Independent advice on buying AI compute, and the supply chain to deliver it
We size the requirement, run the market, negotiate and land the hardware — and say plainly when renting is the better answer.
Honest position
We start from the workload, not the budget. Batch shapes, concurrency targets and the growth curve you actually believe in set the accelerator count. Then we model buy, lease and rent on one three-year horizon and show you where they cross.
Selection
Memory bandwidth decides more outcomes than headline FLOPS, and software maturity decides more than either. We say where the cheaper part is fine and where it costs six months.
Crossover
60%
Sustained utilisation above which owning typically beats renting
Landed cost
5%
GCC common external tariff on most hardware entering the customs territory
Supply
Top-bin accelerators are sold on allocation, not from a catalogue. We hold multi-vendor relationships and take the refurbished market seriously where provenance checks out.
For investors
Utilisation, power price, contract structure and residual value decide the return, and diligence usually skips them. We model all four and stress the cases that break them.
Accelerators
Platforms
Fabric and storage
Consumption models
Commercial
Who we do this for
2–3 weeks
Workload sizing, a buy-versus-rent model in your own figures and a written recommendation; many clients take it and buy elsewhere.
Sizing report, buy-lease-rent comparison, indicative bill of materials
6–14 weeks
Competitive quotation rounds, negotiation, purchase, import and delivery to site with warranty and spares arranged.
Normalised multi-vendor quotations, landed cost plan, warranty and spares strategy
Ongoing
Retained advisory across a multi-phase build: refresh planning, allocation management, negotiation and residual value tracking.
Quarterly utilisation review, refresh plan, residual value model
When utilisation will sit below roughly 60%, when the workload still changes shape every quarter, when the site has no power and cooling headroom, or when the requirement is modest inference a rented endpoint serves well. Buying converts a flexible operating cost into an illiquid asset on a steep depreciation curve.
The group does sell hardware through a separate entity, and we state that at the outset. In practice that works in your favour: we quote multi-vendor, show the comparison including offers we are not supplying, and produce the sizing model before any bill of materials exists. The review alone is a supported outcome.
For inference fleets, development capacity and some fine-tuning work, yes — often the strongest value in the market. The conditions are non-negotiable: verified provenance, documented burn-in, thermal and ECC testing before acceptance, and a warranty backed by somebody who will still exist in two years.
Send us the constraint you are actually up against — budget, latency, regulator, deadline.
