A vendor has already quoted the hardware
Before signing, you want an independent check on whether the proposed cluster matches the workload.
Enterprise engineering
Size the workload before you buy a single GPU.
Buying a GPU server or high-performance computing cluster before sizing the workload creates costly capacity, facilities and operating risk. Kritical assesses the workload, data, benchmark and complete infrastructure so you choose a workable architecture, and know what it costs, before you commit.
On this page
Choose where to start.
Jump to the part that matches the problem or decision in front of you.
Use the arrows or scroll to see every card.
The landscape you are choosing between.
Hosted AI platforms, including Microsoft's Azure AI Foundry and Copilot, make it easy to start without buying hardware, and for many workloads that is the right call. On-premises makes sense when data residency, latency, cost at scale or a specific accelerator investment pulls the other way. You shouldn't have to guess which side of that line your workload sits on, and a vendor with hardware to sell shouldn't guess it for you. We size the workload first and let the answer follow the numbers.
Three situations bring teams to this page.
Data sensitivity rules out a hosted option
The information the model needs to see cannot leave your environment, and a local deployment is being seriously evaluated.
A pilot worked, and now capacity has to scale
What ran fine on a workstation now needs a real production design.
What you get: a decision, not just a recommendation.
Start with the task, users, sensitive information, response time and systems the model must reach. We compare local, hosted and hybrid approaches against those needs. For an on-premises design, we also account for logging, connectors, updates and administrative access so the security decision covers the whole system.
Size the complete workload.
Model size is only one input. Context length, concurrency, training or inference pattern, retrieval, embeddings, datasets and evaluation affect accelerator and CPU selection plus compute, memory, storage and network requirements. We scope a representative benchmark and cluster assumptions before procurement. Power, cooling, space and support requirements, including rack density, HVAC capacity and replacement paths, are part of the same design.
Context and concurrency
Test context length and concurrency against representative use.
Retrieval and embeddings
Include retrieval and embeddings in the workload model.
Compute, memory, storage and network
Size compute, memory, storage and network requirements together.
Facilities and support
Consider power, cooling, space and support alongside accelerator capacity.
Why Kritical for on-premises AI.
We are not selling a GPU box; we are pricing the engineering work that decides whether it is the right box.
We explain the vendor landscape in plain words
Azure AI Foundry, Copilot and the on-premises alternative, compared against your actual constraints.
Benchmark before you buy
A representative benchmark and a limited pilot establish whether proposed hardware is appropriate before a larger purchase.
The whole facility is part of the quote
Rack, power, cooling and support requirements are priced alongside the compute, not discovered after delivery.
How it is priced, and what a first engagement leaves behind.
A first engagement is quoted in writing and can produce a workload brief, a data and security boundary, sizing assumptions, a benchmark plan, a target architecture, a bill of materials, a facility dependency list, an implementation sequence, a verification plan and a support model. Product selection, stock, price and benchmark results remain subject to the current quote and the representative testing performed.
First call
We confirm the task, the data sensitivity and whether hosted, hybrid or on-premises is even the right question.
Get a quote
Send the systems and constraints involved and we quote the work.
Questions before you enquire
Does on-premises mean the system is completely offline?
Only if that is designed and verified. Dependencies such as model downloads, telemetry, identity and external tools must be assessed individually.
Should we buy GPUs before starting?
Start with the workload and a capacity assessment. A benchmark or limited pilot can establish whether proposed hardware is appropriate before a larger purchase.
Can you work with infrastructure we already own?
Yes. Include its specifications, available capacity and operating constraints. Suitability depends on the chosen workload and will need to be tested.
Why not just use a hosted AI platform?
Often you should. We compare hosted, hybrid and on-premises against your actual data sensitivity, latency, cost at scale and control requirements, and say so when hosted wins.
Can this run on VMware Cloud Foundation alongside our existing virtual estate?
Yes. We map the AI workload alongside the virtualisation platform so the accelerator, storage and network decisions connect to the rest of your private cloud rather than sitting in isolation.
Let’s define the next step.
Tell us what you want to achieve, what you have today and when it matters.