# CoreGPU

> CoreGPU rents dedicated NVIDIA and AMD data-center GPU servers by the month: 1 to 8 GPUs per server, one flat price per GPU at least 25% below the market median of specialist GPU clouds, paid in Bitcoin, Ethereum, Monero, USDT (TRC-20) or Solana.

Website: https://coregpu.io/. Full text of the site for language models: https://coregpu.io/llms-full.txt. Live stock and prices as JSON: https://coregpu.io/api/v1/gpus. The facts below are generated from the live catalog (2026-11-05 20:58 UTC).

Languages: every page exists in English (https://coregpu.io/), Simplified Chinese (https://coregpu.io/zh, 简体中文) and Russian (https://coregpu.io/ru, русский), with the same prices and stock; prefix any path with /zh or /ru (https://coregpu.io/zh/pricing, https://coregpu.io/ru/gpus/h100).

## Key facts
- What it is: CoreGPU (coregpu.io) is a GPU cloud provider that rents dedicated NVIDIA and AMD data-center GPU servers by the month, with 1 to 8 GPUs per server.
- Price position: Every CoreGPU price is set at least 25% below the market median, the public on-demand price of the same GPU at specialist GPU clouds multiplied by 730 hours (checked November 4, 2026).
- GPU models: GPUs available: NVIDIA B300, NVIDIA B200, AMD MI355X, AMD MI350X, NVIDIA H200, AMD MI325X, AMD MI300X, NVIDIA H100, NVIDIA RTX PRO 6000, NVIDIA L40S.
- Prices: Prices run from $849 per GPU per month (NVIDIA L40S, about $1.16 per GPU-hour) to $4,899 (NVIDIA B300); the price per GPU is the same at every server size, so an 8-GPU server costs 8 times the GPU price.
- Billing: Billing is monthly from a prepaid US-dollar balance: the first month is paid at order, the month starts at delivery, each server renews at the start of the next month and can be stopped at the end of any month, with no setup fee and no cancellation fee. The price is locked for as long as the server keeps renewing.
- Payment: CoreGPU is paid in cryptocurrency only: Bitcoin (BTC, Bitcoin network), Ethereum (ETH, Ethereum mainnet), Monero (XMR, Monero network), USDT on TRON (TRC-20), Solana (SOL, Solana mainnet). There is no credit card payment. Top-ups go from $100 to $100,000 and earn bonus credit, +5% from $100 up to +30% from $1,000.
- Account: An account needs an email address and a password; two-factor authentication is available, and a password cannot be reset. No identity document is requested.
- Hardware: GPUs are dedicated to one customer, never shared, partitioned or time-sliced. 8-GPU SXM and OAM servers have the full NVLink or Infinity Fabric mesh; every server has root SSH access and local NVMe storage.
- Software: Servers start from Ubuntu 24.04 or a ready image with the NVIDIA driver and CUDA 13, or ROCm 7 on AMD, with PyTorch and vLLM.
- Delivery: In-stock servers are delivered with root SSH access in under 60 seconds after payment. Stock comes in fixed allocations with live counts; sold-out models show the date of their next allocation.
- Support: Support runs on tickets in the CoreGPU console for customers with an active server, answered by an engineer within one business day. There is no email address, phone line or contact form.
- SLA: The Service Level Agreement commits to 99.9% monthly availability of power and network for each server, with service credits of 10%, 25% or 50% of the monthly fee below 99.9%, 99% and 95%.
- Refunds: An order can be cancelled before delivery for a full refund to the balance. Paid balance does not expire and is not refundable except where the law requires it or if CoreGPU ends the service or closes an account without cause; bonus credit cannot be withdrawn.
- Best for: CoreGPU suits workloads that keep a GPU busy most of the month: training and fine-tuning runs, always-on inference, research and development nodes. For a few hours of testing, hourly clouds cost less: monthly billing wins above roughly 540 hours of use per month.
- Machine-readable data: Live stock and prices are public as JSON at https://coregpu.io/api/v1/gpus (no key, CORS enabled), described in OpenAPI at https://coregpu.io/api/v1/openapi.json.
- Stock right now: 204 GPUs free across 8 of 10 models.

