Compute profiles & GPU agents

A compute profile is the size of machine an agent’s container gets while it is running. You pick a profile per agent, so a writing agent can run on a small profile while a transcription agent runs on a GPU.

Available profiles

Profile	ID	vCPU	Memory	GPU	Price / hour	Billed per minute	Typical use
Standard	standard	2	4 GB	—	€0.009	€0.0001	Most agents: chat, research, writing, back office, CRM work.
Performance	performance	4	8 GB	—	€0.039	€0.0007	Coding agents, builds, browser automation, heavier tool use.
High memory	memory	4	32 GB	—	€0.14	€0.0023	Large datasets, analytics, knowledge-graph consolidation.
GPU · L4	gpu-l4	8	32 GB	1× NVIDIA L4 24 GB	€0.59	€0.0098	Local inference, embeddings, transcription, image work.
GPU · A100	gpu-a100	12	85 GB	1× NVIDIA A100 80 GB	€2.90	€0.0483	Fine-tuning, large open models, video pipelines.
GPU · H100	gpu-h100	16	120 GB	1× NVIDIA H100 80 GB	€4.90	€0.0817	The heaviest training and inference workloads.

Prices in EUR, excluding VAT. Compute is metered only while an agent's container is running a task.

Billed per minute, only while running

Agents are billed only while a container is running a task. Idle agents are suspended and their memory stays in git.

Which plans include which profiles

Profile	Launch	Scale	Enterprise
Standard standard	✓	✓	✓
Performance performance	✓	✓	✓
High memory memory	—	✓	✓
GPU · L4 gpu-l4	—	✓	✓
GPU · A100 gpu-a100	—	✓	✓
GPU · H100 gpu-h100	—	✓	✓

Scale and Enterprise can also define custom profiles with a specific vCPU, memory, GPU and disk mix.

Choosing a profile

| If the agent mainly… | Start with | |---|---| | chats, researches the web, writes, handles back office or CRM work | Standard | | writes and builds code, runs browser automation, uses heavy tools | Performance | | loads large datasets into memory, runs analytics, consolidates knowledge graphs | High memory | | runs local inference, embeddings, transcription or image work | GPU · L4 | | fine-tunes models, serves large open models, processes video | GPU · A100 | | runs the heaviest training or inference | GPU · H100 |

The model itself usually runs at your model provider, not in the container — so a typical agent that calls Claude, GPT or Gemini needs Standard, not a GPU. Use GPU profiles when the agent runs models or GPU-accelerated tools inside its container.

Signs that an agent needs a bigger profile:

• the agent header shows memory close to its limit, or runs fail with out-of-memory errors;
• builds or browser sessions time out;
• the same task runs much faster when tried on Performance.

Setting an agent’s profile

• Default profiles for new agents are chosen when you configure your instance and can be changed by asking support or from your account.
• Per agent, choose a profile when creating the agent or change it later. A profile change applies the next time the agent’s container starts; running tasks are not interrupted.

GPU agents

GPU profiles give the container a dedicated accelerator for as long as it runs:

• One GPU per container. The GPU is not shared with other agents or customers while your container holds it.
• Drivers and runtime are provided on GPU workers; your agent’s tools see the GPU through the standard CUDA stack.
• Capacity. GPU pools autoscale like CPU pools, but accelerators are a scarcer resource: when a pool is at capacity, a GPU run waits in the queue until a GPU is free. Enterprise customers can reserve GPU capacity so their GPU agents never wait.
• Cost control. Because billing is per running minute, keep GPU agents short-lived: let them do the GPU-heavy step and hand the rest to a CPU agent through delegation.

Resource limits inside the profile

An agent’s Settings → General also exposes memory and CPU limits and a per-task timeout. On ExoHuman Cloud the compute profile sets the ceiling; settings within the agent can only lower it. See Configuration.