--- title: "Datacenter GPU Servers: A100, L40S & H100 · No KYC" description: "Offshore NVIDIA A100 80 GB, L40S and H100, up to 4 × H100 with NVLink, for training and large models. No KYC, paid in crypto." url: https://offshoreserv.com/offshore-gpu-servers/datacenter lang: en updated: 2026-09-27 source: HTML page at the url above (canonical); this is its Markdown version --- A100 · L40S · H100 SXM5 # A100, L40S & H100. Datacenter GPUs, offshore. HBM or ECC memory and NVLink for training, fine-tuning and serving large models: A100 80 GB, L40S 48 GB and H100 80 GB, up to four H100 in one server. **From $355.99/mo, no KYC.** - 48 to 320 GB of VRAM - Up to 3.35 TB/s per GPU - NVLink on multi-H100 nodes - Ready in 1 to 24 hours ## Five datacenter GPU servers. Every GPU is dedicated to you, with fast NVMe for datasets and checkpoints. | GPU | VRAM | CPU | RAM | Storage | Price | | --- | --- | --- | --- | --- | --- | | **A100 80 GB** (Training and fine-tuning) | 80 GB | 16 vCPU | 128 GB | 3.84 TB NVMe | **$355.99**/mo | | **L40S** (Inference at scale) | 48 GB | EPYC 7443P · 24c | 256 GB | 2 × 1.92 TB NVMe | **$418.99**/mo | | **H100 80 GB** (Frontier training, SXM5) | 80 GB | 24 vCPU | 192 GB | 2 TB NVMe | **$581.99**/mo | | **2 × H100** (160 GB HBM3, NVLink) | 160 GB | 48 vCPU | 384 GB | 4 TB NVMe | **$1,096.99**/mo | | **4 × H100** (320 GB HBM3, NVLink), cluster | 320 GB | 96 vCPU | 768 GB | 30 TB NVMe | **$2,348.99**/mo | Need eight GPUs or InfiniBand between servers? GPU and dedicated customers can ask for a custom build by ticket from the [client area](https://offshoreserv.com/account/support). ## A100 or H100? Both have 80 GB. The H100 has much more bandwidth and FP8 support; the A100 costs less per month. ### A100 80 GB (Best value) Ampere, HBM2e - **Memory bandwidth**: ~2.0 TB/s - **Precisions**: FP16, BF16, TF32, INT8 - **Fits at 4-bit**: Up to about 70B - **Best for**: Fine-tuning, steady inference **$355.99**/mo ### H100 80 GB Hopper SXM5, HBM3 - **Memory bandwidth**: 3.35 TB/s - **Precisions**: FP8 added, Transformer Engine - **Fits at 4-bit**: Up to about 70B - **Best for**: Training, high-throughput serving **$581.99**/mo The **L40S** (48 GB, FP8) sits in between: an inference card for 32B models at 8-bit and for image and video generation. ## Which server for which job? Our starting points for the most common training and serving jobs. | Job | Our pick | Why | | --- | --- | --- | | LoRA fine-tuning up to 14B | L40S or A100 | 48 to 80 GB holds the model, adapters and optimizer state. | | Full fine-tuning of a 7–8B model | 2 × H100 | Mixed-precision training needs roughly 16 bytes per parameter: 112 to 128 GB, more than one 80 GB card. | | Serving 70B at 4-bit | A100 or H100 | About 50 GB with an 8K context leaves room for batching on one 80 GB card. | | Serving 70B at 8-bit | 2 × H100 | About 85 GB with an 8K context, split over NVLink. | | 405B at 4-bit, or training at scale | 4 × H100 | 320 GB of HBM3 in a single node. | Memory figures for the weights are in our [VRAM table](https://offshoreserv.com/offshore-gpu-servers#llm); context windows and batch size add to them. ## Where datacenter GPUs pay off. Jobs where HBM bandwidth, ECC memory and NVLink are worth the higher monthly price. - **Fine-tuning** — LoRA, QLoRA and full fine-tunes on private data that never leaves the server. — A100 - **Serving large models** — vLLM endpoints for 70B-class models with many concurrent users. — [vLLM guide](https://offshoreserv.com/docs/gpu/images) - **Training from scratch** — Research runs on up to four NVLink-connected H100. — [Multi-GPU](https://offshoreserv.com/offshore-gpu-servers/multi-gpu) - **Batch generation** — Large image, video or embedding jobs that run for days. — L40S ## Where datacenter GPUs are available. Each server has its own list; the plans above show exactly where. - [Moldova Chișinău Best value — Outside the