## GPUs and monthly prices
Price per GPU per month in USD (taxes excluded), the same at every server size; market = median public on-demand price x 730 h.
- [NVIDIA HGX B300](https://coregpu.io/gpus/b300): 288 GB HBM3e, 8 TB/s, FP8 4.5 PF, FP4 13.5 PF, NVLink 5 · 1.8 TB/s, SXM · 8-GPU HGX. $4,899/GPU/month (about $6.71/GPU-hour; 8-GPU server $39,192/month; market $6,552, 25% less). Stock: only 9 left. Best for: High-throughput FP4/FP8 reasoning inference and large fine-tunes.
- [NVIDIA HGX B200](https://coregpu.io/gpus/b200): 180 GB HBM3e, 8 TB/s, FP8 4.5 PF, FP4 9 PF, NVLink 5 · 1.8 TB/s, SXM · 8-GPU HGX. $4,449/GPU/month (about $6.09/GPU-hour; 8-GPU server $35,592/month; market $5,979, 25% less). Stock: sold out, next allocation 2026-11-20. Best for: Mainstream pre-training and fine-tuning, mixed AI and FP64 science.
- [AMD Instinct MI355X](https://coregpu.io/gpus/mi355x): 288 GB HBM3E, 8 TB/s, FP8 5 PF, FP4 10.1 PF, Infinity Fabric · 1.08 TB/s, OAM · 8-GPU UBB. $2,449/GPU/month (about $3.35/GPU-hour; 8-GPU server $19,592/month; market $3,285, 25% less). Stock: 21 available. Best for: Memory-bound LLM training and inference on ROCm, strong FP64.
- [AMD Instinct MI350X](https://coregpu.io/gpus/mi350x): 288 GB HBM3E, 8 TB/s, FP8 4.6 PF, FP4 9.2 PF, Infinity Fabric · 1.08 TB/s, OAM · 8-GPU UBB. $2,249/GPU/month (about $3.08/GPU-hour; 8-GPU server $17,992/month; market $3,022, 25% less). Stock: only 6 left. Best for: 288 GB per GPU for long-context inference and big-batch fine-tuning.
- [NVIDIA H200 SXM](https://coregpu.io/gpus/h200): 141 GB HBM3e, 4.8 TB/s, FP8 1.98 PF, NVLink 4 · 900 GB/s, SXM · 8-GPU HGX. $2,699/GPU/month (about $3.70/GPU-hour; 8-GPU server $21,592/month; market $3,606, 25% less). Stock: 34 available. Best for: LLM fine-tuning and FP8 inference on the deepest CUDA ecosystem.
- [AMD Instinct MI325X](https://coregpu.io/gpus/mi325x): 256 GB HBM3E, 6 TB/s, FP8 2.6 PF, Infinity Fabric · 896 GB/s, OAM · 8-GPU UBB. $2,299/GPU/month (about $3.15/GPU-hour; 8-GPU server $18,392/month; market $3,071, 25% less). Stock: sold out, next allocation 2026-12-02. Best for: Large-memory inference: fit 400B+ models on a single node.
- [AMD Instinct MI300X](https://coregpu.io/gpus/mi300x): 192 GB HBM3, 5.3 TB/s, FP8 2.6 PF, Infinity Fabric · 896 GB/s, OAM · 8-GPU UBB. $1,749/GPU/month (about $2.40/GPU-hour; 8-GPU server $13,992/month; market $2,351, 25% less). Stock: 41 available. Best for: Cost-efficient inference of 70B to 405B models. Lowest price per GB of HBM.
- [NVIDIA H100 SXM](https://coregpu.io/gpus/h100): 80 GB HBM3, 3.35 TB/s, FP8 1.98 PF, NVLink 4 · 900 GB/s, SXM · 8-GPU HGX. $1,999/GPU/month (about $2.74/GPU-hour; 8-GPU server $15,992/month; market $2,716, 26% less). Stock: only 3 left. Best for: Budget training, research and CUDA development on proven hardware.
- [NVIDIA RTX PRO 6000 Blackwell Server Edition](https://coregpu.io/gpus/rtx-pro-6000): 96 GB GDDR7, 1.6 TB/s, FP8 1 PF, FP4 2 PF, PCIe 5.0 x16, PCIe 5.0 · up to 8 per server. $1,199/GPU/month (about $1.64/GPU-hour; 8-GPU server $9,592/month; market $1,599, 25% less). Stock: 52 available. Best for: Inference up to ~70B, image and video generation, rendering.
- [NVIDIA L40S](https://coregpu.io/gpus/l40s): 48 GB GDDR6, 864 GB/s, FP8 0.73 PF, PCIe 4.0 x16, PCIe 4.0 · up to 8 per server. $849/GPU/month (about $1.16/GPU-hour; 8-GPU server $6,792/month; market $1,146, 25% less). Stock: 38 available. Best for: Entry inference, diffusion models and 3D rendering at the lowest monthly price.

