you can choose to sell the OctaSpace straight away, or because the OctaSpace is quite undervalued at the moment with the markets tanking, you could say sell 75% of it, keep 25%." That's a currency bet layered on top of a hardware bet.
You don't get customers. The marketplace does, and you compete inside it on two variables. Price is the first. He surveys the OctaSpace marketplace on camera and observes that the 5090 tier is filling up — "there's quite a lot of them online now, right? Hence why I built this two-card rig, because we can see the competition here isn't very high" for dual-GPU configurations specifically. So the positioning play is configuration scarcity: be in a tier where fewer machines are listed. Stability score is the second, and it's the one you actually control. The platform tracks crashes, and "generally, you know, the rigs that are having a slightly less stability score are going to rent for less." Machines that stay up rent at a premium. That's why his undervolting and cooling work matters commercially, not just thermally — an unstable rig is a discounted rig. The practical implication is that this is a pure commodity market. Your machine is fungible with every other machine of the same specification, price is the primary sort, and the platform sets the terms. There is no brand, no relationship, no repeat customer and no way to differentiate beyond uptime and rate. He also notes that idle time isn't entirely dead — "when it's not being rented, it just does idle jobs" — but doesn't quantify what those earn. Our read: commodity marketplaces have a well-understood trajectory. Early participants earn above-normal returns because supply is scarce; as supply arrives, price converges toward the marginal operator's cost — which, for anyone paying retail electricity, is far above his. His own footage shows the 5090 tier already crowding. This is the single strongest argument for treating the two-year payback estimate as optimistic.
Hardware, as listed on camera: EPYC 7B12 motherboard with eight PCIe slots, 128GB DDR4-2133 (eight 16GB sticks), Arctic Freezer 4U M Rev 2 cooler, 1TB NVMe SSD, one Zotac RTX 5090 and one ASUS TUF RTX 5090, a Cooler Master 1600W Platinum V2 PSU plus a Seasonic 1300W for headroom. Total: approximately $8,000-8,500, and he's explicit that the GPUs are about 80% of that. Software: HiveOS as the Linux layer for monitoring and tuning, OctaSpace as the rental marketplace and payment rail. The cost lines, with the numbers he gives: Electricity — $4 to $5 a day for a rig like this under normal use, against $12 a day of revenue. He pays zero due to solar and battery. Depreciation — never mentioned, and it's the largest hidden cost. GPUs lose value fast as new generations ship, and a 5090 in four years will not be a premium rental tier. Token exposure — revenue arrives in OctaSpace tokens whose value fluctuates independently of the compute you sold. Hardware risk — mitigated by buying new for the three-year manufacturer warranty, which he flags as a deliberate decision. His own ROI arithmetic, verbatim: "at an $8,000 investment, if we were rented out at $12 a day, 24/7, 100% up time, 708 days, allegedly that would be. So you divide that by 365, just under 2 years at this rate. Now, let's just assume it's rented out for 50% of the time — you times that number by two. So around a 4-year ROI on this hardware." Our read: run that again with retail electricity. At $4.50 a day of power, net revenue drops to $7.50 a day, and the 100%-uptime payback stretches to about 1,067 days — nearly three years. At 50% utilisation with power costs running whether or not you're rented, it goes past five. On hardware with a three-year warranty and a fast depreciation curve, that is not a business.
Copy the honesty, first. This is a video about a $8,500 crypto-adjacent hardware purchase that voluntarily computes a four-year payback and tells viewers "this isn't for everyone... you've got to be in a position where you're willing to lose if you're going to be into this sort of crypto stuff and the AI rentals." That framing is worth more than the build guide. Copy the entry path he actually recommends, which is not this rig: "if you want to set this up at home after you've watched the video, you can set it up on any sort of graphics card. It doesn't need to be something like this... if you've got a high-end sort of gaming PC and you wanted to actually make some real dollars in passive income, you can just run this overnight." Test the model on hardware you already own before spending anything. That's free validation and it's the correct first step. Copy the stability discipline. Undervolt, cap boost clocks, keep temperatures down, and treat uptime as a revenue variable rather than a technical nicety, because the marketplace prices it directly. Copy the configuration-scarcity logic: look at what's oversupplied in the marketplace listings and position where it isn't. He notes 3090s rent for about 5 cents an hour because they're massively oversupplied. What to change, in one sentence: do not do this without free or near-free electricity. His entire margin exists because he has solar and a battery. Take that away and $4-5 a day of power against $12 a day of revenue at 70% utilisation leaves a return that does not compensate for the capital risk, the depreciation, or the token volatility. Also change the payout handling. Holding 25% of income in a volatile token because it's "quite undervalued at the moment" is speculation, not a business decision. If you run this, convert to fiat on receipt and let the compute business be a compute business.
The payback period is the risk. Two years at perfect uptime with free electricity; four years at realistic utilisation; longer still if you pay retail power. The GPUs carry a three-year warranty. You are betting $8,500 that a specific hardware generation stays commercially rentable for longer than it stays under warranty. Depreciation, which he never mentions. Consumer GPU generations turn over roughly every two years. A 5090 is a premium rental tier today and a discounted one by the time this rig pays for itself. There is no residual value assumption in the arithmetic at all. Commodity price collapse. His own footage shows the 5090 tier filling up on the marketplace, and 3090s renting at around 5 cents an hour after oversupply. That is the future of every GPU tier. Rental rates only go one direction as supply arrives. Token exposure. Income is paid in OctaSpace tokens. You can sell immediately, but you are still exposed between earning and selling, and he explicitly describes holding a portion as a speculation. Crypto-denominated revenue is not the same thing as revenue. Electricity, which for almost everyone reading this is the deal-breaker. $4-5 a day against $12 of revenue, before any of the above. Security and legal exposure. You are giving anonymous strangers compute on hardware in your home, with no visibility into the workload. He accepts this; you should decide consciously rather than by default, and consider what your insurer and your ISP's terms of service have to say about it. And hardware risk. His build had a defective-seeming cooler and a PSU cable that wouldn't post. Two 5090s under sustained load in a residential room is a thermal and electrical situation that deserves respect.
The four founder playbooks — first customers, first dollar, the flywheel, the silent stretch — are in the Unicorn Club, on this idea and on all 1,403 others. The story, the numbers, the sources and every later section stay open to everyone.
Join the Unicorn Club — $5/moCancel anytime, in one click