Where Should You Deploy Your GPU? What Compute Actually Pays

You own the hardware. An RTX 4090GPUGraphics Processing Unit. Originally designed to render video game graphics, GPUs turned out to be exceptionally good at the massively parallel math that AI models need. Modern AI training and inference runs almost entirely on GPUs.Like a factory with 10,000 workers doing the same simple task in parallel, versus a CPU which is more like 10 workers each doing different complex tasks. AI training involves doing simple math a million times per second on a million numbers, which is exactly what the GPU factory is designed for.GPU in the glossary → at home, maybe a couple of 3090s, or a small rack of datacentre cards. The question is where to point it so it earns, and what it actually earns once the electricity bill and the hardware cost come out. The marketplaces all show you a headline rate. None of them show you the two numbers that decide your real income.

The first is utilisation: the fraction of the month your card is actually rented and working. The second is the currency you get paid in. Some networks pay cash. Others pay a token you then have to manage and sell. Those two variables split the entire market into cash marketplaces and token networks, and that split is the spine of this piece.

Here’s the counterintuitive part. The most predictable place to provide compute is the least crypto-native one. Providing to a token network pays the same wage as a cash marketplace, with a managed market exposure layered on top. The word that matters there is risk, not inferiority.

A provider who reprices and sells on receipt turns a token network into something close to a cash wage, minus friction and slippage. A provider who holds the token is making a separate investment bet, and it should be booked as a position rather than as compute income.

This guide’s worst-case baseline is Australian residential power (about AUD 0.30/kWh, roughly USD 0.19 to 0.20/kWh), with US residential and cheap datacentre power scenarios modelled alongside, because the reader is global. The figures are point-in-time; marketplace rates and card prices were last re-priced on 4 October 2026. Where token value comes up, treat it as analysis, not financial advice.

Closed Networks and Who You Can Actually Sell To

Most of the compute market is shut to you. The hyperscalers (AWS, GCP, Azure) and the large GPU neoclouds (CoreWeave, Lambda Labs) own and operate their own hardware. You can’t list a card with them. They matter here only as the price ceiling the open marketplaces undercut: an RTX 4090 rents near $2.79/hr on a comparable AWS instance and an H100 PCIe near $6.16/hr on CoreWeave, against about $0.40 to $0.55/hr and $1.95 to $2.40/hr on Vast.ai for the same silicon.

The venues that actually let a third party list their own hardware split into two groups.

Vast.ai is the open centralised marketplace a new host can still join, and it’s the cash benchmark for everything that follows. RunPod’s Community Cloud was the other one, but it no longer accepts new hosts. Vast pays fiat, per second, with no token to manage. Every token network is measured against that benchmark as that same wage plus an exposure you either carry or neutralise.

Unit Economics: What a GPU-Hour Nets You

Start with the card most readers actually own. An RTX 4090 draws about 400W under load. On Australian residential power that’s about $0.08 of electricity per rented hour; on US residential power (about $0.15/kWh) about $0.06; on cheap industrial or datacentre power ($0.05 to $0.08/kWh) about $0.024 to $0.032. Amortise the card itself (about $2,400 over three years, the 90-day average eBay sold price for a used 4090 on getpcparts.com) and you add roughly $0.09/hr if it runs continuously.

At a Vast.ai host-kept rate near $0.35/hr, the 4090 nets about $0.27/hr after Australian electricity, rising to $0.32/hr on cheap power. Subtract the $0.09/hr hardware amortisation and true net while rented is roughly $0.18 to $0.23/hr. That’s the number per hour. The monthly figure is where utilisation does its damage.

After hardware amortisation, a realistic Australian prosumer 4090 nets roughly $25 to $65 per month at 20 to 50% utilisation. That’s passive income on hardware you already own, not a business. An RTX 3090 earns far less: about $0.10 to $0.17/hr host-kept on Vast, against about $0.07/hr of Australian electricity. After amortisation it nets roughly $0 to $25 per month, which is marginal at best.

The H100 is a different animal and not a home card. On Vast.ai it earns the host about $1.55 to $2.20/hr, but it draws about 700W and amortises at roughly $1.07/hr (about $28,000 over three years). On cheap power ($0.06/kWh, about $0.042/hr electricity), net while rented is about $0.45 to $1.10/hr. At 70% utilisation that’s about $770 to $1,100 a month before amortisation, and roughly $225 to $560 per card after it. On Australian residential power an H100 is uneconomic before you even count cooling.