## Buying
- [Deploy a server](https://coregpu.io/deploy): pick a GPU and a size, choose an image and an SSH key, create an account in the form, top up if needed; the order goes through once the payment is credited.
- [Pricing](https://coregpu.io/pricing): every GPU at 1, 2, 4 and 8 GPUs, on-demand vs spot vs reserved vs monthly, break-even hours per GPU, billing questions.
- [Pay with crypto](https://coregpu.io/pay-with-crypto): networks, confirmation times and mistakes to avoid for BTC, ETH, XMR, USDT (TRC-20) and SOL; minimum top-up and refunds.
- [All GPUs](https://coregpu.io/gpus): specs, live stock and monthly prices side by side.

## Comparisons
- [B300 vs B200](https://coregpu.io/compare/b300-vs-b200): B300 (Blackwell Ultra) is the better inference GPU: 288 GB instead of 180 GB and 1.5× the dense FP4 throughput, for roughly 10% more per month. B200 keeps strong FP64 and the same 4.5 PF of FP8, so it stays the sensible choice for classic training and mixed HPC work.
- [H200 vs H100](https://coregpu.io/compare/h200-vs-h100): H200 and H100 share the same Hopper compute (1.98 PF FP8), so the question is memory. H200 has 141 GB at 4.8 TB/s against 80 GB at 3.35 TB/s: if your model, batch or context does not fit in 80 GB, the H200 saves you GPUs and is faster on memory-bound inference.
- [MI355X vs B300](https://coregpu.io/compare/mi355x-vs-b300): Both carry 288 GB of HBM at 8 TB/s. MI355X has more dense FP8 (5.0 vs 4.5 PF), keeps strong FP64 and costs about half as much per month; B300 leads on FP4 (13.5 vs 10.1 PF) and runs every CUDA-only kernel. If your stack runs on ROCm, MI355X is the value pick.
- [MI300X vs H100](https://coregpu.io/compare/mi300x-vs-h100): MI300X has 2.4× the memory of an H100 (192 vs 80 GB), more bandwidth and more dense FP8, for a lower monthly price. H100 wins when you depend on CUDA-only libraries or tuned Hopper recipes; for serving large models on ROCm, MI300X is the cost-efficient choice.
- [RTX PRO 6000 vs L40S](https://coregpu.io/compare/rtx-pro-6000-vs-l40s): RTX PRO 6000 Blackwell doubles the L40S memory (96 vs 48 GB), almost doubles the bandwidth and adds FP4 Tensor Cores. The L40S remains the cheapest dedicated GPU for small models and classic rendering; for 30B–70B inference and heavy media work, the RTX PRO 6000 is worth the extra.
- [MI355X vs MI350X](https://coregpu.io/compare/mi355x-vs-mi350x): Same chip family, same 288 GB of HBM3E at 8 TB/s. MI355X runs at up to 1,400 W with liquid cooling for about 9% more FP8; MI350X is air-cooled at 1,000 W and cheaper per month. Memory-bound workloads see little difference; compute-bound training benefits from MI355X.
- [B200 vs H200](https://coregpu.io/compare/b200-vs-h200): B200 is a generation ahead: about 2.3× the dense FP8 of an H200, 180 vs 141 GB and 8 vs 4.8 TB/s. H200 costs much less per month and runs every Hopper recipe unchanged. Pick B200 for training throughput, H200 for cost-efficient fine-tuning and inference.
- [MI325X vs H200](https://coregpu.io/compare/mi325x-vs-h200): MI325X carries 256 GB against 141 GB, more bandwidth (6 vs 4.8 TB/s) and more dense FP8 (2.6 vs 1.98 PF). H200 is the safer pick for CUDA-only stacks; MI325X fits much larger models per GPU on ROCm.
