Real servers.
Real prices.
No fluff.

Perfect first server. Host a few sites or a small Node/Python app.

• 1× IPv4 + IPv6
• Snapshot included
• Root access

Same shape as Starter with more RAM headroom for live traffic.

• 1× IPv4 + IPv6
• Daily snapshot
• Root access

Where real apps live. Postgres + Redis + Node API + Nginx — all on one box.

• Dedicated 1 Gbps port
• Auto backups
• DDoS protection

48 GB RAM swallows Postgres + Elasticsearch + Redis simultaneously.

• Dedicated 1 Gbps
• Auto backups
• Optional managed Postgres

16 cores. 64 GB RAM. The "I never want to think about CPU again" tier.

• Dedicated 1 Gbps
• Hardware RAID
• IPMI access

Top-tier shared compute before going bare metal.

• Dedicated 1 Gbps
• Hot snapshots
• Priority support

Cheap bulk SSD for backups + cold media.

• SFTP / SMB / NFS
• Snapshot rollback
• Auto-expand option

Step up if your backups crossed 300 GB.

• SFTP / SMB / NFS
• Snapshot rollback
• Versioned backups

A full terabyte. Where serious archives live.

• 1 Gbps port
• NFS / SMB / SFTP
• Restic-friendly

Storage + enough RAM for actual workloads (file processing).

• 1 Gbps port
• CPU headroom
• Daily snapshots

Storage + heavy compute on the same box. Process video, dedupe huge datasets.

• 1 Gbps port
• 50 GB RAM
• Hardware RAID

Your own physical machine. Nothing shared.

• Bare metal
• IPMI included
• Dual NVMe option

Server-grade EPYC + ECC RAM. The grown-up tier.

• ECC memory
• RAID 1 included
• IPMI / remote console

32 cores of pure 4 GHz boost. Compile + serve at the same time.

• NVMe RAID
• 4.2 GHz boost
• Hot-swap drives

Enterprise bare metal. 64 cores. 192 GB. Quiet operation, ridiculous throughput.

• Hardware firewall option
• Dedicated 10 Gbps upgrade
• Compliance-ready

Cheapest GPU-attached VPS we sell. Light CUDA, hardware video encode, learning environment.

• CUDA-capable
• Bare hypervisor allocation
• Snapshot included

Newest-gen consumer VPS slice. 6 GB GDDR7 for entry image generation + small Q4 LLMs at low cost.

• Blackwell-gen
• GDDR7
• Virtualized GPU slice

Entry GPU box. Image generation, light fine-tuning, video upscale.

• CUDA + cuDNN preinstalled
• Docker ready
• Jupyter optional

RTX 5060 8 GB on 24-core Skylake-SP Platinum. Bare-metal, full card, modern GDDR7.

• Blackwell + GDDR7
• 24-core Platinum host
• Bare-metal dedicated

Entry-level Pascal GPU. Light CUDA acceleration, classic video transcode, low-power lab work.

• Pascal architecture
• Low-power footprint
• Bare-metal dedicated

Quadro-class Pascal entry. 4 GB GDDR5 for light ML and visualization.

• Quadro drivers
• Bare-metal dedicated
• OpenGL + CUDA

Legacy Tesla K80 dual-GPU. Cheap aggregate VRAM for batch CUDA work.

• Dual K80 dies
• 24 GB combined VRAM
• Bare-metal dedicated

Turing-gen workstation GPU. Low-power 4 GB card with modern driver stack.

• Turing architecture
• Low TDP
• Quadro drivers

Consumer Turing card. Affordable CUDA for hobby ML and small models.

• Turing architecture
• CUDA + RTX optional
• Bare-metal dedicated

Turing 6 GB consumer card with dual-Xeon host. Solid entry image-gen + 7B Q4 inference.

• 6 GB GDDR5
• Dual-Xeon host
• Bare-metal dedicated

RTX 2060 with 128 GB system RAM. RT cores + Tensor cores for entry-level production.

• RT + Tensor cores
• 128 GB host RAM
• Bare-metal dedicated

Datacenter Pascal P100. 16 GB HBM2, FP64 powerhouse for scientific compute.

• HBM2 memory
• FP64 strong
• Datacenter pedigree

Ada Lovelace consumer. 8 GB GDDR6 + DLSS 3. Solid all-rounder for inference + render.

