PassivePin

COMPUTE · Last updated: · 11 min read · by Jordan, PassivePin Editorial

GPU compute: Render, Akash, io.net

How AI startups rent your GPU, what realistic monthly earnings look like, and the difference between a render job and an inference job. Includes a cost-recovery calculator.

GPU compute is the highest-earning DePIN vertical, but it is also the highest-friction. You need a recent NVIDIA GPU, a stable internet connection, and a willingness to leave a power-hungry machine running 24/7. The upside: $50 to $500 per month per GPU, depending on the project and the hardware. The downside: hardware depreciation, electricity costs, and the risk that the project you picked turns out to be paying you in tokens that drop 80%.

What GPU networks actually do

The customers of GPU networks are AI startups, VFX studios, researchers, and developers who need GPU compute on demand. They do not want to commit to a year-long cloud contract; they want a few hours of an A100 today, a few days of a 4090 next week, and a few minutes of a H100 to test a new model. The traditional answer is AWS or Vast.ai, but the cost is high and the availability of high-end GPUs is often constrained.

DePIN GPU networks aggregate idle consumer and prosumer GPUs into a single marketplace. When a customer requests compute, the network routes the job to a node that has the right hardware available. The network pays the node operator a share of the customer revenue, in tokens. The customer pays less than AWS; the operator earns more than renting out via Vast.ai; the network takes a cut. The flywheel works as long as the demand keeps growing and the token does not crash.

There are two kinds of jobs: render jobs (Render Network's original product, which is 3D rendering for VFX, gaming, and architecture) and inference jobs (running a trained AI model on new inputs, which is the bulk of demand in 2026). Inference jobs pay more per hour but require more VRAM; a 4090 with 24GB can run most open-source models, but a frontier-model customer will want an A 100 with 80GB or an H100. The mix of jobs routed to your machine depends on what you have.

The three networks worth knowing

Render Network. The most established GPU network, founded in 2017. Render started as a pure 3D rendering network (OctaneRender, Redshift) and pivoted to AI inference in 2024. The $RNDR token is the longest-running GPU-network token and the most liquid. The customer base is split between VFX studios and AI startups, with the AI side growing faster. A 4090 on Render earns $100-$300/month at $RNDR's current price; a 3090 earns $50-$150. The customer demand is real and verifiable; the VFX studios have been around for years.

Akash. A more general-purpose cloud compute network that supports both GPU and CPU jobs. Akash is Cosmos-based, which means the transaction costs are very low but the customer base skews to crypto-native projects. Akash is a good fit for someone who already runs a homelab and wants to monetize excess capacity, including CPU-only nodes for non-AI workloads. Earnings on Akash are typically lower than Render per dollar of hardware, but the demand is steadier.

io.net. A newer network focused exclusively on AI training and inference, with an aggressive customer-acquisition strategy. io.net has signed contracts with several Y Combinator AI startups; the demand is real but more concentrated than Render's. Earnings are highly variable depending on your GPU tier and location. io.net is the highest-risk, highest-reward of the three; the $IO token is newer and more volatile.

You can run all three simultaneously on the same machine, but the resource contention will hurt you. We recommend picking one primary network and using the others as overflow. Render is the safest primary for most operators; io.net is the highest-yield for a hobbyist who can tolerate volatility.

The cost-recovery calculator

A simple model for whether a GPU rig is worth it.

For a single 4090 purchased at $1,800 in mid-2026, running 24/7 on Render, the numbers look like this:

So the math is positive in the base case. But the base case assumes $RNDR stays at its current price. A 50% drop in $RNDR pushes the payback to 22 months and makes the math marginal. A 70% drop (which happened in 2022) makes the math negative; you would have been better off holding the $1,800 in stables and earning yield.

The conservative operator's approach: cash out 50% of every payout into a stablecoin, hold the rest in $RNDR. If $RNDR drops, the stablecoin half cushions you. If $RNDR rises, the held half gives you upside. Over 2-3 years, this approach has historically outperformed both "cash out everything" and "hold everything" by a meaningful margin.

The hardware question

For an operator considering buying a new GPU for DePIN, here is the current ranking by earnings-per-dollar-of-hardware, based on real 30-day data from the directory:

  1. RTX 4090 (24GB) — best balance of price, VRAM, and earnings-per-watt. The default choice for a new build.
  2. RTX 3090 (24GB) — older, cheaper, similar VRAM. Slightly lower earnings but better payback if bought used at $700-$900.
  3. RTX 4080 (16GB ) — newer, lower power draw, but only 16GB VRAM limits the inference jobs you can run.
  4. RTX A5000 (24GB) — workstation card, lower earnings per hour but very reliable 24/7.
  5. Anything with less than 12GB VRAM — not worth it for AI inference; only viable for 3D rendering or simple CPU-offload jobs.

A100, H100, and other data-center cards are not realistically available to individual operators at retail prices. If you are a small data center with access to those cards, the math is different (and much better), but you are reading a different article.

When to skip

Skip GPU compute if any of these are true: electricity is expensive where you live (above $0.25/kWh makes most rigs unprofitable); you cannot tolerate 24/7 fan noise (a GPU rig under load is loud); you do not have stable high-speed internet (jobs that drop mid-render are penalized); you are uncomfortable with the volatility of GPU-network tokens (a 60% drop in $RNDR is possible and historically has happened). For everyone else, a single 4090 is the highest-earning entry point into DePIN.

Continue learning: Storage networks: Filecoin, Arweave, and the long tail →

© 2026 PassivePin. See also: about · methodology · FAQ · editorial policy.