GPUs for Enterprise AI
How GPU memory, throughput, concurrency and model size shape private AI infrastructure decisions.
GPU selection is a capacity-planning problem, not a shopping exercise. Model size, quantisation, context length, concurrency and response-time targets determine what the platform actually needs.
GPUs are usually the largest infrastructure cost in private AI. Understanding how memory and batching behave prevents both expensive overprovisioning and disappointing user experience.
Measure with your model and workload. Published tokens-per-second figures rarely describe a multi-user enterprise service.
There is rarely one universally correct architecture. The best decision is the one that fits the process, risk, people and operating capability of the organisation.
Why it matters
GPUs are usually the largest infrastructure cost in private AI. Understanding how memory and batching behave prevents both expensive overprovisioning and disappointing user experience.
In practice, value appears when the technology becomes part of a governed process: responsibilities are clear, evidence can be checked and the organisation can observe whether outcomes improve.
What it really means
GPU selection is a capacity-planning problem, not a shopping exercise. Model size, quantisation, context length, concurrency and response-time targets determine what the platform actually needs.
The useful distinction is between a technical capability and a production service. Enterprise use requires identity, permissions, data handling, evaluation, monitoring and a clear owner for the result.
How it works in practice
The flow can be described in five steps. Each one needs an explicit purpose, an owner and a way to verify that it behaves as expected.
The exact architecture will vary. What matters is keeping the boundaries visible and making failures diagnosable rather than hiding them behind a fluent answer.
What this looks like in real work
An organisation expects fifty active users but only eight concurrent generations. Testing shows that two GPUs with workload routing provide better resilience and latency than one larger device running at full capacity.
The lesson is not that AI produced an answer. It is that the answer appears inside a controlled process, can be verified and is tied to a measurable outcome.
Common mistakes we see
Sizing by employee count
Active concurrency and workload shape matter more than total licences.
Looking only at VRAM
Memory bandwidth, interconnect and software support also affect throughput.
Assuming linear scaling
Multiple GPUs introduce communication and scheduling overhead.
Ignoring power and cooling
Enterprise deployment must account for sustained load, not desktop peak figures.
When it makes sense
- The process has a clear owner and outcome
- The required information and permissions are understood
- Results can be checked or measured
- The organisation can operate the capability responsibly
When it probably does not
- The problem is still undefined
- A deterministic rule would be simpler and safer
- No one owns data quality or exceptions
- The consequence of failure cannot be controlled
Not using AI can be the right decision. Honest scope is usually more valuable than a technically impressive pilot without a real problem.
How we usually approach it
We work from engineering and operational experience. We do not believe in a universal recipe, but we consistently ask the same questions before building:
Once these answers are reasonably clear, choosing models, infrastructure and integration becomes much easier.
Useful enterprise AI is not defined by how impressive the demo looks, but by how reliably it improves a real process.
Technology evolves quickly and we do not claim to have final answers. These guides capture what we have learned while building and operating platforms, shared with the humility that tomorrow may require a better approach.