What it counts and how it works, explained. By Jillian Kaplan
AI infrastructure has to earn its place in the budget. A DPU is no exception: it should make you money or save you money, and the case has to hold up in front of finance.
CPUs and GPUs have decades of shared vocabulary for that conversation. DPUs do not, and the category is young enough that most published claims come from vendors. So when we built a total-cost-of-ownership calculator for the E1, we kept it small on purpose. It asks for four things about your environment and reports what those inputs produce over a four-year window. The reasoning behind those numbers is what this post is for.
The cards are not free, and the model does not pretend otherwise. Their cost is subtracted before anything is reported as a return, so the ROI and payback figures are what is left after you have paid for them.
The savings are counted per node, so they scale with the size of the deployment. The tool asks for your node count first for that reason.
One slider, two kinds of value
The E1 runs the host’s infrastructure work on the card itself. That frees up GPU time. One slider in the calculator sets how much of that time you expect to use. Bank it, and the model counts it as lower power and cooling cost. Put it back to work, and the model values it as compute at the GPU hourly rate you enter. Each gain flows through one path, never both, so the model cannot double-count the same benefit.
Set the slider to zero and you get the base case: infrastructure savings only. One 800G device replaces several slower NICs, and the power and cooling savings scale with your facility’s PUE. At the hyperscaler defaults and 1,000 nodes, that base case returns $4.1 million over four years and pays back in 42 months, before any GPU value is counted.
The tool opens at 50 percent for hyperscaler training and 25 percent for enterprise. Start at zero, check that the base case holds in your environment, then move the slider to the share you believe. As you drag, value shifts out of power savings and into recovered GPU value. The capital savings do not move, because they do not depend on what you do with freed GPU time.
Where the defaults come from
Defaults load by scenario. Hyperscaler training assumes wholesale power at $0.07 per kWh, a PUE of 1.12, and an 8-GPU HGX B200 reference node. Enterprise training assumes $0.10 per kWh, a PUE of 1.30, and a DGX B200 node. The GPU hourly rate defaults to $2.65 in both. Those four inputs are yours to change. Everything else is held at the reference architecture, and the full set of assumptions is available from us on request.
What the card costs
An E1 costs more than the NICs it displaces. The model carries that difference, and the calculator nets the card cost into the ROI and payback figures it reports. What pays for the card is what it removes from the host, which is CPU capacity along with the per-core licensing and the power that ride on it.
What it reports
The calculator reports four numbers over a fixed four-year window: total value created, payback in months, ROI, and GPU-hours freed. Total value created is the gross figure, before the cost of the cards; payback and ROI are net of it. A breakdown shows where the value comes from, and the fine print states the reference architecture, what the ROI accounts for, and how GPU productivity is valued.