How to Earn EUR 10–50 Per Day with Your Idle Mac GPU (and NVIDIA Too)

By Wingston Sharon | March 2026

Most high-end GPUs are idle more than 20 hours a day. A Mac Studio with an M2 Ultra chip sits on your desk doing nothing for most of its life. An NVIDIA RTX 4090 inside a gaming rig runs at full capacity maybe 3–4 hours in the evening. The rest of the time, it's generating heat and drawing standby power for nothing.

There's now a way to change that.

Agentosaurus runs a distributed AI model evaluation network — essentially a crowd-sourced alternative to single-vendor benchmarking platforms like LMSYS Chatbot Arena. Instead of one organization running all evaluations on centralized hardware, we distribute evaluation jobs to a network of contributor machines and aggregate results using consensus logic.

We pay contributors in $AGENTO tokens for completed evaluations. Based on current token pricing and job rates, that works out to EUR 10–50 per day for hardware that would otherwise be completely idle.

This article covers what's required, how the earnings math works, and how to set up in about 5 minutes.

Why Your GPU Has Value That's Not Being Captured

Let me be specific about the hardware situation first.

A Mac Studio M2 Ultra (192GB unified memory) costs roughly EUR 4,500. It can run Llama 3.1 70B at about 15 tokens/second. On any cloud platform that offers equivalent inference capacity — something close to an A100 40GB — you'd pay EUR 2.80–4.50/hour. The Mac does the same work.

For most Mac owners, this machine is used for software development, design work, video editing. During those activities, the GPU is probably 30–60% utilized. Outside working hours — nights, weekends, long meetings where the Mac is just sitting open — utilization drops to near zero.

The same applies to NVIDIA cards. An RTX 4090 with 24GB VRAM can handle all Tier 1 and Tier 2 evaluation jobs in our network. During gaming, it's at 95%+ utilization. During everything else: idle.

The distributed compute model captures this wasted capacity. When you're not using your hardware, we're using it. When you need it back, the agent pauses jobs.

The Three Evaluation Tiers

Our job queue distributes three categories of evaluation work, matched to hardware capability:

Tier 1 — Basic Evaluations (Mac M1/M2, GTX 1080 8GB+)

Reward: 100 $AGENTO tokens per completed evaluation
Time: 30–45 minutes per job
Tasks: HellaSwag, ARC-easy, Winogrande, BoolQ
Hardware minimum: Mac M1 (8GB unified memory) or NVIDIA GTX 1080 (8GB VRAM)

Tier 1 covers standard language comprehension and commonsense reasoning benchmarks. These are the "easy" evaluations — smaller model sizes, shorter context windows, simpler prompts. The hardware requirements are correspondingly modest.

Tier 2 — Standard Evaluations (Mac M3, RTX 4090 24GB)

Reward: 250 $AGENTO tokens per completed evaluation
Time: 1–1.5 hours per job
Tasks: MMLU, TruthfulQA, ARC-challenge, GSM8K
Hardware minimum: Mac M3 (36GB unified memory) or NVIDIA RTX 4090 (24GB VRAM)

Tier 2 evaluations use larger models (30B–70B parameter range) and more demanding benchmarks. MMLU covers 57 academic subjects. GSM8K tests mathematical reasoning. These require enough VRAM to hold the model plus evaluation context in memory simultaneously.

Tier 3 — Advanced Evaluations (Mac M3 Max/Ultra, A100-class)

Reward: 500 $AGENTO tokens per completed evaluation
Time: 2–3 hours per job
Tasks: HumanEval, MBPP, BigBench-Hard, MATH, GPQA
Hardware minimum: Mac M3 Max (128GB) or NVIDIA A100/H100

Tier 3 is reserved for the hardest evaluations — code generation, competition math, graduate-level science questions. These models are large (70B+), the prompts are long, and the evaluation logic is complex. Not many machines can handle it, which is why the reward is highest.

Earnings Calculator: Being Honest About the Math

The EUR 10–50/day range is real, but the actual figure depends on:

1. Your hardware tier (determines which jobs you qualify for)
2. How many hours you contribute (how long the agent runs)
3. Job queue depth (whether there's work available)
4. $AGENTO token price (determined by market, launching Q2 2026)

Here's a transparent breakdown using current job rates and estimated token price:

Assumed token price: EUR 0.15 per $AGENTO (based on presale round; market price TBD)

Wait — those numbers look high. Let me be more conservative and realistic.