EU · no DSA — London **~33 ms** New York **~113 ms** Singapore **~129 ms**](https://offshoreserv.com/locations/moldova) - [Netherlands Amsterdam Network hub — EU member · DSA applies — London **~7 ms** New York **~87 ms** Singapore **~154 ms**](https://offshoreserv.com/locations/netherlands) - [Iceland Reykjavík Free-speech haven — EEA · outside the EU · no DSA — London **~29 ms** New York **~63 ms** Singapore **~169 ms**](https://offshoreserv.com/locations/iceland) Latency figures are estimates from distance, not measurements. Every jurisdiction has the same prices. [How we estimate latency](https://offshoreserv.com/network) ## The rules, before you pay. What we promise is written into our policies, not just our marketing. - US DMCA notices **Not actioned** — Answered with our policy, never enforced. Only a local court order, or a valid EU notice in EU locations, can require action. - Identity **Email only** — No name, address, phone or ID document, ever. A private or disposable address is fine. - Payment **5 cryptocurrencies** — Bitcoin, Ethereum, Monero, USDT and Solana, paid on-chain from any wallet to your balance. No card processor, no chargebacks. - Transparency **Signed canary** — A PGP-signed warrant canary every quarter and a public count of every request we receive. ## Datacenter GPUs, answered. **Another question?** The full FAQ answers what people ask before an order: payments, privacy, complaints and support. ### Is the H100 the SXM or the PCIe version? Our H100 servers use the SXM5 form factor, with HBM3 at 3.35 TB/s per GPU. The 2 × and 4 × H100 servers connect the GPUs with NVLink. ### Do I get the whole GPU? Yes. Every GPU is dedicated to your server. There is no time-slicing and no MIG partitioning shared with other customers. ### Which CUDA version is installed? The Ubuntu 24.04 image ships with a current NVIDIA driver and CUDA toolkit; run `nvidia-smi` to see the exact versions. The [GPU images guide](https://offshoreserv.com/docs/gpu/images) matches PyTorch builds to them. ### How fast is delivery? Between 1 and 24 hours after you order. If a server is temporarily out of stock in your jurisdiction, we tell you and either hold your order or return the payment to your balance. ### Can I get a refund? GPU servers are not refundable once delivered, because the hardware is set aside for you. Check the specifications and the [GPU images guide](https://offshoreserv.com/docs/gpu/images) before you order. ## Keep reading. Guides, policies and articles picked for this page, written by our team. - [H100 servers 80 GB of HBM3, up to 4 × H100.](https://offshoreserv.com/offshore-gpu-servers/h100) - [A100 80 GB servers 80 GB of HBM2e for fine-tuning.](https://offshoreserv.com/offshore-gpu-servers/a100) - [How much VRAM do LLMs need?Model sizes, quantization and the right GPU.](https://offshoreserv.com/blog/how-much-vram-for-llms) - [AI-ready GPU images Drivers, CUDA and the tools preinstalled on GPU servers.](https://offshoreserv.com/docs/gpu/images) - [Multi-GPU servers Up to 4 × H100 and 320 GB of VRAM.](https://offshoreserv.com/offshore-gpu-servers/multi-gpu) - [RTX GPU servers The best price per token for inference.](https://offshoreserv.com/offshore-gpu-servers/rtx) ## Datacenter GPUs, without a datacenter contract. 1. Pick a GPU and a jurisdiction 2. Pay in crypto, no ID asked 3. SSH in within 1 to 24 hours From $355.99/mo · dedicated GPUs · CUDA preinstalled --- OffshoreServ is an offshore hosting provider: VPS, dedicated servers, Windows RDP and GPU servers in seven jurisdictions (Iceland, Switzerland, Moldova, Romania, the Netherlands, Bulgaria and Malaysia), paid only in cryptocurrency (Bitcoin, Ethereum, Monero, Tether (USDT) and Solana), with no identity checks (no KYC). Prices and plans: https://offshoreserv.com/pricing · Answers: https://offshoreserv.com/faq · Every page: https://offshoreserv.com/llms.txt