How Much Can You Make Renting Your GPU for AI in 2026?

My first real obsession in this space was not Bitcoin. It was GPU mining Ethereum. I was a gamer first, so the idea that the same graphics cards I cared about could also earn while I slept hit something deep. Building those rigs felt incredible. Seating cards into frames, fanning out riser cables, getting the PSU sizing right, watching a fresh rig boot and start hashing for the first time. I had RedPandaMining, Bits Be Trippin, and Mining Chamber playing in the background while I wired everything up. Honestly, ETH mining in its prime was the absolute best this hobby has ever felt.

Then Ethereum moved to proof of stake, the hashing era ended, and a lot of that hardware went quiet or got flipped.

Fast forward to now, May 2026. Over the last few months, we have been getting the same question from people reaching out to us: "Can I actually make money with GPUs, with AI?"

The short answer is yes. The mechanism is just different. You are not solving hashes anymore. You are renting raw compute to people who train and run AI models. This guide breaks down how it works, how much you can realistically expect to make, which cards are worth buying, a proper budget section, and where to list your hardware so it actually gets used.

Why this wave is real, and different from mining

There is a reason the rig-building itch is back. The demand side of AI compute is genuinely starved right now.

A few hard facts shaping May 2026:

• AI companies are pouring well over half a trillion dollars into infrastructure this year, and demand for GPU compute massively outruns the supply that exists.
• NVIDIA has cut consumer GPU production by roughly 30 to 40 percent to redirect wafers and memory toward data center chips.
• A global GDDR7 and high-bandwidth memory shortage has pushed memory toward 80 percent of a high-end card's bill of materials, with memory prices climbing sharply through the first half of the year.
• RTX 5090 cards that launched at 1,999 dollars are now selling well above 3,000 dollars from board partners, with some variants and secondary listings pushing far higher.

Put simply, there are not enough GPUs to go around, and that is exactly what makes an idle card valuable. Every consumer GPU sitting in a closet now has a paying tenant available for it.

Here is the part that matters if you come from mining. This is not the same risk profile. There is no halving eating your reward, no difficulty bomb quietly killing your margin overnight, no surprise consensus change that retires your whole setup. Demand here comes from real businesses that need to train and serve AI models. That said, it is not risk-free either. A new GPU generation can drag down rental rates for older cards, and your earnings rise and fall with marketplace demand. It trades one set of risks for another. The trade is, in my opinion, a better one. And this is not just theory. The clearest proof is who is already doing it. Established crypto miners are openly converting their farms to AI compute, including some of the exact channels I had playing in the background back in my ETH days. RedPandaMining has been documenting his own pivot on camera: buying used multi-GPU servers, loading them with RTX 4090s and a 5090, and posting his real rental earnings as he goes. When the people who rode the last GPU wave are repositioning hard for this one, that tells you the wave is real.

Renting your GPU vs renting a GPU: clear this up first

The phrase "GPU rental" gets used for both sides of the same market, and it confuses people constantly.

You want to be the host, also called the supplier or provider. You own hardware, you list it, and customers pay you for the hours they use it. The other side, the renter or customer, is the AI developer paying to borrow compute. This whole guide is about being the host.

What do those customers actually do with your card once they rent it? Mostly four things:

1. Inference: running already-trained models. Chatbots, AI APIs, agents. This is the bread and butter, and it is steady.
2. Image and video generation: Stable Diffusion, Flux, and similar tools, often at scale.
3. Fine-tuning: adapting a smaller language model to a specific task or dataset.
4. Batch jobs and rendering: transcription, data processing, 3D rendering.

Your card does not care which one it is doing. You just provide the compute and get paid for the time.

The honest math: how much can you actually make?

I am going to try to be as straight with you here because overselling this helps nobody.

The earnings numbers platforms show you the gross revenue. Your real take-home is what is left after the costs that quietly eat into it. The formula looks like this:

Monthly net = (hourly rate x hours per day x 30 x utilization %) - electricity - GPU depreciation - platform fee (roughly 15 to 25 percent)

Those hidden costs together typically swallow somewhere between 30 and 65 percent of gross, depending mostly on where you live. Here is a realistic picture of monthly net profit per card, based on current industry data at roughly the US average power rate of 0.12 dollars per kWh:

GPU

VRAM

Realistic net per month

Notes

RTX 3080

10 GB

around 8 dollars

Barely clears costs. Not worth buying for this.

