🔍 Read the full analysis: The Promise And Business Case For Running AI On-Prem on ThorstenMeyerAI.com
Get smart everyday buys delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
IBM CEO Arvind Krishna said open-weight AI models can cost 10% to 25% as much as frontier models while delivering about 90% of their effectiveness, and can run on a few racks. An Artificial Analysis comparison read on October 26, 2026, put the best downloadable model at 80% of the leading proprietary model; the economics of buying hardware also depend heavily on how much it is used.
IBM CEO Arvind Krishna told investors on October 21 that open-weight AI models can cost 10% to 25% as much as frontier models while being “90% as effective,” making on-premises deployment a potential option for enterprises with suitable workloads. A model comparison published by Artificial Analysis and read October 26 put the leading downloadable model at 80% of the top proprietary model, while hardware cost estimates show that buying servers pays off only if they are used enough.
Krishna made the comparison during the question-and-answer portion of IBM’s third-quarter earnings call. He said an enterprise could begin with “two or three racks” and use open-weight models for routine work such as summarizing email, reviewing contracts and responding to customer calls. He also said organizations in finance may value control over where data and logs are processed. IBM sells on-premises hardware and services, so its cost and performance claims come from a company with a commercial interest in the deployment model.
The Artificial Analysis Intelligence Index scores models across ten evaluations. Its October 26 comparison listed MiMo-V2.6-Pro at 46.3 as the leading model with downloadable weights and Claude Opus 5.5 at 57.6 as the leading proprietary model. That makes the downloadable model about 80% of the proprietary leader’s score on this index, not 90%. The comparison is not a measure of every task or of the cost and quality of a particular enterprise deployment.
The source material’s hardware comparison puts an eight-GPU H200 server at a listed price of $172,643 and AWS rental of the same GPU class at $63.30 an hour. On those figures, server purchase cost equals about 114 days of rental at full utilization, or roughly 15 months at 25% utilization. Those estimates exclude staffing, power and cooling, networking, and other operating costs. IBM separately said its llm-d project can run large models at five to ten times lower cost; that is an IBM claim, not an independently verified measurement in the cited material.
IBM's 90% claim, checked
Open weights reach 80% of the leader
Krishna says open-weight models are "90% as effective". On an independent index the best model you can download today reaches 80% of the best closed one, up from about half two years ago.
Best open model as a share of the best closed
Artificial Analysis Intelligence Index, current version, best downloadable vs. best proprietary model at each date.
What Krishna said
Illustrations from the earnings call, not a named benchmark.
- Late 2024"70" vs. "20 or 30"Krishna, Q3 call, 21 Oct · no named benchmark
- Late 2025"70 to 50"Krishna, Q3 call, 21 Oct · no named benchmark
- Now"90% as effective"Krishna, Q3 call, 21 Oct · no named benchmark
Top downloadable models today
Intelligence Index score and licence type. Read the licence before the leaderboard.
| Model | Index | Licence |
|---|---|---|
| MiMo-V2.6-Pro Xiaomi | 46.3 | MIT |
| DeepSeek V4.1 Flash 1017 DeepSeek | 45.1 | MIT |
| GLM-5.3 Z AI | 44.8 | own commercial licence |
| Step 5 Preview StepFun | 43.7 | own commercial licence |
| Muse Spark Open Meta | 38.9 | licence with conditions |
In points the gap did not shrink: 7.5 points at the end of 2024, 11.3 today. Closed models moved too.
Sources: Artificial Analysis, Intelligence Index v4.3.2 (read 26 Oct 2026), IBM Q3 2026 earnings call transcript (21 Oct 2026), Artificial Analysis on Kimi K3.1 Pro (23 Oct 2026).
Powered by Thorsten Meyer AI
The Cost Depends on Workload
The business case is not simply that open models are cheaper. Companies must match model capability to each task, account for licensing and operating expenses, and compare owned hardware with cloud capacity they can scale or release. Routine text work may not need the most capable model, while complex or high-stakes uses could require a stronger system or more human review.