- [B200 vs H100](https://coregpu.io/compare/b200-vs-h100): B200 (Blackwell) has 2.25× the memory of an H100 (180 vs 80 GB), 2.4× the bandwidth (8 vs 3.35 TB/s), about 2.3× the dense FP8 (4.5 vs 1.98 PF) and FP4 Tensor Cores that Hopper does not have, for about 2.2× the monthly price. Per FP8 PFLOPS the two cost about the same: the B200 wins on memory, FP4 and the number of GPUs a model needs, the H100 on entry price and Hopper-tuned code.
- [H100 vs L40S](https://coregpu.io/compare/h100-vs-l40s): The H100 is a training GPU: 80 GB of HBM3 at 3.35 TB/s, 1.98 PF of dense FP8 and NVLink between the GPUs of an 8-GPU server. The L40S is a PCIe inference and graphics card with 48 GB of GDDR6 at 864 GB/s and 0.73 PF of FP8, for less than half the monthly price. Pick the H100 for training, fine-tuning and large models; the L40S for 7B to 13B inference, image generation and rendering on a budget.
- [MI300X vs H200](https://coregpu.io/compare/mi300x-vs-h200): MI300X carries 192 GB of HBM3 at 5.3 TB/s against 141 GB at 4.8 TB/s on the H200, and 2.6 PF of dense FP8 against 1.98 PF, for about 35% less per month. The H200 wins when you depend on CUDA-only kernels, TensorRT-LLM or recipes tuned for Hopper; for serving and fine-tuning large models on ROCm, MI300X gives more memory and more compute per dollar.
- [B300 vs H200](https://coregpu.io/compare/b300-vs-h200): B300 (Blackwell Ultra) has about twice the memory of an H200 (288 vs 141 GB), 1.7× the bandwidth (8 vs 4.8 TB/s), 2.3× the dense FP8 (4.5 vs 1.98 PF) and 13.5 PF of FP4 that Hopper does not have, for about 1.8× the monthly price. It costs less per unit of FP8 compute and fits much larger models; the H200 stays the best value when a model fits in 141 GB and your code targets Hopper.
- [MI355X vs B200](https://coregpu.io/compare/mi355x-vs-b200): MI355X has 1.6× the memory of a B200 (288 vs 180 GB) at the same 8 TB/s, a little more dense FP8 (5.0 vs 4.5 PF) and FP4 (10.1 vs 9 PF), for about 45% less per month. The B200 has the faster GPU-to-GPU link (NVLink 5 at 1.8 TB/s against 1.08 TB/s) and runs every CUDA-only kernel and TensorRT-LLM. If your stack runs on ROCm, MI355X gives more memory and more compute per dollar.
- [MI355X vs H200](https://coregpu.io/compare/mi355x-vs-h200): MI355X is a generation ahead of the H200: about twice the memory (288 vs 141 GB), 1.7× the bandwidth (8 vs 4.8 TB/s), 2.5× the dense FP8 (5.0 vs 1.98 PF) and FP4 that Hopper does not have, for about 9% less per month. The H200 stays the safe choice when your code depends on CUDA-only kernels, TensorRT-LLM or recipes tuned for Hopper.
- [RTX PRO 6000 vs H100](https://coregpu.io/compare/rtx-pro-6000-vs-h100): The RTX PRO 6000 Blackwell has more memory than an H100 (96 vs 80 GB) and FP4 Tensor Cores, for 40% less per month. The H100 has about twice the dense FP8 (1.98 vs 1.0 PF), twice the bandwidth (3.35 vs 1.6 TB/s) and NVLink between the 8 GPUs of a server, where RTX PRO 6000 cards talk over PCIe 5.0. Pick the RTX PRO 6000 for single-GPU inference up to 70B, image and video generation and rendering; the H100 for training and multi-GPU jobs.

