FAQ

Honest answers to the questions that decide whether this project is worth your watts.

Is this profitable on my machine?

Depends on the machine and your electricity price, and the honest baseline is modest. On an RTX 4090 at $0.10/kWh (July 2026): Salad-class inference ~$3.5/day (static estimate), Vast.ai rental ~$2.9/day (p25 asks, 55% utilization assumed), the best mineable kawpow coin ~$0.05/day, RVN/ERG/ETC and NiceHash GPU sale negative. Older or smaller GPUs scale down from there. Run bun run src/agent-cli.ts plan and read the number: the tool exists to give you that answer, not to promise one.

So mining is dead on consumer GPUs?

At steady state, essentially yes: inference and rental crush it. The nearly-zero NiceHash payings on KAWPOW/AUTOLYKOS/ETCHASH are a real market signal, GPU altcoin hash demand collapsed. Mining only re-enters the plan during emission windows (new PoW listings, hash-bid spikes like the Qubic pattern, profitability spikes), which is exactly what the watcher detects. That optionality is the reason mining support exists at all.

Why does the plan sometimes say IDLE?

IDLE means no destination nets more than $0.01/day after electricity: powering the compute would lose money. It is a recommendation with teeth: the executor honors it by not starting (or by stopping) the compute workload. A machine at IDLE can still earn through its stack (IP-slot and passive networks), which costs approximately nothing to run.

What do tiers A/B/C/D mean?

They grade how real a network's number is. A: measured cash payouts, the only real number. B: token listed with deep DEX liquidity, priced live by contract address. C: listed but thin, or CEX-only (which never counts as liquid here). D: unlisted pre-TGE points, pure expected value; anything tier D in a plan is explicitly marked speculative. Details and formulas in docs/scorer.md.

Are the scorer numbers trustworthy?

They are honest estimates, and the system is built to say so. Every non-measured figure in the registry is commented as a seed with its refinement path, and every score carries flags: pricing path (address vs symbol vs manual CEX with verification date), realizable liquidity percentage, EV parameters, ToS and KYC risks. Comparables are calibrated on measured history, including the failures (Nodepay -99% post-TGE is a standing bear comparable). If you want certainty, only tier A gives it.

Will running this get me banned from networks?

It can if you farm carelessly. Several bandwidth networks (Grass is the canonical case) enforce 1 device / 1 IP / 1 account and ban automation without appeal. The scorer prices this in via the banToS risk multiplier, and the planner attaches an explicit "respect 1 device/1 IP" warning to any stacked network with high ban risk. The stack fills at most your configured ipSlots: the tool never suggests multi-boxing an IP.

Does it bundle proxyware or sell my bandwidth behind my back?

No. Nothing runs without being in the plan, the plan is printed before execution, and v0 ships dry-run by default: drivers log the exact command or API action instead of acting until you set dryRun: false and provide the binary or key. Per-network ToS respect (1 device/1 IP, no proxyware bundling by default) is a stated design constraint of Phase 1.

Can it actually switch workloads today, or only print plans?

Both, with the default being print. The miner driver can really spawn xmrig or t-rex once you configure binary paths and wallets and disable dry-run; the Vast rental driver can really list, reprice (p25 minus 2%) and delist your machines once vastApiKey is set. Inference and DePIN drivers are log-only in this phase. Everything else, benchmark seeds included, degrades to auditable logs.

What hardware is supported?

NVIDIA consumer GPUs from GTX 1660 Super through RTX 5090 (seed benchmarks in src/profiles.ts), the Ryzen 9 7950X for RandomX, and Apple Silicon (stack-only: no CUDA, RandomX not profitable). Unknown NVIDIA GPUs fall back to a deliberately conservative generic profile; unknown x86 CPUs likewise. Detection works on macOS and Linux.

Is there a token?

Not yet. The roadmap targets a Virtuals ACP agent (Revenue Network: USDC revenue plus buybacks) with a launch only once revenue is verifiable on-chain. Nothing is live, nothing is promised: no token exists today and none of the current tooling depends on one. Treat any contract claiming otherwise as fake.

What is the buyback model?

The intended Phase 3 design: the agent earns real USDC revenue (settlement on Base), and part of that revenue funds buybacks under the Virtuals Revenue Network mechanics. It is deliberately sequenced after measurable revenue, not before.

I run a GPU farm. Is there a B2B offer?

Phase 2 is a pilot with 2-3 GPU farms: fee charged on measured uplift only, USDC settlement on-chain (Base), and a public revenue dashboard. If routing does not raise your net revenue, you owe nothing. Contact: open an issue on the repo or reach out through the project's X account for pilot onboarding.