RentServers.comManaged vLLM Endpoints and GPU Clusters on Google Cloud (Your Project, Our SLA)
We engineer, secure, and operate high-throughput GKE, Ray, and vLLM clusters directly inside your GCP organization. Your team retains data sovereignty, VPC perimeters, and cloud billing credits while our desk handles CUDA drivers, NCCL tuning, and 24/7 SLA monitoring.
Private LLM hosting in your own GCP project
Managed inference without handing your prompts, weights or keys to another platform. Google bills the compute to you; we bill one retainer for the engineering.
Your GCP organization
Clusters run in a project under your organization and billing account. Google invoices compute to you at list price or at your negotiated rate, and your committed use discounts, reservations and credits apply.
Our engineering desk
Wells Software Consulting LLC operates the clusters through IAM roles your administrators grant and can revoke. We build and run GKE, Ray and vLLM; you keep the data, the VPC perimeter and the keys.
One flat retainer
We bill a fixed management retainer per node per month: $250 for L4, $800 for A100 80GB and $2,500 for H100. No percentage of spend and no markup on GPU hours.
Three managed packages
Every package runs at GCP list cost on your bill, plus a flat retainer per node from Wells Software Consulting LLC.
Managed Inference (L4 Fleet)
g2-standard-12 (1x L4) through g2-standard-96 (8x L4)
Optimized for Llama 3.1 8B FP8 and Mixtral 8x7B AWQ behind a vLLM or TGI endpoint with an OpenAI-compatible API.
GCP Compute (At List Cost) + $250/node/mo Management Retainer
- GCP list on-demand: $730 to $5,842 per node per month
- 3-year CUD from $329 per node per month
Heavy Inference & Fine-Tuning (A100 80GB)
a2-ultragpu-1g through a2-ultragpu-8g (1 to 8x A100 80GB)
Optimized for Llama 70B AWQ / BF16, on GKE with NVIDIA device plugins and Ray autoscaling for batch inference and fine-tuning jobs.
GCP Compute (At List Cost) + $800/node/mo Management Retainer
- GCP list on-demand: $3,700 to $29,602 per node per month
- 3-year CUD from $2,555 per node per month (1-GPU shape)
Foundation Training & Large Clusters (H100 80GB)
a3-highgpu-8g (8x H100 80GB)
Powered by 3-Year CUDs or GCP Dynamic Workload Scheduler (DWS) calendar blocks. Slurm or Kubernetes with NCCL tuning and GPUDirect networking.
GCP Compute (At List Cost) + $2,500/node/mo Management Retainer
- 3-year CUD: $28,371 per node per month at GCP list
- DWS calendar blocks: $41.60 per node-hour, no multi-year term
Size a managed vLLM deployment on GKE
Pick a model, your peak requests per second, a token split and a commitment tier. You see every node shape for that model, priced as GCP compute billed by Google plus our retainer.
Per-node monthly cost for Llama 3.1 8B FP8 at 10 peak RPS, 512:256 tokens, On-Demand
Node count for your peak RPS comes from a load test of your model and prompts on these shapes. Send the sizing with the request form and the plan comes back with measured throughput per node.
- A100 has no native FP8, so this option serves the model in BF16.
- DWS calendar blocks for a3-highgpu-8g list at $41.60 per node-hour with no multi-year term.
| Node shape | Serving | GPUs | GCP compute | Retainer | Total per node |
|---|---|---|---|---|---|
| g2-standard-12 | FP8 | 1 × NVIDIA L4 | $730 | $250 | $980 |
| g2-standard-24 | FP8 | 2 × NVIDIA L4 | $1,461 | $250 | $1,711 |
| g2-standard-48 | FP8 | 4 × NVIDIA L4 | $2,921 | $250 | $3,171 |
| g2-standard-96 | FP8 | 8 × NVIDIA L4 | $5,842 | $250 | $6,092 |
| a2-ultragpu-1g | BF16 | 1 × NVIDIA A100 80GB | $3,700 | $800 | $4,500 |
| a3-highgpu-8g | FP8 | 8 × NVIDIA H100 80GB | $64,598 | $2,500 | $67,098 |
How the calculator works
- Throughput per node: vllm bench serve, 512 input and 256 output tokens, highest request rate that holds p95 time to first token at or under 1 second, per node.
- Other token splits: per-node RPS × (512 × 0.29 + 256) ÷ (input × 0.29 + output). A prompt token costs about 0.29 of an output token.