RTX 4080 / 5070 Ti class

16 GB

around 35 to 45 dollars

Strong for image generation and small LLMs. Lower buy-in.

RTX 3090

24 GB

around 40 dollars

Solid value if you already own one.

RTX 4090

24 GB

around 50 to 60 dollars

The profit-to-cost sweet spot.

RTX 5090

32 GB

around 85 dollars

Highest consumer earner, highest demand.

A100 80 GB

80 GB

400 dollars and up

Different league. Needs 15,000+ in capital.

Now the single most important lever, and it is not the card. It is your electricity rate.

At cheap-power rates (think Texas or the Midwest, around 0.08 dollars per kWh), an RTX 4090 nets noticeably more, closer to 60 dollars and up. At expensive rates (parts of Europe, 0.30 dollars per kWh and above), that same card barely breaks even or loses money. The rule of thumb: under 0.15 dollars per kWh you have real margin, over 0.25 dollars per kWh the economics for consumer cards get very hard. Know your rate before you do anything else.

For a real-world gut check, look at the dashboards operators are posting publicly. In a May 2026 video, RedPandaMining, one of the mining channels I mentioned up top, showed his live Vast.ai earnings: a verified RTX 5090 averaging about 9 dollars a day, roughly $270 a month, and verified RTX 4090s pulling around 6 dollars a day each. Those are gross figures, before your electricity and the platform's cut, so treat them as a top line rather than take-home. A dialed-in, verified operation lands above the conservative table above; a neglected one lands well below it. And that word, verified, is the whole game. Sitting right next to those earners on the same dashboard was an unverified RTX 4090 making 5 cents a day. Same card, same asking price, separated only by a badge. More on why that gap is so brutal further down.

So the honest verdict on a single card: it is side income, the kind that covers the card's own running costs and a bit more. The real money in this, exactly like mining always worked, is in scale and in better hardware. A four to six GPU AI rental rig changes the math completely. And there is a second return that is easy to forget: because of the structural shortage, the cards themselves are holding or gaining value while they earn. Your hardware is appreciating, not just depreciating. That almost never happened in mining.

Which GPUs are best for AI rentals?

For AI work, two specs matter above everything else: VRAM (how big a model the card can hold) and memory bandwidth (how fast it produces output). CUDA core counts and clock speeds are secondary for this.

The single most useful number to anchor on is 16 GB. That is the real floor for a card that will keep earning over the next two to three years. 24 GB and up gets you into the largest, best-paying jobs. 16 GB still serves a big, steady slice of demand. Below 12 GB, you are mostly locked out. Here is how the realistic options stack up, tier by tier.

The 24 GB and up tier: the headline earners

These are the cards that can hold large language models and pull the highest rates. They cost the most and, fairly, they earn the most.

RTX 5090, the flagship

The current king of consumer AI. The Blackwell architecture gives it 32 GB of GDDR7, around 1,790 GB/s of bandwidth (a 77 percent jump over the 4090), and native FP4 support. That 32 GB is the headline: it is the only consumer card that cleanly runs 32-billion-parameter models at Q4 quantization with room left for context. On language model inference it pushes roughly 35 to 46 percent more tokens per second than a 4090. It earns the most and it is the most in demand.

The downsides are real. It draws 575 watts, needs a 1,000 watt or larger power supply with the proper connector, and it is genuinely hard to find at a sane price right now thanks to the shortage.

RTX 4090, the sweet spot

If I were speccing a rental rig today and value mattered, this is where I would anchor. It has 24 GB of GDDR6X, runs 32B models at Q4, has rock-solid mature driver support, and is always in demand on the marketplaces. It does not earn as much per card as a 5090, but its profit relative to what you pay for it is the best in the lineup. For most people building their first AI rental setup, the 4090 is the answer.

