Private, US-hosted GPUs behind an OpenAI-compatible endpoint. Point your existing tooling at one new base URL and run up to 30B-class models — at a flat monthly price, with no per-token surprise bills and zero data retention. We never train on your prompts.
# Your existing OpenAI code — change ONE line:
from openai import OpenAI
client = OpenAI(
base_url="https://api.nexusaisolutions.io/v1", # ← only this
api_key="nxs_your_key",
)
r = client.chat.completions.create(
model="qwen2.5-32b",
messages=[{"role":"user",
"content":"Runs on hardware you can audit."}],
)
# Same SDK. Same shapes. Your data never leaves the cluster.A MacBook runs an 8B model and thermal-throttles under load. Nexus gives you real capacity on shared private GPUs — for less than renting a cloud instance by the hour, and without your prompts becoming someone's training data.
Up to ~30B at 4-bit today — 3-4× the class your laptop can hold — with 70B on the roadmap. Llama, Qwen, Mistral, DeepSeek, or your own fine-tuned adapter.
Measured ~85 tokens/sec single-stream on a 7B, ~130 tok/s at 4-way concurrency, prefill above 1,200 tok/s. On a private network there's no public-internet round trip — latency is just model speed.
No prompts, no completions, no telemetry leave the cluster. We don't train on your data and we don't retain it. Your fine-tuned adapter stays on your tenant — never pooled, never published.
If your tool speaks OpenAI — Cursor, Continue, Open WebUI, LangChain, your own harness — it already speaks Nexus. Change the base URL, keep everything else.
Full model list & endpoints in the docs.
Drag your monthly usage and pick the model you’d otherwise pay for. We show your flat Nexus cost, your effective per-token rate here, the metered cost of the same volume, and exactly where the line crosses — honestly, both ways.
≈ light chat 1-2M · heavy coding assistant 10-30M · batch/agents 50M+
Published list prices, Sept 2026 — verify current. Edit to your real rate.
3,000 GPU-min included · ~23M tokens
Anchor: ~8,000 output tokens per GPU-minute (blended 7B, measured). 13B/30B consume more GPU-min per token — the meter self-adjusts. Overage past your bucket is $0.02/GPU-min.
What metered APIs can’t give you at any price:
Every plan starts with a free 7-day trial — no card. Upgrade only when it's paying for itself. No free-forever tier: the trial converts, or you're out — clean and honest.
Indie developers & solo builders
Small teams & one-person agencies
Growing product or agency
Your laptop caps at ~8B and throttles under sustained load. Nexus gives you 30B-class models and real concurrency on private GPUs — for less than renting a cloud instance by the hour, and without buying and cooling your own card.
No. No training, no retention, no telemetry. Prompts and completions stay on your tenant and are wiped at job end. A fine-tune you run stays a private adapter on your tenant.
Llama 3.x, Qwen 2.5 (incl. 32B), Mistral, DeepSeek, and your own fine-tuned variants — anything open-weight up to ~30B at 4-bit today, 70B on the roadmap.
Your plan includes a monthly GPU-minute bucket. Past it, you pay a flat $0.02/GPU-min or auto-upgrade — your choice. No opaque per-token metering, no throttling mid-project.
Sign up (no card), get an nxs_… key, point your base_url at our endpoint. Minutes, not a sales cycle.
7 days free, no card. Bring your OpenAI code — change one line.
Start free — no card requiredUS-hosted private GPUs · your data never leaves the cluster