Green inference. Built in Massachusetts.
Provocative is an OpenAI-compatible inference API for open-weight LLMs and embedding models, running on GPUs we own and operate. Our datacenter co-locates compute with on-site direct air capture — inference workloads run alongside hardware that pulls CO₂ out of the atmosphere.
Get started{ .md-button .md-button–primary }
See pricing{ .md-button }
Why Provocative
- :material-server-network:{ .lg .middle } **Owned hardware**
---
We run our own racks of H100, Blackwell, and RTX 5090 GPUs in our Massachusetts facility. Marginal cost is electricity and amortization, not an AWS markup — which translates to per-token prices 30–50% below cloud inference and contractual reserved capacity for teams that need it.
- :material-leaf:{ .lg .middle } **Green datacenter**
---
Compute is co-located with **direct air capture** at the same facility — the same racks that serve your tokens also pull CO₂ back out of the air, shrinking the footprint of every request without leaning on retroactive offset credits.
- :material-map-marker-radius:{ .lg .middle } **Northeast US POP**
---
Inference traffic terminates in Massachusetts, not us-east-1. Customers in the Northeast see lower round-trip latency than they would from any major-cloud region — meaningful for voice agents, IDE autocomplete, and real-time applications.
- :material-shield-check:{ .lg .middle } **Contractual data residency**
---
Because we own the racks, we can guarantee in writing where your prompts and completions are processed. Useful for fintech, healthtech, legal, and EU teams with regulatory exposure.
Drop-in OpenAI SDK
Change two lines and keep your existing OpenAI client:
from openai import OpenAI
client = OpenAI(
base_url="https://inference.provocative.earth/v1",
api_key="pk-prov-YOUR-KEY",
)
response = client.chat.completions.create(
model="llama-3.1-70b-instruct",
messages=[{"role": "user", "content": "Hello!"}],
)
The full migration guide is here.
Models
We serve a curated catalog of open-weight chat and embedding models, including Llama 3.1 (8B / 70B / 405B), Qwen 2.5 Coder 32B and 72B, Mistral Small 3, DeepSeek V3, and the embedding models BGE-M3, E5-Mistral, and Nomic Embed. The full catalog lists context windows, GPU class, and per-token rates.
Next steps