RentServers.comDedicated LLM Inference on Reserved H200 and H100 Nodes
Your open models on two reserved HGX nodes, 8x H200 SXM and 8x H100 SXM, behind one OpenAI-compatible endpoint. Wells Software Consulting LLC runs the vLLM and Ray Serve stack, the monitoring and the model swaps. Nobody else's traffic touches the hardware, nothing is retained, and the SLA is written.
This is a managed service, not a referral. You contract with and pay Wells Software Consulting LLC, and the term sheet itemizes the hardware lease and our management fee.
Two pools, each matched to its models
Large models get the H200's 141 GB per GPU; small models get eight parallel replicas on the H100 node.
Large-model pool
8x H200 SXM 141GB
Llama 3.3 70B FP8 as 4 replicas at tensor parallel 2, Llama 3.1 405B FP8, DeepSeek R1 / V3 FP8 at TP8 or EP8.
Burst overage: $4.95 per active GPU-hour, billed per minute
High-throughput pool
8x H100 SXM 80GB
Llama 3.1 8B and other 7B to 32B models as 8 single-GPU replicas, embeddings and rerankers, or 70B FP8 as 2 replicas at TP4.
Burst overage: $4.25 per active GPU-hour, billed per minute
One monthly price for both pools
A reserved floor, not per-token billing: your capacity is there whether traffic spikes or idles, and your bill is the same every month.
| Item | Price | What it covers |
|---|---|---|
| Monthly floor | $34,000/month | Both pools, 16 GPUs reserved 24/7 (11,680 GPU-hours, about $2.91 per GPU-hour). No per-token charges. |
| H200 burst | $4.95/GPU-hour | On-demand H200 capacity beyond the reserved node, billed per active minute, only on your approval. |
| H100 burst | $4.25/GPU-hour | On-demand H100 capacity beyond the reserved node, billed per active minute, only on your approval. |
| Minimum term | 6 months | Then month to month with 60 days' notice. A 72-hour pilot comes first. |
The same 16 GPUs at serverless list prices
8 H100 and 8 H200 GPUs running 730 hours a month, against the $34,000 floor.
| Platform | H100 per hour | H200 per hour | 16 GPUs, one month | Dedicated floor vs list |
|---|---|---|---|---|
| Modal | $3.95 | $4.54 | $49,575 | 31% lower |
| RunPod Serverless (flex workers) | $4.79 | $5.93 | $62,605 | 46% lower |
Public list prices: Modal: modal.com/pricing, per GPU-second, billed while running, dated October 21, 2026; RunPod Serverless (flex workers): www.runpod.io/product/serverless, per GPU-hour, dated September 27, 2026. Hourly figures are per GPU; Modal's per-second rate is multiplied by 3,600. Serverless platforms bill only while workers run, so a workload active part of the day costs less than the month shown.
What we run for you
The serving stack, operations and engineering are in the monthly floor.
- One endpoint, two pools. Ray Serve handles API keys, quotas and routing in front of two independent vLLM pools. Each model is pinned to the node that suits it, and a fault in one pool leaves the other serving.
- OpenAI-compatible API. Chat completions, completions and embeddings with streaming. Point your existing SDK at a new base URL and key.
- Right-sized replicas. Tensor parallel stays inside each node on NVLink. 70B FP8 runs as four 2-GPU replicas on the H200 node; small models run one per GPU on the H100 node, so more requests run in parallel at lower latency.
- Tenant isolation. Separate deployments for sensitive workloads, no shared prefix cache across tenants, per-key rate limits.
- Model swaps. Change or add models on 7 days' notice, with a shadow-traffic check before cutover.
- 24/7 monitoring. Health checks, latency and error-rate alerting on both pools, with a monthly SLA report from the request metadata.
