Scale-to-zero acceleration for inference and training.
The dedicated, high-performance compute plane for heavy machine-learning workloads, from inference and multi-modal synthesis to training and fine-tuning models, plus intensive numerical execution. Hardware starts on invocation and stops when the work is done, so agents get sub-second acceleration without managing any persistent infrastructure.
Scale to zero
Hardware provisions instantly on invocation and terminates on completion, so idle compute costs nothing.
Hot-mapped model cache
Popular open-weights models such as DeepSeek, Llama, Qwen, Flux and Whisper are pre-staged on NVMe and load into isolated VRAM in under a second.
PyTorch and CUDA enclaves
Run agent-written Python with direct access to torch, triton, flash-attn and torchvision, each run in its own isolated enclave.
Built for training too
Train and fine-tune models, including LoRA, as a job that starts on demand, saves checkpoints and weights to Ahura Store, and ends when training does.
Billed per GPU-second
Metered against your account and charged to your card or drawn from prepaid ACUs, within the spending limits you set.
{ "tool": "gpu.infer", "model": "whisper", "input": "store://recordings/call-0142.wav", "output": "store://transcripts/call-0142.json" }Interactive simulation
Press Run to send the call. The dashboard shows what it would use in a partition. Simulated: nothing is actually executed.
# See what is staged and hot.
→ POST /mcp
{
"jsonrpc": "2.0",
"id": 1,
"method": "tools/call",
"params": {
"name": "gpu.models",
"arguments": {}
}
}# Transcribe a recording. Hardware starts, runs and stops.
→ POST /mcp
{
"jsonrpc": "2.0",
"id": 2,
"method": "tools/call",
"params": {
"name": "gpu.infer",
"arguments": {
"model": "whisper",
"input": "store://recordings/call-0142.wav",
"output": "store://transcripts/call-0142.json"
}
}
}Ready. Press Run.
Partition · N. Virginia · created just now
Services in this partition
GPU
0 calls · 0.00 ACU
Store
0 calls · 0.00 ACU
Spend by service
Ledger
Why it matters
Heavy models need accelerators, and accelerators are usually rented by the hour whether anything is running or not. Agents work in bursts, so most of that spend is idle.
Cost
No idle accelerators, ever, whether inferring or training.
Speed
Sub-second weight loading by zero-copy memory mapping.
Isolation
Each job runs in its own enclave.
Training
Fine-tune and train models as an ordinary tool call.
Autonomy
Agents fund and settle their own compute.
Ahura GPU is for the heavy work. Use Ahura AI for cheap everyday inference turns and Ahura VM for ordinary code.
In depth
Run your own Python in a custom enclave, or call a lightweight endpoint for instant inference on a hosted model.
Describe what a job needs with a Python decorator. Ahura GPU starts an isolated sandbox, injects the code, runs it and returns the result.
Nothing stays running between calls. Idle compute costs nothing and there are no drivers or servers to look after.
Frequently used open-weights models are staged on NVMe and memory-mapped straight into VRAM with zero copies, so weights load in under a second.
DeepSeek, Llama, Qwen, Flux and Whisper are kept hot. Vision models such as Florence, Segment Anything and YOLO run the same way.
Agent-authored scripts reach torch, triton, flash-attn and torchvision directly, inside an isolated enclave.
Run training jobs, from full runs to low-rank adaptation (LoRA) sessions, in the same enclaves. Datasets are read from Ahura Store, and checkpoints and finished weights are written back to it.
Save progress as a job goes, so a long training run can be resumed by a workflow rather than started over.
Weights you train can be loaded for inference on the same platform, so a swarm can adapt and then use the result.
Billed by the GPU-second. Each run is metered against your account balance or card, within the spending limits you set.
Pair it with Ahura VM to prepare data, Ahura Store to hold inputs and outputs, and Ahura Workflows to fan work out in parallel.
4 tools your agent can call. Each one is a standard MCP tool with a typed schema and a price shown in tools/list.
Tool names on this page are indicative. Use gpu.models to see what is available at any time.
Due diligence assistant
Work through a data room and flag what needs a human.
Autonomous contract and document auditing
Audit hundreds of contracts, filings or codebases in parallel.
Real-time vision and video monitoring
Detect, track and read what appears in live video.
Synthetic media and asset generation
Design, render and localise product graphics, audio and assets.
Zero-trust processing of sensitive data
Let an agent work on files it never sees.
Entity resolution
Work out which records are the same thing.
and 1 more