RTX 3090, the value veteran

Here is the card a lot of people overlook. The 3090 has the same 24 GB of VRAM as a 4090. It runs maybe 15 to 20 percent slower on inference, but it sells used for roughly half the price of a 4090. For renting, that VRAM parity is what matters, because 24 GB is the practical floor for attracting good jobs. One real warning: the 3090's memory modules run hot under sustained load. Plan on better cooling, fresh thermal pads, or a budget for it. This is the budget MVP, and I will come back to it below.

The 16 GB tier: the smart-money middle ground

A 16 GB card will not run a 32B language model, but do not write it off. It comfortably handles 7 to 13B model inference and, just as importantly, image and video generation like Stable Diffusion XL and Flux. Image generation is one of the largest and steadiest demand categories on the rental marketplaces, and a 16 GB card does it well. These cards rate roughly three out of five stars on rental demand against the five-star 90-class, and they pull somewhere near half the daily revenue of a 5090, but they cost far less than half as much. On pure payback period, a 16 GB card often breaks even faster than a 5090.

RTX 5070 Ti, the mid-tier sweet spot

If you want a new card in this tier, this is the one. It carries the same 16 GB of GDDR7 as the 5080, has Blackwell features including FP4 support, and draws a modest 300 watts, yet it costs meaningfully less than the 5080. The best balance of buy-in, power draw, and earning in the middle of the stack.

RTX 4080 and 4080 Super, strong if you can find one

The Ada 16 GB option, and historically one of the better payback periods of any card here. The catch in May 2026 is supply: the 40-series is out of production, so stock is drying up and prices have firmed. Worth grabbing at a fair price, not worth overpaying for.

RTX 5080, capable but overpriced for rentals

The 5080 is a fast card, but for rental purposes it sits in an awkward spot. It gives you the exact same 16 GB of VRAM as the 5070 Ti for a noticeably higher price, and on AI work it is VRAM that sets your ceiling, not raw speed. Unless 5080 pricing drops back toward sanity, the 5070 Ti is the smarter buy.

RTX 4070 Ti Super

The other Ada 16 GB card. Same general story as the 4080: a viable 16 GB earner, limited by 40-series supply. A reasonable pickup if the price is right.

12 GB and below: tread carefully

The RTX 4070 and RTX 5070 ship with 12 GB. They will run smaller models and they technically list on the marketplaces, but demand for 12 GB cards is thin and the rates reflect it. Treat these as experiment-and-learn hardware, not income hardware.

The 8 GB cards, including the RTX 5060, RTX 4060, and the 8 GB version of the 5060 Ti, are not worth listing for AI rentals. 8 GB is a hard wall: it cannot do SDXL at full resolution or hold a useful language model, so it simply will not attract paying jobs. The one borderline exception is the 16 GB version of the 5060 Ti, though it has become hard to find and now prices close to a 5070, which undercuts its value.

Fewer expensive cards, or more cheaper ones?

This is the real question once you decide to build a rig, and there is no single right answer. The honest tradeoff:

Going with 24 GB and up means fewer cards, more capital tied up in each one, and a bigger hole if one dies. In exchange you get access to every category of job, higher utilization, higher rates, and hardware that is actually appreciating in today's shortage, which softens the failure risk because even a used survivor is worth a lot.

Going with 16 GB cards means more units, less money at risk per card, and built-in redundancy: one failure takes down a smaller share of your earning power. The cost is that each card earns less and sees lower utilization, because fewer jobs qualify for it. Three 16 GB cards will not reliably out-earn two 4090s the way the sticker math suggests.

My take: if you have cheap power and want maximum return per rig, anchor on 24 GB cards. If you are more worried about concentration risk, want a lower entry point, or are building rigs that need to suit a range of customer budgets, a 16 GB build is a perfectly sound call, and the 5070 Ti is where I would start. There is no wrong answer here, only the one that fits your risk tolerance.

A100 and enterprise cards

A different game entirely. An A100 80 GB can net 400 dollars or more a month, but it costs 15,000 dollars and up, and it expects real infrastructure: proper power, cooling, networking. This is a business decision, not a hobbyist one. Worth knowing it exists, but it is not where most readers should start.