Realistic assumptions:
- Job queue doesn't run at 100% capacity (beta network, limited jobs)
- You contribute 8 hours/day while sleeping or at work
- Token price at launch is closer to EUR 0.05 (more conservative)
- Network starts sparse; grows over time

The EUR 10–50/day headline comes from the moderate-to-optimistic range once the network reaches reasonable job depth. During early beta, expect the lower end.

Important caveat: $AGENTO token price at launch (Q2 2026) is unknown. Token prices are volatile. The EUR figures above are estimates based on current presale pricing and should not be treated as guaranteed returns. Treat this as compute contribution with token compensation, not a guaranteed income stream.

What the Agent Actually Does

The b9agent is a lightweight daemon that:

1. Registers your hardware — GPU UUID, VRAM, compute capability
2. Receives job assignments from the evaluation queue
3. Downloads the model (one-time, cached after first run)
4. Runs the benchmark in an isolated environment
5. Signs the result with a cryptographic key tied to your hardware
6. Submits to the network and receives token reward

The agent uses Ollama under the hood for model loading. It runs in the background at low priority — your machine's foreground applications take precedence. If you start a GPU-intensive task (video export, gaming, local inference), the agent automatically yields resources.

You don't need to run a node 24/7 or maintain uptime commitments. Contribute when you want. Stop when you need the compute back.

Setup: 5 Minutes

Step 1: Register at admin.agentosaurus.com

Create an account and register your hardware:
- Your machine will receive a hardware fingerprint (GPU UUID + VRAM check)
- You get a 1-hour TTL authentication token
- Hardware verification takes 2–3 minutes

Step 2: Configure credentials

# Add to your shell profile (~/.zshrc or ~/.bashrc)
export BETA9_TOKEN="jMmOVFQ..."       # From registration
export GATEWAY_HOST="100.72.101.23"    # Tailscale gateway IP

Step 3: Download and run the agent

# Download the agent binary
curl -O https://registry.agentosaurus.com/b9agent
chmod +x b9agent

# Start in background
./b9agent --token $BETA9_TOKEN --pool-name external &

The agent connects to the Beta9 gateway over Tailscale (WireGuard encrypted). Your machine joins the distributed evaluation pool. Job assignments start flowing within minutes if the queue has work available.

Step 4: Monitor your earnings

# Check agent status
./b9agent status

# View completed jobs and pending rewards
./b9agent history --days 7

Technical Architecture: Why This Actually Works

The naive concern with distributed AI evaluation is trust: how do you know the results are accurate if you can't verify the evaluators?

We solve this with a multi-layer verification system:

Hardware attestation: Each evaluator's GPU UUID is cryptographically bound to their results. You can't fake GPU hardware in a way that passes our verification without physical access to the machine. This prevents cloud VM spoofing (someone spinning up 100 VMs to game rewards).

Consensus requirement: Every evaluation result requires agreement from at least 5 independent evaluators. Results are only accepted if 80% of evaluators agree within a 2% tolerance. A single bad actor can't corrupt an evaluation — they'd need to control 20%+ of the evaluating machines simultaneously.

Halving reward schedule: The first evaluator to complete a job earns 100% of the reward. The second evaluator earns 50%. The third earns 25%. This incentivizes speed while maintaining the redundancy required for consensus. You want to be first, but you can't be first unless other evaluators confirm your result.

Result verification: Every evaluation run produces a deterministic output given the same model, prompt, and configuration. Our verification layer cross-checks outputs for statistical consistency. Outliers trigger manual review.

This architecture means the network is useful from the start — even with 50 contributors, you get statistically reliable evaluation results. As the network grows to 500+ evaluators, reliability improves further.