## Guides
- [GPU cloud pricing index: November 2026](https://coregpu.io/blog/gpu-cloud-pricing-index-november-2026): Median on-demand prices for B300, B200, H200, H100, MI300X and more across specialist GPU clouds, converted to monthly cost, with how prices moved this year.
- [Which GPU should you rent in 2026? The best GPU for each workload and budget](https://coregpu.io/blog/best-gpu-to-rent-2026): Which GPU to rent for LLM inference, fine-tuning, pre-training, rendering or HPC in 2026: picks by model size and budget, value per dollar, NVIDIA or AMD.
- [What is a dedicated GPU server? Bare metal vs cloud GPUs](https://coregpu.io/blog/what-is-a-dedicated-gpu-server): A dedicated GPU server gives you whole GPUs that no one else shares. How it differs from cloud GPU instances and fractional GPUs, and what to check.
- [Hourly vs monthly GPU rental: when does monthly win?](https://coregpu.io/blog/hourly-vs-monthly-gpu-billing): A plain calculation of when a flat monthly GPU price beats on-demand, spot or reserved billing, with break-even utilisation for H100, H200 and B300.
- [How much GPU memory do you need for LLM inference and fine-tuning?](https://coregpu.io/blog/how-much-gpu-memory-llm): Formulas and tables to size GPU memory for serving and fine-tuning 7B to 405B models in BF16, FP8 and FP4, and which GPU configurations fit.
- [How to validate a new 8-GPU server before a long run](https://coregpu.io/blog/validate-a-new-8-gpu-server): A checklist to run before a long training job: GPU health, ECC errors, NVLink and PCIe topology, NCCL bandwidth and a short burn-in, for NVIDIA and AMD.
- [Why GPU clouds sell out, and how allocations work](https://coregpu.io/blog/why-gpu-clouds-sell-out): Why flagship GPUs keep selling out, what an allocation is, and how CoreGPU publishes live stock and restock dates instead of hiding availability.
- [GPU cloud glossary](https://coregpu.io/glossary): HBM, NVLink, SXM vs PCIe, FP8 and FP4, GPU-hour, break-even and the other terms behind GPU prices.

