Serverless AI
Pay-Per-Token, Scale-to-Zero

AGICY's serverless platform achieves sub-100ms cold starts for the most popular models through our pre-warmed model pool architecture. Frequently requested models like Llama 4 Maverick, Mistral Large 2, and DeepSeek V3 are maintained in a ready state across the Tenstorrent Galaxy fleet, meaning the first request after an idle period experiences no perceptible cold start penalty. For less common or custom fine-tuned models, cold starts typically complete within 200–500ms as the model weights are loaded from NVMe storage into RISC-V memory. Subsequent requests to a warm endpoint experience the same sub-50ms time-to-first-token latency as dedicated inference endpoints.

You are billed for every token processed by the inference engine — both input tokens (your prompt and context) and output tokens (the model's response). Pricing varies by model size tier, with small models (7B–13B parameters) starting at €0.10 per million input tokens and Contact for pricing output tokens. There are no charges for idle time, network egress, or API calls that don't produce tokens. Billing is calculated per-request and aggregated into your monthly invoice with detailed breakdowns by model, endpoint, and time period. Real-time usage dashboards and budget alerts help you monitor and control spend. Volume discounts apply automatically above 100M tokens per month. There are no minimum commitments or reserved capacity fees.

Default rate limits are set at 1,000 requests per minute and 100,000 tokens per minute per endpoint, which is sufficient for most production workloads. These limits are soft limits and can be increased upon request — enterprise customers routinely operate at 10,000+ requests per minute with custom rate limit configurations. The platform's auto-scaling architecture supports bursts up to 50× the baseline request rate with no degradation in latency, as new compute capacity is provisioned within milliseconds from the pre-warmed pool. There are no hard upper bounds on total throughput; the platform scales horizontally across the entire Tenstorrent Galaxy fleet as needed. Contact our team to discuss custom scaling configurations.

Yes — AGICY's serverless platform provides the same data sovereignty guarantees as our dedicated and on-premises deployment modes. Every serverless inference request is processed on RISC-V silicon physically located in AGICY's EU-sovereign facility in Cyprus. Your prompts and completions never leave EU jurisdiction. The serverless control plane — including the request router, auto-scaler, and billing system — is also hosted entirely on EU-sovereign infrastructure with no dependencies on US cloud providers. There is zero exposure to the US CLOUD Act because AGICY is a Cyprus (EU) entity. We don't log your prompts, don't train on your data, and provide full audit trail transparency through our See our Trust Center for full audit trail documentation.