The SLA and data terms, in writing
These commitments go into the contract and are measured from the request metadata every month.
| Commitment | Target | Detail |
|---|---|---|
| Uptime | 99.5% per pool, per month | Service credits: 10% below 99.5%, 25% below 99% of the monthly floor |
| Time to first token | p95 under 1,000 ms | 70B class models, 2,000 input tokens or fewer, within contracted concurrency |
| Stream speed | 35 tok/s per stream (70B), 100 tok/s (8B) | Minimum per stream within contracted concurrency |
| Hardware replacement | 24 to 72 hours | Failed GPUs or nodes replaced by the data center operator under our lease |
| Data | Zero retention, no training | Metadata only (token counts, latency, key ID), kept 30 days; TLS 1.2+, encrypted disks, US only, deletion certificate at exit |
From request to production
Nothing is billed until you sign the term sheet.
- 1Send the workloadModels, prompt and output lengths, peak concurrency and latency target. The form below takes two minutes.
- 2Pilot on your promptsWe benchmark your real traffic shape on both pools with vllm bench and run a 72-hour pilot with monitoring live.
- 3Sign and launchYou get a term sheet with the floor, burst rates, SLA and data addendum. Once signed, the endpoint moves to production with a 6-month minimum.
Questions
Something else? Email hello@rentservers.com.
Who is the contract with?
Wells Software Consulting LLC. This is a managed service, not a referral: you sign one agreement with us, we lease and operate the two nodes, and we invoice you monthly. The term sheet names the US data center, the hardware and the supplier as the only subprocessor before you sign.
What does the $34,000 a month cover?
Both pools, all 16 GPUs, reserved for you around the clock: 11,680 GPU-hours a month, about $2.91 per GPU-hour. It includes the hardware lease, networking, the serving stack, monitoring, model swaps and engineering time. There is no per-token charge on top of the floor.
What is overage?
Burst capacity beyond the two reserved nodes, added on demand when you ask for it: $4.25 per active H100 GPU-hour and $4.95 per active H200 GPU-hour, billed per minute. Nothing is added without your written approval, and a monthly cap goes in the contract.
How is our margin disclosed?
The term sheet itemizes the monthly floor into the hardware lease we pay the data center operator and our management fee, and states the markup applied to burst capacity. You see both lines before you sign.
Which models can we run?
Open-weight models you choose: Llama, DeepSeek, Qwen, Mistral or your own fine-tuned weights. 70B class models run FP8 on the H200 pool; 8B to 32B models, embeddings and rerankers run on the H100 pool. You can swap models on 7 days' notice. You confirm your license terms, including the Llama Community License user threshold.
Is anything retained or used for training?
No. Prompts and outputs are not stored and never used for training. We keep request metadata only (token counts, latency, API key ID) for 30 days for billing and the SLA report. Traffic is TLS 1.2 or later, disks are encrypted, and the nodes stay in the US. At exit you receive a deletion certificate.
Why two separate pools instead of one 16-GPU cluster?
Tensor parallel traffic belongs on the NVLink inside one node. Splitting one model across an H100 node and an H200 node runs it at H100 speed, strands H200 memory and lets either node take it down. Two pools behind one front door give each model the right hardware and keep a fault in one pool away from the other.
What happens if a node fails?
The other pool keeps serving. The supplier replaces failed hardware within 24 to 72 hours under our lease, and downtime counts against the 99.5% uptime commitment, with credits paid automatically. If you need failover to a hosted endpoint under the same data terms, we set that up before launch.
What is the minimum term?
6 months. After that the service runs month to month and either side can end it with 60 days' notice. A 72-hour pilot on your real prompts runs before the term starts.
How is this different from Managed Inference on Google Cloud?
Managed Inference on Google Cloud runs clusters inside your own GCP organization: Google bills you for compute and we charge a flat retainer per node. Dedicated Managed Inference is turnkey: we provide the hardware and the endpoint for one monthly price. Managed Inference on Google Cloud
NVIDIA, H100, H200 and HGX are trademarks of NVIDIA Corporation. Llama is a trademark of Meta Platforms, Inc. Modal and RunPod are trademarks of their owners. Wells Software Consulting LLC is not affiliated with them.