The budget-friendly section: best bang for your buck

Not everyone is dropping 3,000 dollars on a 5090, and you do not need to. Here is how to get into AI rentals without overspending.

Best budget pick: a used RTX 3090. This is the standout. You get the full 24 GB of VRAM, which is what unlocks the bulk of paying jobs, at roughly half the cost of a 4090. Spend a little of what you save on cooling, the VRAM thermals are a known issue, and it will run reliably 24/7. If you are buying specifically to rent on a tight budget, two used 3090s will out-earn one mid-tier new card and give you redundancy on top.

Entry ticket: RTX 4060 Ti 16 GB (around 399 dollars). This card lets you experiment. It handles 7 to 8 billion parameter models and SDXL image generation fine. Be honest with yourself about it though: its bandwidth is low and demand for 16 GB cards is thinner, so the rates are lower. Treat any income from it as learning money, not a paycheck. It is a great way to understand the platforms before you commit real capital.

Borderline, and what to skip. Cards in the 10 to 12 GB range like the RTX 3080 technically qualify on most platforms, but after electricity, depreciation, and fees they barely break even. Only run one if you already own it and your power is cheap. Anything older than the 3000 series, or under 10 GB of VRAM, is not worth the time. The combination of an old CUDA generation and small VRAM simply locks you out of most jobs.

Budget reality check. If your goal is actual income rather than tinkering, the smartest budget move is not the cheapest single card. It is putting two used 3090s to work. Same 48 GB of combined capacity as two 4090s, far lower cash outlay, and you are diversified if one card has an issue.

Where to rent your GPU for the highest returns

Picking the platform is half the game. Here are the that matter, and who each one suits.

Vast.ai, the big marketplace

The largest GPU rental marketplace and where most serious hosts end up. You set your own price, the platform fee is the lowest of the bunch (roughly 10 to 15 percent), and demand is broad. It is more hands-on than the others, and there is one non-negotiable step: get verified. Verified listings get jobs, unverified ones mostly sit idle. If you want maximum control and maximum net earnings, this is the one.

RunPod, the standardized option

Think of RunPod as Vast.ai with more guardrails. Pricing is more standardized, the environment is more consistent, and there is a little less for you to manage. Net earnings come in a close second to Vast.ai. A good choice if you want fewer moving parts.

Salad, the easiest on-ramp

Salad is built for exactly our crowd: gamers with idle PCs. It is a simple desktop app, genuinely the lowest-effort way to start. The trade-off is that it pays the least, and it pays in "balance" you redeem for gift cards or crypto rather than a straight bank transfer. One thing to know: some of the workloads on Salad include adult content generation, and you can opt out of those in the settings if that matters to you.

io.net and Render, the crypto-native route

These will feel like home if you come from mining. They are decentralized compute networks that specifically recruit crypto miners and idle rigs, and they pay in their own token or in USDC. The familiar territory is a plus. Just go in clear-eyed that getting paid in a token means you are also taking on that token's price swings.

Most experienced hosts list on two or three platforms at once and route idle capacity to whichever has demand. Just check each platform's terms first, since a few require exclusivity.

Setting up and actually maximizing your earnings

The platforms make joining easy. Earning well takes a bit of discipline. The providers who out-earn everyone else all do the same handful of things.

Don't starve your GPU: the rest of the rig

The card is the headliner, but it does not run alone. The CPU, RAM, and storage around it will not earn you a single extra dollar, but get them wrong, and they will quietly cost you money. A great GPU paired with a weak system loses jobs to a great GPU paired with a balanced one, because marketplaces like Vast.ai benchmark and list your whole machine, and renters filter on those specs. Data loading and preprocessing lean on the CPU and RAM, so if those are starved, your GPU sits idle waiting and the renter has a worse experience on your box. Build the supporting cast so it never bottlenecks the star:

