NVIDIA’s CES presentation on 5 January 2026 introduced Rubin as a platform built around six co-designed chips, connecting compute, networking and data movement. The announcement put the surrounding system at the centre of the performance story. The commercial question for an infrastructure business is how much of that system it can deploy, operate and sell.
Define the unit of investment
An accelerator price is a useful input to procurement. It is an incomplete basis for approving a service. Our view is that the investment unit should be the smallest complete environment capable of delivering the intended workload: compute, memory, storage, fabric, power, cooling, software and support.
That definition changes comparisons. A configuration with a lower hardware price may require more integration work or a longer qualification period. A tightly integrated platform may reduce some engineering effort while concentrating dependencies on a particular vendor. Both consequences belong in the decision paper.
Put the dependencies on one schedule
Consider a delivery in which the equipment arrives before the facility can energise it. Asset ownership has begun; useful capacity has not. Financing, support and space commitments may already be running while customer acceptance remains some distance away.
We would ask for a single schedule covering facility readiness, delivery, installation, software qualification and customer acceptance. Each milestone should have an accountable owner and evidence of completion. Where dates depend on third parties, the plan needs an explicit allowance for delay and a decision on who carries its cost.
Underwrite a usable service
The final review should connect the technical design to a recognisable customer requirement. Which jobs will run? How long can they wait? What performance must be repeatable? What happens during maintenance? How will the environment be supported after the initial engineering team moves on?
A useful sensitivity analysis then varies utilisation, time to acceptance, power cost and the price customers are prepared to pay. Benchmark results can inform that analysis, but they do not establish demand or operating margin.
The approval record should connect the complete service, its dependencies and its operating economics. That gives engineering, finance and delivery teams the same basis for the next decision.