Hardware Requirements (Honest)

Mac requirements:
- Minimum: Mac Mini M1 (8GB unified memory) for Tier 1 only
- Recommended: Mac Mini M3 (24GB+) for Tier 1 and Tier 2
- Best: Mac Studio M2 Ultra / Mac Mini M3 Pro (96GB+) for all tiers

NVIDIA requirements:
- Minimum: GTX 1080 (8GB VRAM) for Tier 1 only
- Recommended: RTX 3090/4090 (24GB VRAM) for Tier 1 and Tier 2
- Best: A100/H100 for all tiers including Tier 3

Network requirements:
- Upload: 10 Mbps minimum (for submitting results and telemetry)
- Stability: Stable connection preferred; agent handles reconnection automatically
- VPN: Tailscale handles network connectivity; no open ports required on your end

Storage requirements:
- 100GB free disk space recommended
- Models are cached after first download; subsequent jobs of the same model family skip the download

Operating system:
- macOS 14+ (Sonoma) for Mac contributors
- Ubuntu 22.04+ or Windows 11 (WSL2) for NVIDIA contributors
- Docker available as fallback for other Linux distributions

Beta Access: Currently Limited

We're running a closed beta with 50 initial contributor slots. This is intentional — we want to validate the evaluation pipeline, token distribution mechanism, and consensus logic before opening to a larger network.

Beta contributors get:
- Priority allocation: First access to higher-tier jobs when they become available
- Locked token price: Beta contributors purchase $AGENTO at presale price (significantly below launch price)
- Network governance rights: Early contributors vote on evaluation standards and reward parameters
- Whitepaper access: Full tokenomics documentation before public launch

If you have hardware that meets Tier 2 or Tier 3 requirements, we prioritize your application.

Apply: admin.agentosaurus.com/gpu-contributor

What Happens to the Results

This isn't compute contribution for its own sake. The evaluations power the public Agentosaurus leaderboard — a distributed, tamper-resistant alternative to existing AI benchmarks.

Every model that gets evaluated through our network:
- Receives a consensus score across 5+ independent evaluators
- Has hardware-attested results (provably not from cloud VMs gaming the benchmark)
- Joins a public database of reproducible evaluation results

For the AI research community, this addresses a real problem: current benchmarks can be gamed by the model providers who submit to them. When Anthropic submits Claude to LMSYS Chatbot Arena, the evaluation is run on infrastructure that Anthropic knows about. When your Mac M3 evaluates Claude at 2am, Anthropic doesn't know that evaluation is happening. The results are harder to engineer.

You're contributing to infrastructure that makes AI benchmarking more trustworthy. The token rewards are how we compensate you for that contribution.

Common Questions

Q: Will this slow down my Mac during work hours?

No. The agent runs at background priority and yields immediately when foreground applications need resources. You can also configure working hours:

./b9agent --schedule "09:00-18:00=pause" --schedule "18:00-09:00=active"

Q: What if I lose internet connectivity mid-evaluation?

The agent handles disconnection gracefully. Partially completed evaluations are held locally and submitted when connectivity returns. You don't lose credit for work already done.

Q: Can I run this on a cloud VM?

You can, but the economics don't work. Cloud compute costs more than you'd earn from the network. The model is designed for idle hardware that has already been paid for.

Q: Is this legal?

Yes. You're contributing compute to a distributed network in exchange for token compensation. This is similar to running a mining node or participating in distributed computing projects like Folding@home, except you're compensated in tokens rather than points. The legal status of $AGENTO tokens in your jurisdiction depends on your local regulations — we are not a securities issuer, and token compensation is for compute services rendered, not investment returns.

Q: When does token trading start?

Q2 2026, pending MiCA compliance confirmation. Tokens earned during beta are held in escrow until the token launch. Early contributors have preferred access to the initial liquidity pool.

The Bigger Picture

This isn't just a side hustle.

The distributed GPU network is the infrastructure layer that makes Agentosaurus economically viable at scale. We can't build a 1,000-node evaluation network with owned hardware — the capital requirements are too high. But we can build it by compensating the thousands of people who already own high-end hardware and aren't using it 20 hours a day.

If this works, you end up with:
- An AI evaluation network that's genuinely decentralized (no single organization controls results)
- A leaderboard that's harder to game than anything currently available
- Infrastructure with EU data residency by default (contributors are in EU, results stay in EU)
- Economic alignment between infrastructure contributors and platform users

The EUR 10–50/day is real. The network effects are the larger bet.

Apply for beta access: admin.agentosaurus.com/gpu-contributor

Read the tokenomics whitepaper: See our MiCA compliance analysis

Questions: contributors@agentosaurus.com

Wingston Sharon is the founder of Agentosaurus. Views expressed are his own. $AGENTO token price estimates are projections based on presale pricing and should not be interpreted as investment advice. Cryptocurrency tokens are high-risk instruments; never invest more than you can afford to lose entirely.