• System RAM: aim for at least your GPU's VRAM, and ideally 1.5 to 2x it. A 24 GB card wants 32 GB minimum, with 48 to 64 GB comfortable. A multi-GPU rig should cover every card's VRAM plus headroom, so plan on 128 GB or more.
• CPU: you do not need a flagship. A modern mid-range chip is fine. What matters is core count, since each GPU's workload needs threads feeding it. Avoid old or very low-core CPUs that choke on data loading, and scale cores up as you add cards.
• Storage: a fast NVMe SSD is close to mandatory. Renters download large datasets and model weights onto your machine, and slow disk drags the whole job down. Budget for 1 TB or more of NVMe.
• Power supply: on a 24/7 rig this is a reliability and safety item, not an afterthought. Use a quality unit with real headroom above your total system draw, not just the GPU's rating. On 4090 and 5090 builds, seat the 12VHPWR power connector fully, every time, because a partly seated connector is a genuine fire risk.

Verification is the difference between real money and almost nothing

If you take one operational lesson from this guide, take this one. On Vast.ai your machine has to pass the platform's checks and build a track record before it earns a verified badge. Until it does, two things quietly bury you: the marketplace ranking pushes your listing down, and most renters filter for verified hosts only, so you are simply invisible to them.

How big is the gap? Go back to that operator dashboard from earlier. A verified RTX 4090 was pulling around 6 dollars a day. An unverified RTX 4090, the same card at the same asking price, was making 5 cents a day. That is identical hardware earning either roughly 180 dollars a month or roughly 1 dollar a month, decided entirely by a badge. Your card does not matter if it is not verified.

So treat verification as step one, before pricing tweaks or anything else. And if you are scaling into a multi-GPU rig, remember that every single card has to clear verification before it earns. Budget your first few weeks near zero and push to get the whole fleet verified as fast as you can.

Hardware checklist: a 24 GB or larger NVIDIA card ideally, a stable upload connection of 50 Mbps or more, Linux preferred (Windows works), Docker support, real PSU headroom, and cooling that keeps the card under 80 degrees Celsius all day.

Use a dedicated machine. Do not rent out your daily gaming rig. Uptime is everything to the marketplace ranking algorithms, and a shared machine cannot deliver it.

Price to win first, then raise. List slightly below the going market rate for your card to attract early jobs and build a reliability score. Once your reputation is solid, nudge the price up.

Protect your uptime. Aim for 95 percent or better during your listed hours. A UPS to ride out brief power blips is a cheap insurance policy for your ranking.

Be patient with month one. Expect roughly 30 to 40 percent of your steady-state earnings in the first month while the platform learns you are reliable. Things stabilize around month three. Do not judge this by week one.

Keep an eye on the basics. Watch temperatures, keep your NVIDIA drivers current so you stay compatible with incoming jobs, and remember this is taxable income. Track your electricity, depreciation, and internet costs, because they are deductible.

So, is renting your GPU for AI worth it?

It depends entirely on your situation, so let me break it down honestly.

• You already own a 3090, 4090, or 5090 and your power is cheap: yes, easily. This is close to free money on hardware that is otherwise sitting idle.
• You game on that card daily and you pay 0.30 dollars per kWh: probably not. The wear and the power cost will not be worth the modest return.
• You want this to be real income: then think in rigs, not single cards. A four to six GPU AI rental rig is where the numbers get genuinely interesting, the same way mining always rewarded scale.
• You just miss the rig-building days: this is the closest thing we have to that era, and honestly the hardware holds its value far better than it did in mining, because the shortage driving prices is structural rather than speculative.

Closing thoughts

This is exactly why we have been fielding so many GPU questions lately. The same instinct that made building ETH rigs so satisfying is back, just pointed at AI compute now. The difference is that this time the demand is coming from real businesses, the hardware is appreciating instead of bleeding value, and you do not need to chase a halving clock.

If you are putting together an AI rental rig and want a hand speccing the right cards, power supplies, and frames for it, that is what we do at Endless Mining. Reach out and we will help you build something that earns.

The figures in this guide are realistic estimates based on current market data and will shift as the market moves. The biggest variable is your local electricity rate, so always run the numbers for your own situation before buying hardware. This is not financial advice.