So the breakeven rule of thumb: a consumer card on Australian residential power has to clear roughly $0.08 to $0.10/hr just to cover electricity, and below about 20 to 25% utilisation at typical rates the net contribution is trivial. Datacentre cards only make sense on sub-$0.10/kWh power. Vendor calculators tend to assume optimistic utilisation; that’s the assumption to stress-test before you believe any projection, including the ones here.

Platform by Platform: How Each One Pays

The cash benchmark first, then the token networks, each with the specific way its tokenTokenA digital unit of value or access rights tracked on a blockchain. Tokens can represent ownership in a project, a right to use a service, a share of future revenue, or simply a tradable asset with no underlying claim.Like a physical poker chip a casino issues. The chip itself has no value. What makes it worth something is what it lets you do at the casino, what the casino has promised, and how much other people will pay you for it.Token in the glossary → risk is carried or removed.

Vast.ai: the cash leader for prosumers

Vast.ai pays per GPU-hour at a host-set price, billed per second. Vast describes its cut two ways. Its renter pricing docs promise “no markup layered on top of host-set prices”, while its May 2026 host earnings article says “Live GPU prices are typically about 25% above what hosts earn.” We take the host article at its word, so the host-kept rates here are roughly 80% of the live renter ask.

Payout is fiat, rolling, no minimum, crypto accepted. Representative host-kept rates run RTX 3090 about $0.10 to $0.17/hr, RTX 4090 about $0.29 to $0.59, RTX 5090 about $0.30 to $0.60, A100 about $0.50 to $0.85, and H100 and H200 about $1.55 to $4.00. A four-GPU RTX 5090 rig at 80% utilisation grosses roughly $700 to $1,400/month.

Demand is the deepest on the open market, with about 120,000 developers active on the renter side. Token risk: none. Start here unless you have a specific reason not to.

RunPod Community Cloud: closed to new hosts

RunPod’s Community Cloud runs on third-party hosts paid in fiat per second. Its docs now say: “Runpod is no longer accepting new hosts for Community Cloud. Existing Community Cloud resources remain available.” The note is absent from a December 2025 Wayback capture of that page and present in May 2026. A new card owner can’t list there, so RunPod stays in this guide as a renter-price benchmark and as a venue for hosts already on it.

Current renter prices run RTX 3090 about $0.22 Community against $0.50 Secure, RTX 4090 about $0.34 against $0.74, RTX 5090 about $0.69 against $0.99, A100 from about $1.19/hr and H100 from about $1.99/hr. Secure Cloud runs in RunPod’s partner datacentres. RunPod Hub pays repository maintainers up to 7% of the compute revenue their published repos generate, in RunPod credits. Token risk: none for existing hosts.

Akash: the decentralised option priced in dollars

AkashDePINDecentralised Physical Infrastructure Networks. Protocols that use token incentives to coordinate real-world physical infrastructure like GPU compute, wireless networks, storage, mapping sensors, or bandwidth.Like crowd-sourced ride-sharing but for physical hardware. Uber incentivises drivers with dollars. DePIN incentivises hardware operators with tokens. The network grows because individuals choose to contribute capacity in exchange for rewards.DePIN in the glossary → runs a reverse auction. Tenants post requirements and a price, providers bid, the lowest qualifying bid wins, and the lease settles on-chain. Since Burn-Mint EquilibriumBurn-Mint EquilibriumA tokenomics model where network fees burn tokens while new tokens are minted and paid to suppliers. The system tries to balance burns and mints so circulating supply stays roughly stable when usage scales.Like a business that spends a dollar of revenue for every dollar of wages it pays. Money flows in and out at the same rate, so the total cash in the company stays flat. The rate of flow tells you how big the business is.Burn-Mint Equilibrium in the glossary → went live on 23 March 2026, leases are priced and paid in ACT, a USD-pegged compute credit minted by burning AKT, which the provider redeems back to AKT at the market price. That keeps the lease income dollar-denominated until you redeem. Demand is real but small and recently softening: Akash’s own Q3 2025 figures showed 367 GPUs and utilisation above 50%, then Messari counted 198 GPUs in use in Q4 2025 and 84 of 334 (33.7% utilisation) in Q1 2026. H100s rent about $2.00 to $2.70/hr on Akash’s own GPU price feed. What holds Akash back is the Kubernetes provider stack you have to run and the thin, declining demand. Token risk: small if you redeem and sell promptly, because only the AKT leg after redemption floats. Full assessment in our Akash review.

