Furiosa NPU-Powered AI Inference API

RanoAI
AI Inference, Powered by NPU

Fast, efficient AI inference on Furiosa NPU hardware.6 models · up to 120B params · 60% cheaper than GPU cloud.

3x Power Efficiency vs GPU
60% Lower Token Cost
Gemma 4 Multimodal
Furiosa NPU Chips
example.py
# OpenAI-compatible — drop-in replacement
from openai import OpenAI

client = OpenAI(
  base_url="https://api.ranoai.com/v1",
  api_key="rno-xxxxxxxxxxxxxxxx"
)

# gemma-4-31b · gemma-4-26b · gemma-4-12b
# qwen-3.5-35b · qwen-3.5-27b
response = client.chat.completions.create(
  model="gemma-4-31b",
  messages=[{"role":"user",
    "content":"Hello!"}]
)

# $0.25 / 1M input tokens
NPU cluster online — avg latency ~280ms
AI Data Center
NPU cluster — all systems operational
Infrastructure

Next-Gen NPU
Built for Inference

NPU chips paired with Google and Alibaba's best open models — fast, affordable inference that GPU clouds can't match on cost.

Furiosa RNGDNPU · Furiosa AI · 48GB HBM3

Korea's second-generation NPU built for large-scale AI inference. RNGD delivers exceptional throughput per watt with a dataflow architecture purpose-built for transformer workloads.

Model Library6 Models · 3 Families

GPT-OSS-120B, Gemma 4 (31B · 26B · 12B), and Qwen 3.5 (35B · 27B) — six state-of-the-art models with multimodal and long-context capabilities, all served on NPUs.

Benchmarks

RanoAI NPU vs RTX 5090 & RTX PRO 6000

Gemma 4 27B inference · Furiosa RNGD (HBM3) vs NVIDIA Blackwell (GDDR7)

RanoAI NPU
GPU
Worst
Throughput
TPS — Gemma 4 27B, batch 32
1,920
1,050
568 †
RanoAI NPU
RTX PRO 6000
RTX 5090 †
Efficiency
TPS / Watt
6.4
1.75
0.99 †
RanoAI NPU
RTX PRO 6000
RTX 5090 †
Mem. Bandwidth
TB/s
~3.9
1.79
1.79
RanoAI NPU
RTX 5090
RTX PRO 6000
Power Draw
TDP per card (W) ↓
600W
575W
~300W
RTX PRO 6000
RTX 5090
RanoAI NPU
Token Cost
$ / 1M output ↓
$2.80 †
$1.40
$0.50
RTX 5090 †
RTX PRO 6000
RanoAI NPU

↓ lower is better for Power Draw and Token Cost  ·  † int4 quantization required (fp16 exceeds 32 GB VRAM)  ·  Source: Cloudrift, NVIDIA specs

How It Works

From API Key to Response
in 4 Simple Steps

Drop-in replacement for OpenAI. Point to our endpoint and let NPUs handle the rest.

💻
Your App
any OpenAI SDK
☁️
RanoAI API Gateway
api.ranoai.com
NPU Cluster
Furiosa RNGD
🤖
6 Models
GPT-OSS-120B · Gemma 4 · Qwen 3.5
Simple, transparent pricing

Pay Only for
What You Use.

No subscriptions. No commitments. NPU efficiency means lower cost per token — no GPU markups.

First 1M tokens free — no credit card required
ModelParamsContextInput / 1M tokensOutput / 1M tokens
GPT-OSSgpt-oss-120b
120B128K$0.60$1.80
Gemma 4gemma-4-31b
31B128K$0.25$0.75
gemma-4-26b
26B128K$0.20$0.60
gemma-4-12b
12B128K$0.08$0.25
Qwen 3.5qwen-3.5-35b
35B128K$0.28$0.85
qwen-3.5-27b
27B128K$0.18$0.55

Need higher rate limits or dedicated capacity?