• Ada architecture
• DLSS 3
• Bare-metal dedicated

RTX 2060 on 40-core Skylake-SP. GPU paired with serious host compute.

• 40-core host
• RT + Tensor cores
• Bare-metal dedicated

Datacenter Volta V100 32 GB. HBM2, Tensor Cores Gen1. Still a workhorse for training.

• 32 GB HBM2
• Tensor Cores Gen1
• NVLink-ready

Ampere RTX 3060 Ti. 8 GB GDDR6 with strong tensor throughput per dollar.

• Ampere architecture
• Tensor Cores Gen3
• 2 TB storage

Ampere RTX A5000 pro card. 24 GB GDDR6 ECC. Workstation-class reliability.

• 24 GB GDDR6 ECC
• NVLink
• Quadro-class drivers

RTX A4000 on dedicated bare metal. 16 GB ECC, NVLink, full pro driver stack.

• 16 GB GDDR6 ECC
• NVLink
• Bare-metal dedicated

16 GB GDDR7 Blackwell-gen workstation card. ECC VRAM. Visualization + light ML.

• ECC VRAM
• Quadro-class drivers
• OpenGL + Vulkan

16 GB VRAM. Real ML territory — fine-tune image-gen models, run 7B LLMs.

• 16 GB GDDR6
• PyTorch / TF / JAX ready
• NVLink option

24 GB pro-grade Blackwell-gen + 56 GB system RAM. The sweet spot for production inference.

• Quadro-class driver
• 24/7 SLA option
• Tier-1 datacenter

48 GB GDDR7 Blackwell-gen. Run unquantized 30B LLMs, image-gen + short-form video warm together.

• 24 GB GDDR6
• NVLink-ready
• Multi-instance GPU

Latest-gen consumer flagship with workstation cooling. Brutal performance.

• 32 GB GDDR7
• 1 Gbps uplink
• Liquid-cooled chassis

Dedicated bare-metal RTX 4090 with 256 GB system RAM. No sharing.

• Dedicated GPU
• 256 GB system RAM
• Triple-tier storage

48 GB VRAM = chat + image both warm at the same time. The "honest dedicated" tier.

• 48 GB GDDR6 ECC
• NVLink
• 24/7 dedicated

96 GB GDDR7 Blackwell-gen flagship workstation card. ECC VRAM. Insane VRAM headroom.

• 96 GB GDDR7 ECC
• 1 Gbps uplink
• Workstation drivers

Bare-metal RTX 5090 32 GB with 256 GB RAM + triple-tier storage. The honest dedicated 5090 tier.

• Latest-gen flagship
• 256 GB host RAM
• Triple-tier storage

Enterprise-class A40 — datacenter pedigree, 48 GB ECC, passive cooled.

• Datacenter card
• ECC + RAS
• Up to 7 MIG instances

Datacenter A100 40 GB. The card every ML team learns on.

• Tensor Cores Gen3
• HBM2e memory
• 600 GB/s NVLink

80 GB HBM2e — fits Llama 70B unquantized, all media warm together.

• 80 GB HBM2e
• Full A100 architecture
• Datacenter SLA

The flagship. H100 80 GB Hopper. Best single-GPU performance available.

• Hopper architecture
• Transformer Engine
• 3 TB/s HBM3

Two 4090s side by side. Parallel inference or model + media on separate cards.

• 2× consumer flagship
• Dual PCIe Gen4
• CUDA peer-to-peer

Three V100 32 GB datacenter cards. 96 GB combined HBM2 for distributed training.

• 3× datacenter V100
• HBM2 across all
• NVLink option

Three A5000 pro cards. 72 GB combined ECC VRAM. Great mid-tier training cluster.

• 3× pro Ampere
• 72 GB ECC total
• NVLink-ready

Two newest-gen 5090s. 64 GB combined VRAM. Liquid-cooled.

• Latest-gen pair
• 44 cores host
• GPUDirect ready

Three A6000s. 144 GB total. Cheaper than 2× A100 80 GB with way more VRAM.

• 3× pro cards
• NVLink possible
• ECC across all

192 GB VRAM. 512 GB system RAM. Heavy enterprise workload tier.

• Quad-GPU
• 512 GB host RAM
• 20 TB storage

Training-cluster grade. Four A100 80GBs, 320 GB combined VRAM. Train your own foundation model.

• 4× datacenter A100
• NVLink Bridge
• Optional InfiniBand