io.net: idle rewards plus job revenue

io.net pays two streams. Hourly block rewards come from a 300M IO supplier pool released over about 20 years (95% to GPU suppliers), gated by staking 200 IO per chip and at least 5 hr/day uptime, and paid whether or not your card is working. On top of that, you earn compute-job payments when a card is actually booked. Higher-end cards capture the job revenue; consumer cards mostly earn the idle block rewards. The change worth watching is the Incentive Dynamic Engine (IDE), which is meant to shift suppliers to demand-driven dollar-target payouts designed to hold supplier ROI stable regardless of the IO price. Its burn went live on-chain in June 2026, but the demand-driven supplier payouts are still not active: epoch rewards remain a fixed drip. Token risk: IO price risk until those payouts switch on. See our io.net review.

Render: high-VRAM cards and a payment delay

Render node operators earn availability plus completion rewards under Burn-Mint Equilibrium, priced in OctaneBench-hours (Tier 2 priority at 100 OBh per RENDER, Tier 3 economy at 200 OBh per RENDER), with a 5% protocol fee per job and payment released about 7 to 10 days after the creator approves the work. It started as a rendering network, and reported operator income runs about $300 to $800/month per high-end GPU in rendering contexts. The 2025 Compute subnet and a Salad integration (about 60,000 GPUs) extend it toward AI. Best fit for OctaneRender-capable high-VRAM cards. Token risk: RENDER price risk plus a payment delay, which is harder to manage than the others because you can’t sell on receipt if receipt is 10 days out. Provider walkthrough in how to earn with Render.

Bittensor, Aethir, Phala: niche, enterprise, or high-skill

Three networks that look like compute income but aren’t a route for a single consumer card:

• Bittensor (TAO). Per-block emissionsEmissionsNew tokens created and distributed by a blockchain protocol over time as rewards to validators, stakers, or miners. Emissions fund network security and participation at the cost of diluting existing holders.Like a company that pays employees partly in newly printed shares. Every year the total number of shares goes up, which means existing shareholders own a slightly smaller slice of the same company unless the company grows faster than the printing.Emissions in the glossary → split 41% miners, 41% validatorsValidatorA computer that runs the full blockchain protocol, verifies transactions, and proposes new blocks. Validators are the workers that keep a Proof of Stake network running, and they earn rewards for doing the work correctly.Like a notary public who witnesses and stamps legal documents. Validators witness transactions, check they follow the rules, and stamp them into the permanent record. A notary who commits fraud loses their license. Validators work the same way, except the license is staked tokens that get slashed on misbehaviour.Validator in the glossary →, 18% subnet owner. The first halvingHalvingA protocol event that cuts the rate of new token emissions by half. Halvings are scheduled in advance, happen automatically at fixed intervals, and are a core mechanism for enforcing declining token supply growth over time.Like a savings account where the interest rate is contractually cut in half every four years. You still earn interest, but the rate drops on a known schedule, and the issuer can't change it without breaking the contract.Halving in the glossary → around 12 to 14 December 2025 cut emissions from roughly 7,200 to 3,600 TAO/day. Earning on a compute or inference subnet (such as Targon, SN4) means building a competitive subnet miner and out-competing on validator-scored quality. That’s ML engineering, not GPU hosting. Don’t treat it as plug-in compute.
• Aethir (ATH). Cloud Hosts earn 80% of customer service fees plus ATH. Aethir’s self-reported Cloud Host economics assume up to 95% utilisation and aren’t independently audited, and the role needs enterprise capital, KYC, staking, and SLAs. US residents can’t monetise. Checker-node earnings are trivial. Not a single-card route. Our Aethir review has the detail.
• Phala (PHA). TEETEETrusted Execution Environment. A hardware-secured region of a CPU or GPU where code runs in isolation, so even the machine's operator can't read what's happening inside. TEEs give decentralised AI inference privacy guarantees.Like a bank vault inside a bank. The bank owns the building, staffs the lobby, and runs the security cameras. But what's inside the vault is invisible to everyone, including the bank staff, unless the customer opens it.TEE in the glossary → GPU workers earn by hardware weight times uptime, paid hourly in vPHA, with vPHA staked as collateralCollateralTokens posted as security against an action and returned when the action completes. Collateral is held, not spent. It backs a position or a session and comes back to the owner once the obligation clears.Like a rental bond. You hand it over before you move in, the landlord holds it while you occupy the place, and you get it back when you leave the property in order.Collateral in the glossary → per GPU. The differentiator is verifiable privacy via confidential computingConfidential ComputeHardware-enforced computation where data and code are encrypted in memory and only the authorised application can access them. The machine's operator cannot read what the application is doing even though they own the machine.Like renting space in a bank vault. The bank owns the building and runs the security, but what you put in the vault is invisible even to the bank staff. Only you have the key.Confidential Compute in the glossary →, not raw yield, which is modest. See the Phala review.

