- Sadeepa Wijesekara of HPE told The TWIML AI Podcast that owned AI hardware can break even below 60% use. HPE sells that hardware.
- She also said a hybrid mix is wiser than going fully on-premises, and that teams should itemise token spend per agent and use case.
Sadeepa Wijesekara runs the engineering team behind HPE's Private Cloud AI product. On The TWIML AI Podcast she argued that running your own AI hardware can pay off even when it is not busy all the time. She works for the company that sells that hardware. So the break-even figure is a seller's claim, and a buyer should test it before relying on it.
What was said
Host Sam Charrington asked how any company could beat OpenAI or Anthropic on the cost of running models. Wijesekara said to compare the same model, or the same GPU, in the cloud and on-premises. Do it over three or five years, she said, and the difference shows.
She said HPE's calculators are available to customers through its sales team. They assume 60-70% utilization, meaning the share of time the GPUs are busy, and still show a return against public frontier models. Break-even, she said, is probably below 60%, because the price gap is large. She did not claim you need full use.
Her example was scheduling. Developers use the system by day. At night, agents can scan a whole code base for security flaws and propose fixes. Wijesekara said that is how you get the best return on the GPUs.
She described the end state this way: "you become your own token factory inside your premise." She also warned that one headline figure for token spending hides which agent or task is driving the bill.
Why it matters
Our reading: renting by the token and owning hardware are both choices with a price. The break-even point depends on inputs the seller's calculator sets, such as how busy your GPUs will be. If your own work cannot be queued overnight, the 60-70% assumption may not hold.
Her own advice to itemise spend per agent and task gives buyers a way to check. Measure where your calls go. Then ask the vendor to rerun the sums with your numbers, not theirs.
The other side
Wijesekara did not argue for owning everything. She agreed that the frontier labs will keep the better models, and said the strategy should be hybrid, not just on-premises. She also said most companies should not assemble data-centre parts themselves.
In the excerpts, the host raised the scale objection and asked whether the 60-70% figure was low, but no one examined what else the calculator assumes. She said she was not part of HPE's own rollout. She described sizing the system as a calculated guess that the team then scaled up.
She also suggested that teams may be worrying too much about cost, and that risk questions deserve more attention. That is her view, and it is consistent with selling a private system.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- The TWIML AI Podcast: Buying Tokens vs Building an AI Factory with Sadeepa Wijesekara - #785 (2026-11-17)