## Use cases
- [GPU Servers for LLM Pre-Training, Billed Monthly](https://coregpu.io/solutions/training): B300, B200 and MI355X 8-GPU servers for LLM pre-training: up to 2.3 TB of HBM per node, NVLink or Infinity Fabric, a flat monthly price, 25% below market.
- [GPUs for LLM Fine-Tuning: H200, MI350X, B300 Monthly](https://coregpu.io/solutions/fine-tuning): Rent GPUs for LoRA, QLoRA or full fine-tuning by the month: 141 GB to 288 GB per GPU for weights, optimizer states and long sequences, 25% below market.
- [GPU Servers for LLM Inference at a Fixed Monthly Cost](https://coregpu.io/solutions/inference): Serve LLMs on dedicated GPUs with a fixed monthly cost: L40S to B300 servers sized to your model, no surprise bill when traffic spikes, 25% below market.
- [GPU Servers for Research Labs and Universities](https://coregpu.io/solutions/research): Dedicated GPU servers for research labs and universities: no queue, a monthly price that fits a grant, receipts for every payment and 1 to 8 GPUs per node.
- [GPU Servers for 3D Rendering, VFX and Simulation](https://coregpu.io/solutions/rendering): RTX PRO 6000 Blackwell and L40S servers for rendering, VFX, simulation and video generation: RT cores, up to 96 GB per card, billed by the month.
- [8-GPU Servers for AI Labs, Billed Monthly](https://coregpu.io/solutions/ai-labs): Full 8-GPU B300, B200, H200 and MI355X servers for AI labs, billed monthly instead of a multi-year contract, with receipts, a DPA and public prices.
- [GPU Servers for AI Startups: Monthly, No Commitment](https://coregpu.io/startups): GPU servers for AI startups: start with one GPU, grow to eight, resize at every renewal. No 12-month contract, no setup fee, 25% below the market median.
- [Enterprise GPU Servers, Billed Monthly](https://coregpu.io/enterprise): Enterprise GPU servers billed monthly: single-tenant hardware, a DPA, receipts with your company details and public security documentation, 25% below market.

## Docs
- [Quickstart: from checkout to your first GPU job](https://coregpu.io/docs/quickstart): Connect to your CoreGPU server over SSH, check the GPUs and their interconnect, and run your first PyTorch job in a few minutes.
- [Templates and images](https://coregpu.io/docs/images): CoreGPU server templates: vLLM, SGLang, Ollama, PyTorch, JupyterLab, ComfyUI, NVIDIA NGC and more. What each installs, how to open it, how to use your own image.
- [Billing, renewals and cancellations](https://coregpu.io/docs/billing): How billing works at CoreGPU: a prepaid USD balance topped up in BTC, ETH, XMR, USDT (TRC-20) or SOL, top-up bonuses, monthly renewals and stopping.
- [Inventory API reference](https://coregpu.io/docs/api): Public JSON API for live CoreGPU stock and prices: list every GPU or fetch one model, with availability, restock dates, a field reference and curl examples.
- [CoreGPU CLI (early access)](https://coregpu.io/docs/cli): The coregpu command-line tool: list live stock, deploy a server, renew and stop from your terminal. Currently in early access.

## Company and policies
- [About CoreGPU](https://coregpu.io/about): how prices are set, what is published, key facts.
- [Support](https://coregpu.io/contact): console tickets for customers with an active server; no email, phone or contact form.
- [Security](https://coregpu.io/security): single-tenant servers, SSH by key or by a root password hashed in the browser, secure erase between customers, 2FA.
- [Data centers](https://coregpu.io/data-centers): power, cooling and network requirements for each GPU platform.
- [Service Level Agreement](https://coregpu.io/legal/sla): 99.9% monthly availability, hardware replacement, service credits.
- [Terms of Service](https://coregpu.io/legal/terms): prepaid balance, renewals, cancellation and refunds.

## Machine-readable
- [llms-full.txt](https://coregpu.io/llms-full.txt): the pages above as plain Markdown in one file.
- [Inventory API](https://coregpu.io/api/v1/gpus): live stock and prices as JSON, no key, CORS enabled; one GPU at /api/v1/gpus/{slug}.
- [OpenAPI description](https://coregpu.io/api/v1/openapi.json): the inventory API for agents and tools.
- [Sitemap](https://coregpu.io/sitemap.xml) and [blog feed](https://coregpu.io/blog/feed.xml).

## Optional
- [Changelog](https://coregpu.io/changelog), [System status](https://coregpu.io/status), [Privacy Policy](https://coregpu.io/legal/privacy), [Acceptable Use Policy](https://coregpu.io/legal/acceptable-use), [Data Processing Addendum](https://coregpu.io/legal/dpa).