Validated reference architectures for the AI workloads enterprises actually run, built by people who built converged infrastructure the first time round. Configured to order through our Dell, HPE and distribution relationships. We publish what each tier is for and what it is not for, because the most expensive mistake in AI infrastructure is buying the tier above your workload.
The spectrum
Enterprise AI work falls on a spectrum, and the infrastructure follows from where you sit on it. Most organisations use existing models, a smaller number optimise them on their own data, and almost none build their own. The three pods map onto the first two. The third is a different conversation entirely.
Serve pre trained open models, answer questions over your own data with retrieval. Where most enterprise GenAI actually lives.
Adapt open models on your own data with periodic fine tuning, on top of everything the inference tier serves.
Train a foundation model from scratch. A different discipline, a different budget, and almost never the real requirement.
A pattern worth noticing: even the large infrastructure vendors' own AI workload maps place the bulk of enterprise generative AI at the inference end, not the training end, and their standard pod style bundles are built around inference and retrieval rather than training. The honesty on this page is not a contrarian take, it is where the workloads genuinely are.
The tiers
For serving open weight models, internal chatbots, document Q&A and retrieval over internal data. The tier most enterprises actually need.
Inference plus periodic fine tuning capacity, for organisations adapting open models on their own data. Explicitly not a training cluster.
The regulated tier. Sovereign inference with the security and data residency wrap for financial services, healthcare and critical infrastructure.
How buying works
Sizing conversation first, then a validated bill of materials configured to order. No pre built stock, no published SKUs going stale on a web page, and no obligation to buy the sizing outcome from us.
Specific components and pricing live in the sizing engagement, where they can be matched to your workload and negotiated properly, with our technology acquisition discipline behind the purchase.
The anti pitch
Most enterprise AI workloads are inference. Serving models, answering questions over your own documents, processing calls and content. That is what the estate will actually do all day, and it is what the infrastructure should be sized for.
Training grade infrastructure is the wrong buy for a chatbot estate. Not a stretch, not future proofing: the wrong buy, at several times the cost, with utilisation that will embarrass everyone at the first review.
Expansion headroom should be priced as an option, not sold as a default. If the workload grows, buy the growth when it is real. The target of this honesty is oversizing as a practice, not any vendor: the same Dell and HPE platforms we configure are excellent when sized to the workload in front of them.
Questions
A published, validated combination of compute, accelerators, storage, networking and software, sized and tested for a defined class of AI workload. A good one saves you designing from scratch and protects you from overbuying. The test of a good one is whether it also states what it is not for, because an architecture that fits everything has been sized for the largest case and priced accordingly.
Most enterprises need the Inference Pod: it covers serving models, chatbots, document processing and retrieval, which is the overwhelming majority of real enterprise AI. Move up only on evidence: the Fine Tuning Pod when you are genuinely adapting open models on your own data, the Private AI Pod when regulation or data residency demands sovereign inference. Start lower than you think and prove the need to climb.
It depends on the sizing, which is the honest reason we publish no prices: a published number is either padded or wrong. The sizing conversation produces a validated bill of materials configured to order through our Dell, HPE and distribution relationships, priced for your workload, with expansion offered as a costed option rather than built in by default.
Yes. The sizing outcome is yours, and there is no obligation to purchase through us. We think our vendor side negotiating experience gets you a better configured deal, but the design does not depend on it, and we would rather you deploy the right architecture than the right invoice.
Almost never. What enterprises run in practice is inference and retrieval, with some periodic fine tuning. Foundation model training is a different discipline with different economics, and buying training grade infrastructure to serve chatbots is the single most expensive category error in enterprise AI today. Size for the workload you have, and keep expansion as an option you exercise on evidence.
The advisory behind the pods is AI Infrastructure Advisory, and the buying discipline is AI Commercial Defence.
One sizing conversation and you will know which pod fits, what goes in it, and what it should cost. Or that the API is still the right answer, if that is what your numbers say.
Book a sizing conversation