Renting Your Home GPU to AI Companies in 2026 — Can You Actually Make Money?
The line “rent your home GPU during idle hours and earn from AI companies” has circulated since 2022. By 2026 the question is whether it really works. How much do you net? When do you lose money? This article cracks open the actual revenue models on the three main DePIN projects — RNDR, Akash and io.net — across typical cards like the RTX 4090, 5090 and A6000, after subtracting electricity, depreciation and opportunity cost. If you already read the DePIN beginner guide, treat this as the hands-on financial math companion.
How the three networks actually differ
RNDR, Akash and io.net keep getting grouped together but they take very different work.
| Project | Main workload | Pricing | Entry barrier |
|---|---|---|---|
| RNDR (Render) | 3D rendering, video post, Stable Diffusion inference | RNDR tokens per job | Requires node approval |
| Akash | Containerized general compute (incl. LLM inference and training) | USD/hour (USDC settled) plus AKT incentives | Docker compose deploy |
| io.net | Distributed GPU clusters, AI training | USDC/hour (dynamic market) plus IO incentives | Low, online in hours |
RNDR is rendering revenue. Akash is cloud compute revenue. io.net is AI compute revenue. The customer profiles and revenue stability differ a lot.
Per-card real-world earnings (early 2026)
Monthly earnings across the dominant consumer and workstation cards:
RTX 4090 (24GB):
- RNDR: about $80-130/month (depends on job arrival rate)
- Akash: about $90-170/month (assuming 30-50% utilization)
- io.net: about $110-220/month (peak demand windows)
RTX 5090 (32GB):
- RNDR: about $130-210/month
- Akash: about $150-280/month
- io.net: about $180-360/month
RTX A6000 (48GB) workstation card:
- RNDR: about $200-340/month
- Akash: about $280-460/month
- io.net: about $340-580/month
RTX 4070 / 4060 and similar mid-range: essentially unprofitable, monthly earnings often below electricity. Their small VRAM (12/8GB) cannot host mainstream LLM inference.
What survives after costs
Gross is not net. The cost stack:
Electricity: a 4090 at full load draws about 450W. Twenty-four hours full load = 10.8 kWh. At US residential $0.16/kWh that is $1.73/day = $52/month. At China residential ¥0.6/kWh it is about $27/month.
Depreciation: a new 4090 costs about $1600, three-year straight-line depreciation = $44/month. If you already own the card for gaming, this is not an incremental cost.
Network and opportunity cost: stable 1Gbps fiber runs $30-50/month. Opportunity cost is real but hard to quantify, since a fully loaded card cannot game or run your own AI projects.
Take a 4090 on io.net earning $180 gross/month at US electricity:
- Gross: $180
- Electricity: -$52
- Network: -$10 (marginal if you already have fiber)
- Depreciation: -$44 (excluded if card is sunk cost)
- Net: $74/month (or $118 excluding depreciation)
China electricity drops electricity to $27, net about $99/month (or $143 excluding depreciation).
| Scenario | 4090 net | 5090 net | A6000 net |
|---|---|---|---|
| US electricity + depreciation | $74 | $130 | $260 |
| US electricity, no depreciation | $118 | $190 | $360 |
| China electricity + depreciation | $99 | $165 | $310 |
| China electricity, no depreciation | $143 | $225 | $410 |
When you lose money
Loss path 1: bought the card specifically to rent at high electricity prices. A new 4090 outrunning its depreciation under US electricity has under 50% odds. Even at China electricity, the 5090 only achieves 12-15% annualized “physical ROI”, which does not crush equities or DeFi LPs.
Loss path 2: utilization assumptions too high. The monthly numbers above assume 50-70% real utilization. Starting in late 2025 DePIN GPU networks entered “many cards, fewer jobs”. On io.net 4090 utilization dropped to 30-35%, so projected revenue shaves about 40%.
Loss path 3: token volatility ignored. RNDR, AKT and IO are all highly volatile, each dropped 50%+ in 2024. If part of your revenue is denominated in those tokens, the gap between book and realized USD can be wide.
Loss path 4: regulation and tax. Renting GPU compute is business income in most jurisdictions and taxable. Token settlement adds crypto tax exposure. Real tax burden often eats 20-30% of net.
Loss path 5: hardware failure. Consumer GPUs running 24/7 at full load fail 3-5x more often than gaming use. A single repair or replacement can wipe half a year of net.
Is it better to just sell the card?
A question many DePIN believers skip. If you sell your 4090 today you get $1100. If you rent it for 12 months and sell, residual value is about $750.
Take “sell now” as the baseline:
- Sell now: $1100 cash, zero cost, zero risk
- Rent 12 months then sell: about $750 residual + 12 x $74 (US with depreciation) = $1638
- Net advantage: $1638 - $1100 = $538/year (about 49% annualized)
Looks good, but the math depends on:
- io.net token price and job arrival rate not declining
- Hardware failure rates staying normal
- You not adding hidden time and ops cost
If tokens halve and utilization halves, net could drop from $538 to below $100, annualized into single digits. At that point selling the card and buying an index ETF wins.
Who should really run a DePIN GPU node
Back to first principles. The fitting profile:
- Already own a 4090/5090/A6000 for other reasons, so rented idle time is incremental zero cost
- Relatively low electricity (most of China, Southeast Asia, Nordic hydro regions)
- Already running AI projects, comfortable with Docker, Linux and CLI, near-zero deployment overhead
- Patient for 6+ months, not expecting monthly stability
- Not relying on this income as primary, treats it as alpha
If three or more of those do not apply, skip the operation or get exposure indirectly by holding RNDR/AKT/IO tokens, which carries its own price risk. Compare with the lighter-hardware path in Grass decentralized bandwidth, which has lower entry but lower per-unit yield too.
Outlook for DePIN GPU in 2026
On a longer horizon, DePIN GPU compute faces three forces:
- Hyperscalers build their own faster: Anthropic, OpenAI and xAI pour billions into H200, B100 and B200 clusters every quarter, compressing the addressable decentralized compute market
- Hardware generational leaps: B100/B200 and Chinese H100 equivalents deliver 2-3x performance per watt, weakening the economics of home 4090s
- New demand emerging: AI inference shifts toward edge and privacy-sensitive scenarios where long-tail customers pay a premium for model isolation on DePIN networks
Combine those forces and DePIN GPU as a category will keep getting squeezed over the next 12-24 months, while in the 3-5 year horizon it survives as an AI compute supply-side complement. The honest framing for 2026 is transition rather than gold rush. Earning $100-300 extra per month from an existing 4090 or 5090 is realistic. Buying a rack of new GPUs hoping to profit at a node operator is almost certainly a losing bet.