One clarification that saves an argument: Venice (VVV/DIEM) is not a provider marketplace. It’s the demand side. Users stake VVV for emission yield and can lock staked VVV to mint DIEM, each staked DIEM worth $1/day of API credit. Venice has never disclosed who runs its GPU fleet or how those providers are paid, and you can’t become a Venice compute provider. It belongs in this market as the consumer counterpart that buys compute. The Venice review covers the demand-side economics.

The Morpheus Case Study: On-Chain Provider Income

This is the part nobody else publishes, because it needs an on-chain indexer pointed at the contract, and we run one. Morpheus is worth a deep look for one reason: every claim about its provider economics is verifiable on-chain, and what’s verifiable contradicts what you’d assume. It doesn’t pay the most today.

How a provider gets paid: in MOR drawn from the 24% Compute pillar of emissions, via the Yellowstone model. You run the Morpheus-Lumerin node on Base, stake MOR, and register a modelModelA trained neural network that takes inputs (text, images, audio) and produces outputs (more text, classifications, generated content). In DeAI the model is the thing that actually does the work.Like a very experienced apprentice who has spent years watching thousands of masters make furniture. They can't explain how they know when a joint is right, but they can make a chair that looks and functions like a Chippendale. The training is invisible. The output is what matters.Model in the glossary → through on-chain bids. The router sends users to the lowest bid, and payment is bid price times session length. Worth knowing before you set a bid: the contract sizes session duration as the consumer’s stipend divided by your bid price, so bidding higher earns you the same MOR over a shorter session rather than more MOR, until the seven-day cap binds. Almost all traffic runs on the staked-inference path, where the user’s MOR collateral is returned in full at session close and nothing is burned. Direct payment, where the session’s cost comes out of the user’s own deposit, is a rounding error by token volume. Providers are paid whether or not any external customer pays. The full payment trace is in how Morpheus pays for inference.

Now the size of the pot. The 24% compute allocation accrues about 2,879 MOR a day into the pool providers are paid from. That’s the headline. The on-chain reality is smaller by an order of magnitude.

Over the trailing 30 days our indexer records 10,968 MOR earned by all providers combined, about 366 MOR a day against the 2,879 the pillar accrues. The rest accrues on the schedule and goes unclaimed. These are our own indexed figures (Base RPC event logs plus Alchemy contract reads on the LumerinDiamond at 0x6aBE1d282f72B474E54527D93b979A4f64d3030a). Official Morpheus sources confirm the mechanics but don’t publish provider counts or per-provider earnings.

Supply has spread out since June, when 47 of the last 50 settled sessions came from a single provider address. Over the last 30 days 24 providers closed sessions, the busiest took 24.8% of them and the top three 51.8%. Our index counts 43 provider addresses that have ever closed a session.

Here’s where the common misconception needs correcting, and it cuts in Morpheus’s favour. The Yellowstone bid market prices MOR volatility in. You set MOR-denominated bids and can raise them as MOR falls, so a managed provider isn’t forced to serve below cost. The design intent is competition driving inference price toward base electricity, with loss-making providers exiting until equilibrium returns. Two residual exposures remain:

• Bids sit on-chain, so a stale bid during a fast MOR drawdown can transiently serve below cost until you reprice. Gas on Base is cheap, so this is second-order.
• The daily compute budget is MOR-denominated, so the network’s total dollar subsidy scales with the MOR price. The binding constraint today is volume and provider concentration, not price.

Fact: Over the trailing 30 days, Morpheus providers earned about 366 MOR a day, spread across 24 providers, against about 2,879 MOR a day accruing to the compute pillar (OYM indexer, refreshed daily).

Take: A new provider’s expected pure income today is low. Thin session flow is the reason, and the payment token has little to do with it. The bid market does let you manage MOR price risk; what it can’t manufacture is customers. The whole pot is small, and on the open models a home GPU can serve it’s priced by resellers whose marginal cost per session is close to zero, so the cheapest listed bid takes most of each book.

