On‑prem AI 'factories' claimed 28%–90% savings

Social posts on Tuesday pushed a familiar claim in AI infrastructure: that companies running persistent AI agents can spend less by owning or leasing dedicated compute instead of paying public-cloud rates. The most widely shared figure in those posts was a 28% to 90% reduction in operating cost for “on-prem AI factories,” a term now used across the industry for dedicated AI infrastructure deployed in private facilities or customer-controlled sites. The posts did not appear to include a public spreadsheet or audited model behind that exact range. But the underlying argument — that steady, high-utilization AI workloads can be cheaper on owned or reserved infrastructure than on metered cloud — is consistent with recent vendor analyses and product launches. Lenovo said in a 2025 total-cost-of-ownership paper that cloud pricing can become expensive for long-running AI jobs, while on-prem systems can become more efficient over time when GPU use is consistent. ### Where did the “AI factory” language come from? AWS used the phrase publicly at its re:Invent 2025 conference. Amazon said in December 2025 that it was offering “AWS AI Factories” for implementing AI infrastructure in customers’ existing data centers, alongside new Trainium3 systems and agent products. The phrase is broader than one company’s branding. In current usage, it generally refers to clusters of GPUs, storage, networking and orchestration software set up for repeated AI training or inference, often in enterprise data centers, colocation sites or regional facilities. (lenovopress.lenovo.com) DataBank, in a May 1, 2026 analysis, described AI infrastructure decisions as a choice among public cloud, on-premises and colocation, with GPU utilization, power and cooling, and compliance all affecting total cost. (aboutamazon.com) ### Why would persistent agents be cheaper off cloud? Persistent agents run continuously. That matters because public cloud bills by time used, while owned or contracted infrastructure spreads costs across a longer period. Lenovo’s paper said cloud is well suited to short-term or bursty workloads, but that usage-based pricing can drive up long-term costs. The paper said on-prem systems can gain cost efficiency through consistent utilization, and it cited very large GPU-hour requirements for modern model training. (databank.com) DataBank made the same point in different terms. Its May 2026 analysis said AI breaks older IT cost models because workloads involve persistent GPU utilization, long-running training and inference cycles, and added charges for storage, networking and data egress in public cloud. ### Does outside research support the exact 28% to 90% range? (lenovopress.lenovo.com) The exact range remains unverified from a primary-source financial model. No public filing, cloud provider document or independently published benchmark located in this review set out that specific 28% to 90% band. What outside material does support is direction, not that precise number. (databank.com) GMI Cloud said on April 27, 2026 that on-prem becomes more cost-effective when teams run roughly 10 to 11 GPUs continuously at full utilization, and said sustained workloads can produce multi-year savings despite higher upfront spending. Google Cloud’s pricing page also says spot pricing can cut some GPU costs by 60% to 91% versus on-demand rates, which shows how widely cloud economics can swing depending on commitment and interruption risk. (lenovopress.lenovo.com) ### What costs get left out of the simple social-media claim? Power, cooling, staffing and hardware obsolescence can change the math. Lenovo said its 2025 paper intentionally excluded some secondary costs, including networking, facility overhead and routine IT operations, to keep the comparison focused on core infrastructure. DataBank said on-prem also carries risks: heavy upfront capital spending, long depreciation cycles, overprovisioning for peak demand, and retrofit costs for power and cooling. (gmicloud.ai) Its colocation analysis said some buyers use third-party facilities to avoid a full facility build-out while still amortizing dedicated GPUs over time. ### What should readers watch next? (lenovopress.lenovo.com) Cloud pricing pages and vendor TCO papers are the next place to look for harder numbers. Google Cloud’s current GPU pricing page, AWS’s P5 instance page and vendor analyses from Lenovo and DataBank provide the clearest published benchmarks for comparing metered cloud with dedicated infrastructure. (cloud.google.com) (databank.com)