§ 01 Inference · Overview

Low-Latency AI Inference
on Sovereign Infrastructure

Deploy open-weight AI models on EU-sovereign RISC-V infrastructure with sub-100ms latency. Our OpenAI-compatible inference API targets 612,000 tokens per second at peak throughput — with zero data egress and full GDPR compliance. No vendor lock-in. No US CLOUD Act exposure. Just pure, sovereign inference performance.

Peak tok/s

612K

Sustained tok/s

540K

Time to First Token

<50ms

Cheaper than B200

5×

§ 02 Modes · 3 deploy paths

Three Deployment Modes

Choose the sovereign inference deployment that fits your workload — from serverless scale-to-zero to air-gapped on-premises installations.

Serverless Inference

Pay per token with automatic scaling. Your managed inference endpoint scales to zero when idle — you only pay for what you use. Ideal for bursty workloads, prototyping, and startups who need real-time AI without upfront commitment. Each serverless inference request is processed on sovereign RISC-V silicon with sub-100ms latency and full EU data residency.

Pay-Per-Token

Dedicated Inference

Reserved inference throughput with guaranteed SLA. Your dedicated inference endpoint runs on isolated Tenstorrent Galaxy servers with predictable token-per-second performance. 99.9% uptime SLA, dedicated support, and consistent low-latency inference for production workloads that demand reliability.

Guaranteed SLA

On-Premises Inference

Air-gapped sovereign inference in your own facility. We will ship Tenstorrent Galaxy servers to your data center, install the AGIOS inference stack, and provide ongoing support. Design target: data stays inside your building's contracted residency boundary. For civil protection, banking, and classified workloads requiring the highest level of data sovereignty.

Air-Gapped

§ 03 Performance · Throughput

Inference Performance

Benchmark-leading inference throughput on RISC-V silicon. Every token processed on sovereign EU infrastructure.

Metric	AGICY (Tenstorrent Galaxy)	NVIDIA B200
Peak Throughput	612,000 tok/s (projected)	Varies by config
Sustained Throughput	540K+ (projected) tok/s	Varies by config
Time to First Token	<50ms	Varies
Cost per 1M Tokens	Contact for pricing	Varies by contract
Power per Server	8–10 kW operating (air-cooled) †	Higher (liquid-cooled)
Data Sovereignty	EU-only, designed for GDPR	US CLOUD Act exposure
Vendor Lock-In	None (open RISC-V ISA)	CUDA lock-in

§ 04 API · OpenAI-compat

OpenAI-Compatible API

Drop-in replacement for OpenAI, Azure, and Anthropic APIs. Switch your inference endpoint to AGICY with a single line change.

# Switch from OpenAI to AGICY — one line change
from openai import OpenAI

# Just change the base_url and api_key
client = OpenAI(
base_url="https://api.agicy.ai/v1",
api_key="your-agicy-key"
)

# Everything else stays the same
response = client.chat.completions.create(
model="llama-4-maverick",
messages=[{"role": "user", "content": "Explain sovereign AI"}],
stream=True,
)

Full support for streaming, function calling, tool use, and JSON mode. Compatible with LangChain, LlamaIndex, and every major AI framework. Our inference API supports all major open-weight models including Llama 4 Maverick, Mistral Large, DeepSeek V3, Qwen 3.5, and Gemma 3 — deployed on sovereign RISC-V silicon with sub-50ms time-to-first-token latency.

✓ Streaming

✓ Function Calling

✓ Tool Use

✓ JSON Mode

✓ Vision

✓ Embeddings

✓ Batch API

✓ Python SDK

§ 05 Questions · FAQ

Frequently Asked Questions

Common questions about sovereign inference on AGICY infrastructure.

What models are available for sovereign inference?

AGICY supports all major open-weight models including Llama 4 Maverick, Llama 4 Scout, Mistral Large 2, DeepSeek V3, Qwen 3.5, and Gemma 3. All models run on sovereign RISC-V infrastructure with no data egress to third parties. You can also deploy your own fine-tuned models through our Model Marketplace.

How does the OpenAI-compatible API work?

Our inference API is a drop-in replacement for the OpenAI API. Simply change your base_url to api.agicy.ai/v1 and use your AGICY API key. All existing OpenAI SDK calls, LangChain integrations, and LlamaIndex pipelines work without code changes. We support streaming, function calling, tool use, JSON mode, and embeddings — all processed on sovereign EU infrastructure.

What latency can I expect from sovereign inference?

Our low-latency inference delivers sub-50ms time-to-first-token (TTFT) for most models, with sustained inference throughput of 540K+ (projected) tokens per second across the fleet. For dedicated inference endpoints, we guarantee 99.9% uptime SLA with predictable token-per-second performance. Real-time AI applications see consistent sub-100ms end-to-end latency from our Vasilikos facility in Cyprus.

Is my data truly sovereign during inference?

Design target: every inference request is processed on RISC-V silicon in our planned EU-sovereign facility in Cyprus. Architecture aims so prompts and completions stay in EU jurisdiction. AGICY is a Cyprus (EU) entity with no US parent — CLOUD Act / FISA exposure is a design-avoidance goal, not a live certification determination. We don't log prompts, don't train on your data, and provide full audit trail transparency.

§ 06 Start · Deploy

Start Inferencing on Sovereign Infrastructure

Get your API key and deploy AI models on EU-sovereign RISC-V compute in minutes. OpenAI-compatible. No vendor lock-in.

Get API Access →View Pricing

🔧 Technology Stack ⚡ Sovereign Compute 🏪 Model Marketplace 🛡️ Trust Center