{"resourceId":"lenovo-onetera-on-premises-local-government-ai-cost-pilot","versions":[{"version":"external-37f249f6f32382192fab1a1c8640af6abc2507e2aef0b02f4eeaee6596d7ff6e","resource":{"id":"lenovo-onetera-on-premises-local-government-ai-cost-pilot","title":"Lenovo customer story: Onetera's 40-day on-premises validation pilot on a ThinkSystem SR675 V3 reports up to 98% cost savings for token-heavy local-government workloads","organization":"Lenovo; Onetera","sector":"Local government AI infrastructure (on-premises inference economics)","geography":"North America","publishedAt":"Undated customer story","publicationDate":null,"eventDate":null,"sourceName":"Lenovo customer success stories","sourceLabel":"Vendor case study","sourceUrl":"https://www.lenovo.com/us/en/case-studies-customer-success-stories/onetera/","evidenceClass":"vendor-claim","outcomeClass":"emerging","topics":["infrastructure","governance-procurement"],"finding":"Lenovo says Onetera worked with the Lenovo AI Center of Excellence to validate the breakeven and token economics of on-premises infrastructure for its municipal AI services. LLMs were deployed on a Lenovo ThinkSystem SR675 V3 server (NVIDIA processors listed), every pipeline stage stayed inside the infrastructure boundary, and during the 40-day pilot only one of the server's eight GPUs was actively used. Stated outcomes are cost savings of up to 98% for token-heavy workloads (against high cloud API pricing), lower cost making agentic AI more accessible to local governments, and control over where data is stored and flows.","sledRelevance":"Why now: it is the infrastructure-cost claim behind Onetera's Oct 25 on-premises pitch to cities. Relevant to state and local IT leaders weighing cloud API spend against owned GPU servers, with the caveat that the pilot was run by the vendor, not a government.","evidence":"Vendor customer-story summary (linked PDF not opened). 'Up to 98%' is a best case with no stated workload mix, comparison price, utilisation assumptions or amortisation period; one GPU of eight was used, so full-server cost allocation is unclear.","architectureImplications":"Supports buying a multi-GPU server and running open models locally for high-token workloads; 1-of-8 GPU utilisation implies large headroom or overprovisioning.","governanceImplications":"Cost comparisons against cloud APIs depend on assumptions public buyers should require in writing, including hardware refresh, staffing and power.","securityPrivacyImplications":"Keeping all pipeline stages inside the infrastructure boundary supports data-sovereignty requirements; the controls themselves are not described.","caveats":"Undated; vendor-run pilot by a software company, not a city; 'up to' figure; methodology in an unopened PDF; hardware vendor and customer both benefit from the claim.","streamIds":["nvidia"],"roles":{"sales":"Interpretation: the 98% figure will attract attention, but it is an 'up to' best case from a vendor-run 40-day pilot with one of eight GPUs active. Use it to open a total-cost-of-ownership discussion with city IT and finance about when owned GPU servers beat cloud APIs, and insist on the prospect's own volumes. It supports a data-sovereignty argument more firmly than a savings promise.","engineering":"Interpretation: the configuration is a ThinkSystem SR675 V3 multi-GPU server running open LLMs locally. That one GPU handled the pilot suggests modest workloads fit far smaller hardware, so right-size before quoting an eight-GPU server. Pre-sales should model utilisation, redundancy (one active GPU is a single point of failure), power and cooling, and compare against API costs at realistic token volumes. Also check open-model licences for the models served.","delivery":"Interpretation: delivering on-premises AI for a city means operating hardware: patching, monitoring, capacity planning and failover. Define proposed acceptance criteria around cost per completed case against documented prior API spend over a full billing cycle, and confirm the city or a managed provider can staff operations before purchase. Plan capacity tests before purchase, since the pilot left seven of eight GPUs idle."},"roleDetails":{"sales":{"customerProblem":"Token-heavy agent workloads make cloud API costs unpredictable for local governments that also want data kept in their own environment.","stakeholders":["City CIO / IT director","Finance or budget office","Procurement","Program owners using AI services"],"discoveryQuestions":["What are current monthly token volumes and API costs?","Who would operate and patch on-premises GPU servers?","What refresh cycle and power costs should be included?"],"valueHypothesis":"Interpretation: steady high-volume workloads may cost less on owned hardware than via APIs; the size depends on utilisation and is not established for any city here.","potentialEngagement":"Total-cost-of-ownership comparison for one AI workload, cloud API versus on-premises GPU server.","unsupportedClaims":["98% savings for a city","That an eight-GPU server is needed","That on-premises is cheaper at low or bursty volumes"]},"engineering":{"fit":"Agencies with steady, high-token workloads and data-residency requirements.","architectureIntegration":"Open LLMs served on a local multi-GPU server inside the city network, replacing calls to external APIs.","prerequisites":["Server-room capacity with power and cooling","Operations staff or managed service","Workload volume data"],"constraints":["Pilot used 1 of 8 GPUs, so sizing evidence is thin","Methodology in unopened PDF","Vendor-run, not government-run"],"security":"Keep all stages inside the boundary as claimed and add standard server hardening, access control and log retention.","proofOfValue":["Measure cost per completed task over a full month on real volumes","Load-test to find the GPU count actually needed"]},"delivery":{"work":"Install, size and operate a local inference server and migrate one workload from API to local serving.","dependencies":["Facilities (power and cooling)","Hardware vendor support","Model licence review"],"ownership":"City IT infrastructure team, or a managed service provider under contract.","skillsAdoption":"IT staff need GPU server operations and model-serving skills.","governanceCheckpoints":["TCO review signed off by finance","Licence review for open models"],"acceptanceCriteria":["Proposed: measured cost per completed task below prior API cost over one billing cycle","Proposed: availability target met with documented failover"],"risks":["Under-utilised capital purchase","Single-GPU failure","Staff skill gaps"]}},"retrievedAt":"2026-10-26T03:02:15Z","enrichedAt":"2026-10-26T03:14:13Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Owning GPU servers requires operations skills many small governments lack.","procurementImplications":"Cities weighing on-premises against API pricing should ask for total cost of ownership, including capital, maintenance, staff and energy, not per-token comparisons alone.","operatingModelImplications":"On-premises AI shifts cost from variable API spend to fixed capital and operations, which suits steady high-volume workloads more than sporadic ones.","sourceVerification":{"openedUrl":"https://www.lenovo.com/us/en/case-studies-customer-success-stories/onetera/","referenceExcerpt":"Onetera worked with the Lenovo AI Center of Excellence (CoE) to validate the accelerated breakeven and favorable token economics of on-premises infrastructure.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}