Dell pitches on-prem inference to Bay Area customers whose token budgets run out in weeks
Dell's Helen O'Sullivan says some Bay Area customers run through token budgets in weeks that they allotted for months. Her fix keeps frontier models and adds open-source models and on-prem infrastructure.
The Product Desk
Wood says existing CPUs may have headroom for some next-generation workloads, and open-weight sandbox tests help assess needs. O'Sullivan saw heavier CPU use by agents but says customers are not quite there; many buyers remain GPU-focused. Dell and Intel weigh capital versus operating expense.
- capability Enterprises' existing CPUs May have headroom for some next-generation apps and workloads, Intel's Wood says, claim 2
- contradiction Customers, per O'Sullivan She saw heavier CPU use by agents, but says customers are 'not quite there yet', claim 6
- constraint GPU-focused buyers Many remain focused on GPUs even as their workloads evolve, O'Sullivan notes, claim 7
- capability Companies testing models Sandbox tests of open-weight models can help assess what workloads require, Wood says, claim 3
- decision Dell and Intel Weigh existing infrastructure and capital versus operating expense models before recommending a plan, claim 5
| Who | How | Kind | Claim |
|---|---|---|---|
| Enterprises' existing CPUs | May have headroom for some next-generation apps and workloads, Intel's Wood says | capability | 2 |
| Customers, per O'Sullivan | She saw heavier CPU use by agents, but says customers are 'not quite there yet' | contradiction | 6 |
| GPU-focused buyers | Many remain focused on GPUs even as their workloads evolve, O'Sullivan notes | constraint | 7 |
| Companies testing models | Sandbox tests of open-weight models can help assess what workloads require, Wood says | capability | 3 |
| Dell and Intel | Weigh existing infrastructure and capital versus operating expense models before recommending a plan | decision | 5 |
What happened
- O'Sullivan added that many buyers remain focused on GPUs even as their workloads evolve.
- Intel's Tim Wood said many enterprises already have CPUs in their data centers with headroom for some next-generation workloads.
- The interview ran on theCUBE, a paid media partner for Dell's AI Leadership Symposium, and Dell sponsors the event coverage.
Why it matters
- contradiction O'Sullivan calls the budget squeeze immediate, yet the one hardware change she describes, heavier CPU use by agents, is something she says customers have not fully reached.
- decision Whoever owns an agent budget has to choose, workload by workload, which ones get an open-weight sandbox test before anyone orders hardware.
- cost Moving inference shifts part of the bill from tokens to the capital or operating-expense plan Dell and Intel weigh, and that cost lands on the infrastructure budget.
The person holding an exhausted token budget gets a hybrid answer from Dell. "You still use those frontier models, but you use the open source and infrastructure on-prem to at least eliminate some of the token stress and token maxing going on right now," Dell's Helen O'Sullivan said [1]. Frontier models stay in the mix, and the relief on offer is partial [1].
SiliconANGLE ties the pressure to agentic workloads multiplying and says on-premises infrastructure and open-weight models are "moving back into the conversation" [11][9]. On the hardware side, O'Sullivan said: "We saw the rebalance and the heavier utilization on the CPU when it comes to agents. Customers are getting there," and added: "They're not quite there yet from the scope I'm seeing." [6] The report does not say how many customers are affected or whether any has moved a workload [10].
Both speakers work on the sales side: O'Sullivan in business development at Dell Technologies, Wood as Intel's sales director for the Northwest region [8]. The budget problem itself is described only by O'Sullivan [12]. Wood's advice is to test first. He said testing open-weight models in sandbox environments can help companies assess what their workloads require [3].
We'd sort each agent workload that is draining the budget on two questions. The first is whether an open-weight model holds up in a sandbox. The second is whether the company already runs a data center with spare capacity, and Wood said a lot of enterprises have CPUs there with headroom [2].
A workload that passes both is the first candidate to move. One that passes the model test but lacks capacity becomes the capital versus operating expense comparison that Dell and Intel say they weigh before recommending a plan [5]. One that fails the model test stays on the frontier model, and the practical control there is a spend cap. The cost of the sort is engineering time spent before any token is saved, and a workload that fails the test keeps paying per token.
What to watch
- Whether Dell or Intel puts a customer name, a budget figure or a migration count behind the weeks-versus-months claim.
- Whether the CPU shift O'Sullivan expects for agents shows up in what buyers order, given her point that many remain focused on GPUs.
- Whether anyone publishes sandbox results for open-weight models on agent workloads that are draining cloud token budgets.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence22
- Adoption
- Insufficient
- Hype gap+35
- Incentives85
- Confidence60
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
"You still use those frontier models, but you use the open source and infrastructure on-prem to at least eliminate some of the token stress and token maxing going on right now," O'Sullivan said.
- [2]
Tim Wood of Intel said: "A lot of enterprises already have CPUs in their data center. There may be existing resources and headroom in the infrastructure they already have that's capable of doing some of these next-generation apps and workloads."
- [3]
Wood explained that testing open-weight models in sandbox environments can help companies assess what their workloads require.
- [4]
TheCUBE is a paid media partner for Dell's AI Leadership Symposium, and Dell is the sponsor of theCUBE's event coverage; SiliconANGLE says neither Dell nor other sponsors have editorial control over content on theCUBE or SiliconANGLE.
- [5]
Dell and Intel are both taking a consultative approach, weighing existing infrastructure and capital versus operating expense models before recommending a plan.
- [6]
O'Sullivan said: "We saw the rebalance and the heavier utilization on the CPU when it comes to agents. Customers are getting there," and added: "They're not quite there yet from the scope I'm seeing."
- [7]
Many buyers remain focused on graphics processing units even as their workloads evolve, O'Sullivan noted.
- [8]
Helen O'Sullivan is AI business development manager and solutions specialist at Dell Technologies Inc.; Tim Wood is sales director for the Northwest region at Intel Corp.
- [9]
On-premises infrastructure and open-weight models are moving back into the conversation alongside frontier models in the cloud.
- [10]
The SiliconANGLE report does not state how many customers are affected by token budget exhaustion or whether any customer has moved a workload on-prem.
- [11]
As agentic workloads multiply, cloud token bills are consuming allocations faster than planned, prompting companies to rethink where inference should run.
- [12]
"A lot of our customers I support in the Bay Area, they're dealing with this real immediate struggle of running through their token budget in a matter of weeks that they had allotted for months," Helen O'Sullivan of Dell said.
Get this ledger inside ChatGPT. Ask about any Clarity Today story and the answer comes with its figure, every claim with its verdict, and the sources. Free, no account. Add it to ChatGPT or Claude
Sources
1 independent publisher whose own reporting we read for this story.
- siliconangle.comDell and Intel weigh the costs of enterprise AI scalability
1 article · October 12, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
- Enterprise AI inference costsFollow
- Agentic AI and tool useFollow
- Open-Weight Model PublishingFollow
- On-premises AI infrastructureFollow
Entities
- theCUBEFollow
- Helen O'SullivanFollow
- SiliconANGLEFollow
- Dell Technologies Inc.Follow
- Tim WoodFollow
- IntelFollow