So who should run a Morpheus provider node? Someone who believes in the protocol, wants a seat early in a thin market, or values the TEE-attested private-inference stack enough to subsidise their own entry. That’s a strategic position, and a defensible one. It’s not a reliable wage, and anyone selling it as one is reading the emissions allocation instead of the settlement data. If you want the mechanics before the conviction, the consumer node guide and the Lumerin provider setup are the on-ramps. (Analysis, not financial advice.)

Provider Ranking on Pure Earnings

Ranked for a single prosumer running consumer cards on Australian residential power, judged on realised income alone. Your hardware, power price, and tolerance for managing a token will reorder this.

1. Vast.ai. Best predictable cash, full host-set price retained, instant fiat withdrawal, deepest demand. Start here.
2. RunPod Community Cloud. Comparable cash for hosts already listed. It no longer accepts new hosts, so a new card owner can’t start here.
3. Akash. Leases are priced in USD-pegged ACT, so the income carries little token price risk if you redeem and sell promptly. Needs Kubernetes, and demand is modest and recently declining.
4. io.net. Reasonable once the IDE dollar-target payouts are fully live. Until then it carries IO price risk and consumer cards mostly earn thin idle rewards.
5. Render. Only with OctaneRender-capable high-VRAM cards, and you’re accepting RENDER price risk plus a 7 to 10 day payment delay.
6. Phala and Bittensor. Niche or high-skill. Phala for confidential-compute believers, Bittensor only if you can build a competitive subnet miner.
7. Aethir. Enterprise scale only, US residents excluded.
8. Morpheus. Lowest expected pure income for a new typical provider today, because session flow is thin and cheap resellers set the price on open models. Deploy for conviction.

Venice doesn’t appear because you can’t provide to it.

The caveats that govern the whole table:

• Token-price risk is manageable but real on every token network. Sell on receipt and it’s friction plus slippage; hold and it’s a separate investment position.
• Utilisation risk is the biggest swing factor, and it’s the number vendor calculators inflate.
• Aethir and Venice figures are self-reported or undisclosed, so treat them as claims, not facts.
• Technical barriers vary widely: Vast is near plug-and-play (RunPod Community no longer takes new hosts), Akash needs Kubernetes, Morpheus needs the Lumerin node plus a hosted LLM, Bittensor needs ML engineering.
• Australian residential power makes H100s uneconomic at home and squeezes consumer margins. All prices and schedules are point-in-time, last re-priced 4 October 2026.

The Companion: A Cross-Platform Income Tracker

There’s no honest cross-platform provider-earnings comparator. Every tool that exists is a siloed per-vendor calculator: the Vast.ai host earnings calculator, the Akash Provider Earn calculator, the Render profitability calculator, the io.net worker dashboard, the Morpheus MOR calculator. GPU-mining calculators like WhatToMine are for proof-of-work coins, not AI compute. DePIN directories list projects but never compute net provider earnings. That gap is the tool worth building, and it ships alongside this article.

The design is a parametric model: you set GPU model, electricity price, and a utilisation assumption, and it returns net monthly income across platforms, with cash payouts and token payouts visually separated and a token-price slider for MOR and the rest. The opinionated choices are deliberate. Decentralised platforms come first, Vast and RunPod sit in as the cash benchmark, token-price sensitivity is made explicit instead of hidden, and a live Morpheus column is fed by our own on-chain indexer. That live column is the part nobody else can copy.

Two honesty constraints keep it from lying to you. Utilisation is unobservable per user, so the tool is a parametric model with a user-set utilisation assumption, not a live earnings oracle, and it says so. And the model decays as rates, fees, and emission schedules change, so it carries a data-refresh discipline: token prices via API are trivial, marketplace rates are the part that needs tending. The article’s numbers are the tracker’s backend, which means when one moves, so does the other.

The tracker also lives on its own at /tools/gpu-income-tracker if you want to bookmark or share it directly.

So, is renting your GPU worth it? On owned hardware you’d otherwise leave idle, on cheap power, at decent utilisation, yes, as passive income measured in tens to low hundreds of dollars a month. As a business bought on purpose, only with datacentre cards on sub-$0.10/kWh power. And if you’re choosing a token network, choose it with the exposure mapped, sell on receipt unless you actively want the bet, and check the settlement data before you trust the emissions headline. That’s the whole game.