- Node count: ceil(peak RPS × 1.3 ÷ per-node RPS), so peak load runs near 75% of capacity, plus 1 spare node (N+1).
- GCP compute: nodes × the GCP list price per node-hour for the tier × 730 hours, billed by Google to your account. Negotiated discounts lower it.
- Retainer: nodes × $250 (L4), $800 (A100 80GB) or $2,500 (H100) per month, billed by Wells Software Consulting LLC.
- Before launch we load-test your model and prompts on the chosen shape and resize if measured throughput differs.
GCP prices: us-central1 list per node-hour from Google Cloud accelerator-optimized VM pricing and Google Cloud Dynamic Workload Scheduler pricing, checked October 21, 2026, 730 hours per month. Other regions price differently. A2 Ultra commitment prices above one GPU scale the published a2-ultragpu-1g rate (committed use discounts); Google quotes those through its sales team.
GCP GPU cluster management, end to end
What the per-node retainer covers on every package.
- GKE GPU clusters. Private GKE clusters with GPU node pools, Google-managed NVIDIA drivers, autoscaling and node auto-repair.
- Managed vLLM and TGI endpoints. OpenAI-compatible endpoints, with weights and quantization (FP8, AWQ or BF16) chosen for your latency target, and rolling upgrades.
- Ray on GKE. KubeRay clusters for batch inference and fine-tuning on A100 and H100 node pools.
- Training clusters. a3-highgpu-8g nodes with GPUDirect networking, NCCL tuning, and Slurm or Kubernetes scheduling.
- Network and security. Private endpoints, VPC firewall policy, VPC Service Controls perimeters and customer-managed encryption keys where your policy calls for them.
- Quota and capacity. We file, justify, and track GCP quota increases and DWS reservations directly inside your billing account. No allocation promises, just expert procurement navigation.
- 24/7 monitoring. Endpoints and GPU nodes are watched around the clock with automated health checks and alerting. Uptime and response targets are written into your retainer agreement before launch.
Questions
Something else? Email hello@rentservers.com.
Who pays Google for the GPU compute?
Your organization does. The GPU instances, disks and network run in a project under your GCP billing account, so Google invoices you directly at list price or at any discount you have negotiated. Wells Software Consulting LLC invoices only the management retainer.
Do you mark up GPU hours?
No. The retainer is a flat fee per node per month: $250 for L4 nodes, $800 for A100 80GB nodes and $2,500 for H100 nodes. Compute appears on your Google invoice, not ours, and it is not a percentage of your spend.
Can we use our committed use discounts and cloud credits?
Yes. Commitments, reservations and credits live in your billing account, so they apply to these clusters automatically. Commitments are between you and Google; we size them with you before you sign.
What access do you need to our organization?
IAM roles on the projects we operate, granted by your administrators and revocable at any time. We do not need billing administrator rights, and access is scoped to those projects.
How do GCP quota and GPU capacity work?
Google grants quota to your project and decides on capacity, which can run short in a zone even after quota is approved. We file, justify, and track GCP quota increases and DWS reservations directly inside your billing account. No allocation promises, just expert procurement navigation. For capacity that has to be there on a date, we set up reservations or DWS calendar blocks.
How do you size the node count?
Per-node throughput comes from vllm bench serve runs of each model on each node shape, at a p95 time to first token of 1 second or less. We size for your peak requests per second with 30% headroom, add one spare node (N+1), and load-test your own prompts before launch.
What happens if we end the retainer?
You revoke our access and everything we built stays in your project, under your control. There are no commitments in our name to unwind. Notice terms are set in the retainer agreement.
How is this different from Managed AI Hosting?
Managed AI Hosting runs small business chatbots and document Q&A tools on infrastructure we manage. Managed Inference on Google Cloud runs your models on GPU clusters inside your own GCP organization. Managed AI Hosting is priced separately.
Request a deployment plan
Send the sizing from the calculator with your region and launch window. Nothing is billed until you sign a retainer agreement.
- 1We confirm the briefWithin one business day we confirm the model, region, launch window and the access you will grant.
- 2We load-test and size itWe run your model and prompts on the candidate node shapes and send a plan with measured throughput, node count, GCP list cost and the retainer.
- 3We build and operate itYou approve the plan and grant IAM roles on the project. We build the cluster, test the endpoint and keep it running under the retainer.
Google Cloud, GKE and Compute Engine are trademarks of Google LLC. Llama is a trademark of Meta Platforms, Inc. Mixtral is a trademark of Mistral AI. Wells Software Consulting LLC is not affiliated with these companies.