gpubox

A slice of a GPU with a machine around it.

A persistent Linux microVM with SSH and a public IPv6, a share of a real NVIDIA or AMD card, and the model already resident on it. Billed by the second, nothing while stopped. A private 9B endpoint from $66 a month, Qwen3.8 27B from $97, Qwen3.8-Flash-Next from $167.

Private beta. Leave your email, invites go out in order. One email when your box is ready, then nothing.

what it looks like

one 24 GB card, live

█ model weights, once per card▒ a box's private VRAM· free

In every box

GPU shareA weighted share of a real card: CUDA on NVIDIA, Vulkan on AMD. Alone you get all of it.

llama-serverOpenAI, Anthropic and Responses APIs on port 8080. Point Claude Code, Codex, OpenCode or your IDE at it.

model residentCatalog weights are loaded once per card and shared read-only. A box is up in about 15 seconds.

ssh, scpA real Ubuntu machine. Bring your own runtime if it speaks CUDA or Vulkan.

IPv6 per boxRouted, public, free. IPv4 on request. HTTPS at box.gpubox.io.

stop, fork, resumeDisk snapshots. Fork a warm box for every agent task.

LoRA adaptersShared base, private adapter. One finetune per customer, cheap.

per-second billingGPU share, private VRAM, VM size. Stopped is $0.

Catalog at launch

model	weights	from / month
Qwen3.5 4B	2.7 GB	$39
Qwen3.5 9B	5.7 GB	$66
Gemma 4 12B	7.1 GB	$70
gpt-oss-20b	11.6 GB	$70
Devstral Small 2 24B	14.3 GB	$75
Qwen3-Coder-30B-A3B	18.6 GB	$75
Qwen3-Coder-Next 80B-A3B	48.4 GB	$167
GLM-4.7-Flash 30B-A3B	18.3 GB	$75
Qwen3.8 27B	16.5 GB	$97
gpt-oss-120b	62.8 GB	$167
Nemotron 3.5 Super 120B-A12B	87 GB	$167
Qwen3.8-Flash-Next 125B-A6B	75 GB	$167
DeepSeek V4 Flash 284B-A13B	160 GB	$254
whisper large-v3-turbo	0.9 GB	$35
bge-m3	0.6 GB	$33

Q4 GGUF. Or pull your own from Hugging Face. Prices with a nano VM at --gpu 1 to 4 by size. Context is private VRAM: at q8_0 a 9B box holds 128k tokens in 2.2 GiB, Coder-Next a million in 13 GiB. Flash-Next keeps its 51B n-gram table in the box's RAM. DeepSeek V4 Flash spans two cards, shared the same way, so a box takes shares on both.

Three meters

$0.03

per GPU share-hour, 24 GB class

$0.012

per GiB-hour of private VRAM

$0.012

per VM-hour, 1 vCPU 2 GiB

Catalog weights are not metered. A box alone on a card gets the whole card. Under contention, time divides by your share, and your share is the floor: we sell at most 24 of a card's 32 shares.

An always-on private 9B

gpubox, --gpu 2	$66
Google G4, 1/8 card, EU	$449
Koyeb, L4	$511
Hugging Face endpoint, L4	$584
RunPod, RTX 4090 secure cloud	$650

On-demand list prices, October 2026, 730 hours.