GPU servers for Web3: ZK proving, AI agents and inference

GPU servers for Web3 teams mostly run two kinds of work: zero-knowledge proving, and AI, from agents to inference behind dApps. We rent whole NVIDIA GPUs by the month, paid from a prepaid balance that you top up in Bitcoin, Ethereum, Monero, USDT on TRON or Solana. This page covers how to size a GPU for ZK proving, what our Acceptable Use Policy allows, and how crypto top-ups, the bonus and refunds work.

Choosing a GPU for ZK proving

Depending on the proof system, provers spend much of their time in multi-scalar multiplication (MSM), number-theoretic transforms (NTT) and hashing. That is integer and modular arithmetic over finite fields, so CUDA cores, memory bandwidth and VRAM matter, while the FP8 and FP4 tensor-core figures quoted for AI say little about proving speed. Several proving stacks ship CUDA backends, RISC Zero, SP1 and the ICICLE library among them; check your prover's documentation for supported GPUs, minimum VRAM and host RAM.

VRAM caps the largest circuit or segment a GPU proves in one pass, while host RAM and CPU cores matter for witness generation and recursion. The RTX 5090 suits one prover with modest memory needs, the RTX PRO 6000 runs larger circuits or several provers side by side, and the H100 adds HBM bandwidth and the most RAM of the three. Prover performance varies widely by stack and GPU generation, so rent one GPU for a month, measure proofs per hour on your own workload, then scale the configuration with the best cost per proof. Our RTX 5090 vs RTX PRO 6000 comparison helps with the first choice.

Running provers around the clock

Provers that serve a rollup or sell capacity on a proving network usually run 24/7, which is where a flat monthly price fits: an RTX 5090 at $449 a month works out to about $0.62 per GPU-hour when it proves around the clock, and an RTX PRO 6000 at $1,390 to about $1.90. A few practices keep them healthy:

• Run the prover under systemd or in a container with a restart policy, and pin the CUDA and driver versions it was built against.
• Watch GPU memory, temperature and errors with nvidia-smi or NVIDIA DCGM, as in our GPU monitoring guide, and alert on stalled jobs, not only on crashes.
• Keep signing keys and wallet credentials off the prover where you can; it only needs the data it proves.
• If the prover talks to a coordinator or sequencer on another of your servers, ask for private networking between them.

What you can run on GPU servers for Web3

• Allowed: zero-knowledge proving and other legitimate Web3 workloads, such as AI agents, model inference and on-chain data analytics.
• Allowed only with our written agreement: proof-of-work cryptocurrency mining, and reselling or sublicensing the service.
• Prohibited: attacks and intrusion, malware, phishing, spam, fraudulent schemes, and anything that breaches export control or sanctions laws.

Validators and RPC nodes are usually bound by CPU, disk and network rather than GPU, so a GPU server is rarely the right tool for them. The full rules are in our Acceptable Use Policy.

AI agents and inference for Web3 apps

Self-hosting a model keeps prompts, wallet context and user data on a server dedicated to you, rather than sending them to a third-party model provider. Size it like any inference server: weights take roughly parameters × bytes per parameter, plus room for the KV cache.

• An 8B model in BF16 needs about 16 GB, which leaves half of an RTX 5090's 32 GB for cache.
• A 70B model in FP4 needs about 40 GB and runs on an RTX PRO 6000 with room for concurrency.
• A 70B model in FP8 needs about 70 GB: tight on an H100, comfortable on an H200.

vLLM and SGLang serve these models with continuous batching, and single-GPU plans include 20 TB of outbound traffic a month with no egress fees. More sizing detail is on our page about GPU servers for LLM inference.

Paying in crypto, and how refunds work

• Bitcoin BTC
• Ethereum ETH
• Monero XMR
• Tether USDT On TRON (TRC-20)
• Solana SOL

Prices and your account balance are in US dollars. You top up the balance with any of these coins, from $100 to $10,000 per top-up: the crypto amount is calculated from the USD amount at an exchange rate locked for a limited time, the sender pays network fees, and the top-up is credited automatically once the network confirms the payment. An underpaid top-up is credited once the remainder arrives. Card payments are coming soon. More on our pay with crypto page.

Every full $100 in one top-up adds 5% in bonus credit, up to 50% from $1,000, so a $1,000 top-up gives $1,500 of credit. Bonus credit is spent after paid credit, has no cash value and is not refundable. Servers and their monthly renewals are paid from the balance, with an invoice for every server payment and renewal and a receipt for every top-up in your dashboard.

Unused paid credit can be refunded from the Billing page of your dashboard at its US dollar value, in USDT on TRON (TRC-20) or in the cryptocurrency you used, to a wallet address you control, within 10 business days of approval; see our refunds help page. Most orders need only an account and a funded balance. We may ask for verification for large orders, for fraud or sanctions screening, or when the law requires it.

Getting started

On the deployment page, choose a configuration, set up the OS image, optional preinstalled software and your login (root password or SSH key), and create a free account or log in, all in one form. You pay the first month from your balance; if it is short, you top up in the coin you prefer in the same form, and the order is held for 3 hours and paid automatically as soon as the top-up is credited. Root SSH access is delivered in under 10 minutes after payment, with the IP and SSH details in your dashboard. Before you order, answers are in our help center, docs, GPU pages and calculators. Our Ubuntu images include NVIDIA drivers, CUDA, Docker and the NVIDIA Container Toolkit, so you can build a CUDA prover without extra driver work.