Dedicated LLM Serving

Dedicated LLM Serving reserves GPU hardware exclusively for your contract. You purchase the complete throughput of the hardware and use it however you want — within whichever model fits on the GPUs you reserve.

Unlike the shared LLM Serving Service, there are no rate-limit ceilings and no shared contention. End-to-end latency, time-to-first-token, and tokens-per-second become predictable functions of your model and prompt, not of other tenants’ traffic.

Compare to Shared Side-by-side comparison of the two service models and when to pick which.

Request a Quote Contact the AIFS team for a tailored Dedicated LLM Serving offering.

Why Choose Dedicated LLM Serving

1. No usage limits

The shared service publishes per-plan RPM (requests per minute) and TPM (tokens per minute) ceilings. On Dedicated LLM Serving, those ceilings do not apply — you can consume tokens up to whatever the reserved hardware can physically produce.

Performance may degrade if you push the hardware near its limit, but that ceiling is a property of your hardware, not a contractual quota. Reserving more GPUs lifts it.

2. Predictable performance, no contention

On the shared service, you compete for capacity with every other tenant calling the same open-source model. At peak load, end-to-end latency and time-to-first-token vary; effective throughput can be lower than the published TPM.

On Dedicated LLM Serving, your GPUs serve only your requests. Performance is deterministic and reproducible — the same prompt produces the same latency profile run after run. This is what makes it possible to meet tight SLAs on latency or throughput.

3. Any compatible model

Any model currently offered on the shared T-Cloud catalog can be deployed on a dedicated instance, provided it fits within the reserved hardware. You can also bring private or fine-tuned variants that aren’t in the public catalog — for example, a LoRA-tuned Llama or a domain-specific Mistral.

4. Predictable cost structure

You pay a fixed fee for the reserved hardware and a fixed fee for the managed inference service, instead of per-token charges that move with traffic. The cost structure doesn’t change month-to-month based on usage.

5. Custom SLAs

The shared service’s standard SLA (99.9% API availability during business hours, no committed resolution times) covers reachability only — it does not cover throughput, latency, time-to-first-token, or RPM/TPM. On Dedicated LLM Serving, you can negotiate custom SLAs that go beyond the standard product, including 24/7 coverage, faster response times, stricter availability targets, and — uniquely — performance SLAs tied to the reserved hardware (latency, time-to-first-token, sustained throughput). Discuss your needs with your Telekom account representative.

Commercial Structure

A Dedicated LLM Serving offering is composed of three commercial elements:

GPU infrastructure — a fixed monthly fee per node or per GPU equivalent, depending on the GPU type you choose (for example NVIDIA DGX B200) and the contract term.
Managed inference service — a fixed monthly fee on top of the infrastructure, covering operation of the inference stack, monitoring, model management, and support.
Custom SLAs — negotiated separately as part of the contract, covering availability, response times, and (uniquely to Dedicated LLM Serving) performance characteristics such as latency, time-to-first-token, and sustained throughput on the reserved hardware.

Contract terms, volume discounts, and SLA scope are negotiated as part of the offer. There are no published list prices for Dedicated LLM Serving — every offer is tailored to the chosen GPU type, node count, contract term, and SLA scope.

When to Move to Dedicated LLM Serving

The shared LLM Serving Service is appropriate for most production workloads. Move to Dedicated LLM Serving when one or more of the following apply:

Your workload is bursty. Short windows of very high token demand that would otherwise exceed shared RPM/TPM ceilings.
You need contractual latency or throughput SLAs. Time-to-first-token, end-to-end latency, or sustained-throughput commitments that the shared service does not — and structurally cannot — provide. On shared, the published RPM/TPM are best-effort ceilings, not SLA-backed performance figures.
You serve many users with strict latency SLAs. The Agentic plan publishes very high RPM and TPM ceilings, but they remain best-effort on shared infrastructure. There is no guarantee that bursty traffic from a large user base will land within the time-to-first-token or end-to-end-latency targets you owe your own users. Dedicated LLM Serving reserves hardware exclusively for your traffic, eliminating the shared-contention risk.
You want to host a private or fine-tuned model. Models that are not in the shared catalog — including LoRA or DPO-tuned variants.
Your compliance posture requires single-tenant infrastructure. No shared compute with other customers.
You want a predictable cost structure. A fixed fee for the hardware and the managed service, instead of token-based billing.

If none of these apply, stay on the shared service — it’s simpler, cheaper at low to moderate volume, and self-service via the marketplace.

Hybrid Deployments

Many enterprise customers combine both:

Dedicated LLM Serving for the latency-sensitive, high-throughput hot path (for example, a customer-facing chatbot or a production agent).
Shared LLM Serving Service for everything else — internal tooling, batch jobs, evaluation, exploration.

Both expose the same OpenAI-compatible API and the same SDKs. Switching a request between Dedicated and Shared is a configuration change, not a code change.

How to Order

Dedicated LLM Serving is not available via the T-Cloud Marketplace — it requires a tailored offer. To get a quote:

Email the AIFS team at ai@t-systems.com with your expected workload profile (peak request rate, sustained token volume, model preferences, latency targets, contract term).
The team will produce a tailored price indication based on your inputs.
You receive a formal offer including the chosen GPU configuration, the infrastructure and inference-service fees, contract term, and the negotiated SLA scope.

For very large or compliance-driven deployments, the AIFS team can also assess installation on a customer-owned environment (§ 3.4.3 of the Service Description), subject to feasibility checks. Mention this in your initial enquiry if relevant.

Shared vs Dedicated LLM Serving — Compare both service models side-by-side
Plans & Pricing — Pricing for the shared Essential / Professional / Agentic plans
Rate Limits — How RPM/TPM work on the shared service
Service Levels — Standard SLA terms (and how Dedicated LLM Serving can extend them)
Service Description (PDF) — Binding service description document