Once an AI pilot becomes a product, a dedicated GPU server gives you a cost you can plan and control over where your data is processed. A dedicated NVIDIA RTX PRO 6000 server brings that option within reach of mid-sized workloads. That was the main takeaway from our webinar The Enterprise AI Blueprint: Scaling the NVIDIA RTX PRO 6000 Blackwell Server Edition, with NVIDIA and Stelia AI, a GPU service provider and Leaseweb customer that runs AI workloads for enterprises and smaller businesses.
TL;DR
Move an AI workload to dedicated infrastructure once it runs daily, handles sensitive data or costs more per month on an API than a server would. For mid-sized open models, the NVIDIA RTX PRO 6000 puts 96 GB of memory on one card, so the model runs on a single GPU, and MIG lets several teams share it. For models of roughly 70B to 405B parameters, choose H200-class servers with NVLink.
Prefer to watch? Here’s the full recording.
Two currencies: cost and control
Token prices keep falling, but the cost of running AI agents is rising. Gartner predicts that inference costs per agentic workflow will increase more than fivefold through 2028, and calls this the inference paradox [1]. Agents reason, call tools, and replan, so each task consumes far more tokens than a chatbot answer, and per-token API pricing grows with every successful feature you ship. Dedicated GPUs turn that into a fixed cost, and the business case depends on utilization: compare options against your real usage pattern, including quiet periods and peaks.
Control is the second currency. Teams that adopt AI tools ad hoc put intellectual property and sensitive data at risk. IBM’s 2026 Cost of a Data Breach Report found that shadow AI, meaning AI tools staff use without approval, was involved in 43% of security incidents at the organizations it studied, up from 20% a year earlier [2]. Data location is the first layer of control. Legal jurisdiction, who operates the infrastructure, what you can audit, and how the hardware supply chain is secured decide whether you keep acting on your own terms when suppliers, prices, or rules change. That is what we mean by digital independence. Leaseweb has helped customers keep that control since 1997.
Our recommendation for AI: start on APIs, then move a workload to dedicated infrastructure once it runs daily, handles sensitive data, or costs more per month on an API than a server would.
Why an RTX PRO 6000 server matters: one model, one card
The RTX PRO 6000 Blackwell Server Edition puts 96 GB of GPU memory on a single card, twice the 48 GB of the L40S. A mid-sized open model that used to be split across two cards now runs on one, so you pay for, monitor and maintain one GPU per model. Until now, the next step up was an H100- or H200-class server, built and priced for much larger models. Stelia AI’s CTO described the effect in the webinar: the lower entry cost lets them offer AI-powered products to their customers more affordably, and shortens the payback on the investment.
The rest of the card’s design builds on that:
- Dev, staging, and production on one server. Multi-Instance GPU (MIG) splits the card into up to four isolated 24 GB instances, where your deployment environment supports it. Give each environment or team its own instance on one dedicated server, and you raise utilization on hardware you already pay for.
- AI and visual work on the same hardware. The card runs inference, rendering, virtual desktops, video, and simulation. When a project moves from generating images to rendering them, the same server takes the new workload.
- Fine-tuning and serving in one place. As Stelia AI’s CTO explained, fine-tuning now runs on this inference-class GPU. Teams adapt an open model and serve it from the same machine on one budget line.
- A fixed cost you can plan. Dedicated servers run on monthly or yearly contracts, so the price stays the same however many tokens your agents consume. The more of that capacity you use, the lower your cost per token.
Which GPU server setup fits your workload
Match the GPU to the workload. Here’s how common situations map to a setup:
| Your situation | Setup | Why it fits |
|---|---|---|
| Serving one mid-sized open model in production (RAG, copilots, document processing) | 1× RTX PRO 6000 | The model and its working memory fit on a single card |
| Several smaller models, or teams that need separate environments | 1× RTX PRO 6000, split with MIG | Up to four isolated instances on one server |
| Steady, growing traffic on a proven model | 2–8× RTX PRO 6000, running parallel copies | Throughput scales by adding replicas; without NVLink, you add copies rather than splitting one model |
| AI alongside rendering, virtual desktops, video or simulation | RTX PRO 6000 | One card covers both workload types |
| Large models of roughly 70B–405B parameters, or memory-heavy serving | 4–8× H200 with NVLink | Fast GPU-to-GPU links for models that span cards |
| Pre-training large models from scratch | Multi-node GPU cluster | Needs fast interconnect across many servers |
It fits SaaS vendors adding AI features, media teams generating images and video, design and engineering firms, and regulated sectors such as finance and healthcare, where open models with published weights and training data give compliance teams something to audit.
If your model needs more than one card, our post on multi-GPU servers covers how to size 4- and 8-GPU setups.
How to test whether an RTX PRO 6000 server fits
Before you commit, answer four questions.
- How much GPU capacity does the application need day to day?
- What response times must it hold under load?
- What control do you need over data, access and configuration?
- How will costs change as usage grows?
Then run your actual model on the hardware with representative requests, and measure response quality, latency, throughput, and cost together. Those numbers make a stronger case for a purchase than any spec sheet.
Deciding where each AI workload runs
Treat “API or own infrastructure” as a policy you set per workload, based on cost, data sensitivity and the control you need. For mid-sized open models and mixed AI and graphics workloads, a dedicated RTX PRO 6000 server is the place to start.
Leaseweb offers NVIDIA RTX PRO 6000 GPUs on dedicated servers in Canada, the United Kingdom, the Netherlands and Germany, on monthly or yearly contracts. You choose the location, so you know which jurisdiction your data falls under before you deploy. For specs and configurations, see our GPU servers.
Our GPU specialists can set up that test with you as a proof of concept.