From the Dell edition of October 30, 2026
Dell sales chief says 80% of tokens have shifted to open models on cost and inference is moving to the edge; independent gateway data only partly supports the claim
Dell Technologies (interview reported by SiliconANGLE/theCUBE) · Technology vendor (applies to SLED buyers) · Global (Dell Technologies New York Symposium)
- Publisher
- SiliconANGLE
- Original publication
- 2026-10-23
- Source retrieved
- 2026-10-31
What happened
Kyle Dufresne, Dell senior vice president of sales, said that within three months the share of tokens generated by frontier models fell from 80% to 20%, with 80% now generated by open models 'because of cost'; that inference is growing faster than training and spreading from cloud to data centers to PCs and the far edge; that CFOs report exploding token costs; and that memory supply will grow about 20% next year against roughly 200% more demand, so many customers buy earlier. Independent gateway data inspected for scrutiny (OpenRouter, Oct. 2026) shows open-weight models at 71.7% of top-50 model tokens on its platform in late September 2026, up from 26.8% a year earlier; a separate summary of ETR survey data puts open-weight models at 34% of enterprise token usage.
Why it matters
If the cost shift is real, agencies and universities paying per-token for cloud AI may find on-premises or edge inference on owned hardware worth evaluating, especially for data that cannot leave their control. The evidence does not show the shift in the public sector, and the 80% figure is not sourced.
Evidence and measured results
Executive interview claims without disclosed data. OpenRouter's platform data directionally supports a large open-weight shift among routed developer traffic, but enterprise survey data cited by WEREA shows a much lower share (34%) and spend still concentrated on closed models; neither covers SLED.
Limitations and uncertainty
Vendor sales executive with an interest in on-premises demand; no source for the 80% figure; paid media partnership disclosed; partially overlaps an archived Oct. 21 SiliconANGLE item on Dell supply constraints, but the token-economics claim is new. Gateway data over-represents developers and startups.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-10-31; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Public-sector AI users face rising metered inference costs and want to keep sensitive data under their control.
- Stakeholders
- CIO/CTO
- CFO or budget office
- Data governance lead
- Procurement
- Program owners using AI
- Discovery questions
- What is current monthly spend on metered AI services?
- Which tasks are high-volume and lower-risk?
- Which data cannot leave agency control?
- Do state rules restrict certain open-weight models?
- Value hypothesis
- For high-volume, lower-risk tasks, self-hosted open-weight models on owned hardware may lower unit cost; this must be measured locally.
- Potential engagement
- A cost-and-placement assessment comparing current cloud AI spend with an on-premises open-model pilot for one workload.
- Unsupported claims
- That 80% of tokens have moved to open models in any market segment
- That on-premises inference is cheaper for every workload
- Any SLED-specific adoption figures
Pre-sales engineering
Role takeaway
- Fit
- Relevant to agencies with sustained inference volume and data-control needs.
- Architecture and integration
- Open-weight models served on PowerEdge or AI PCs alongside existing cloud AI services, with routing by task and data sensitivity.
- Prerequisites
- Workload and cost baseline
- Model quality benchmarks per task
- GPU/memory sizing
- Constraints
- Hardware lead times and memory price inflation
- Model license and provenance restrictions
- Scarce MLOps skills
- Security
- Self-hosting moves model patching, access control and logging to the agency; apply the same controls as other sensitive systems.
- Proof of value
- Compare quality of open versus current models on agency tasks
- Compare full cost per task including operations
- Test latency at the edge versus cloud
Delivery
Role takeaway
- Work
- Run a single-workload pilot of a self-hosted open model and compare it with the current service.
- Dependencies
- Hardware availability
- Model selection and license review
- Security review
- Ownership
- Agency IT with the program owner of the pilot workload.
- Skills and adoption
- Requires MLOps, GPU operations and evaluation skills; plan training or partner support.
- Governance checkpoints
- License and provenance approval
- Security hardening review
- Go/no-go on measured cost and quality
- Acceptance criteria
- Proposed: quality on the pilot task within an agreed tolerance of the current model
- Proposed: documented full cost per task
- Proposed: rollback tested
- Risks
- Quality drop on harder tasks
- Hidden operating costs
- Supply delays
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
Supports evaluating a placement policy across cloud, data center and edge for inference; open-weight models make local hosting possible but add hosting, patching and sizing work.
Governance
Who approves, reviews and stays accountable for outcomes?
Model choice now carries cost, licensing and provenance questions; open-weight model licenses vary and some popular models come from Chinese labs, which may trigger state restrictions.
Security and privacy
What data, permissions and controls need testing?
Local inference can keep sensitive data in-house; self-hosting shifts model security, patching and access control to the agency.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Self-hosting open models requires MLOps skills many public IT shops lack.
Procurement
What should contracts, pricing and exit terms secure?
Memory-driven price rises and the advice to buy earlier conflict with annual budget cycles; agencies may need multi-year or contingency authority.
Operating model
Which teams own the service once it runs?
Shifts cost from metered cloud tokens to capital assets and internal operations; needs a FinOps-style view of AI cost.
Publication history
- 2026-10-30Dell · Issue 024 resources
Stable resource ID: dell-exec-open-model-token-shift-edge-inference-siliconangle-20261023