Ownership can offer more control over infrastructure and data handling, but it also shifts responsibility to the deploying organization. Krishna acknowledged that specialist staff for model kernels and networking can be a bottleneck outside the most technically sophisticated enterprises. A rack that is lightly used can erode the apparent savings; a heavily used system may make the purchase price easier to justify. The comparison needs a company’s actual workloads, staffing and utilization assumptions, not just list prices.
For readers weighing deployment options, the model score is only one input. Licensing terms determine whether weights can be used commercially and where they can be deployed. Hosted open models may also offer low token prices without requiring a company to own hardware. On-premises deployment is consequently one possible operating choice, not an automatic cost reduction.
Popular right now
Philips 4400 Series Fully Automatic Espresso Machine with LatteGo Milk System
Vitamix Propel Series 510 Professional-Grade Blender, 3 Programs, Black
Shark AI Ultra Robot Vacuum with Self-Empty Base
As an affiliate, we earn on qualifying purchases.
How the Model Gap Is Measured
Krishna described the quality gap using illustrative scores: in late 2024, he said, a top model might score 70 while an open model scored 20 or 30; by late 2025, the comparison was 70 to 50. The source material says these were not tied to a named benchmark. Artificial Analysis offers a separate comparison using the same ten evaluations across the models it tracks.
On that index, the best downloadable model was reported at 51% of the proprietary leader at the end of 2024, 74% at the end of 2025, 68% at the end of June 2026 and 80% on October 26, 2026. The gap in points did not steadily narrow: the source reports 7.5 points at the end of 2024 and 11.3 points in October 2026. The percentage comparison rose because the models’ scores and leaders changed over time.
A possible exception remains conditional. Kimi K3.1 Pro scored 55.1, close to Claude Opus 5.5’s 57.6, but its weights had not shipped as of the source’s October 26 review. Moonshot had promised them “by Nov 4.” Until weights are available, it is not a downloadable model for this comparison. The source also notes that licensing varies among models, and says a 2.8-trillion-parameter model would itself require substantial computing resources.
Open Weights and Real-World Costs
The cited benchmark does not establish that open models perform at 80% or 90% of proprietary models on every business task. The score is an index result, and the source does not supply independent tests of the specific workloads, error rates or review requirements at individual companies. It is also unclear how often a typical enterprise would keep an owned server busy enough to meet the stated payback estimates.
The cost comparison uses listed hardware and cloud prices, not a full assessment of ownership. Staffing, facility requirements, energy, networking, maintenance and financing could change the result. The source also reports hosted model prices that vary widely, so token prices alone do not settle whether on-premises is cheaper. The promised Kimi weights and their licence were still unresolved in the material reviewed.
Tests That Could Change the Case
Companies considering deployment will need to test candidate models against their own tasks, set acceptable quality and privacy requirements, and calculate total costs using realistic utilization and staffing assumptions. They also need to review each model’s licence rather than treating “open weights” as a uniform permission to use or modify a system.
The next developments to watch include whether Kimi K3.1 Pro’s weights become available as promised, what licence accompanies them, and whether independent benchmarks and enterprise deployments confirm the claimed savings. IBM’s statements on its hardware and llm-d project remain company claims; the source material does not provide an independent cost audit of those systems.
Key Questions
Did the benchmark confirm IBM’s 90% effectiveness claim?
No. The Artificial Analysis comparison read on October 26, 2026, put the leading downloadable model at 46.3 and the leading proprietary model at 57.6, or about 80% on that index. The score does not measure every task, and IBM’s 90% figure was not tied to a named benchmark in the source.
Does running AI on-premises always cost less?
No. The cited hardware comparison suggests purchase cost could equal around 114 days of cloud rental at full utilization, but about 15 months at 25% utilization. Both estimates exclude staff and other operating costs, so actual economics depend on usage and the full cost of running the system.
What does open-weight mean in this comparison?
The source uses the term for models whose weights can be downloaded. Availability does not make all models equivalent: their licences can set different conditions on commercial use, modification and deployment.
Was Kimi K3.1 Pro available to download?
Not as of the October 26, 2026, review in the source material. It scored 55.1 on the index, but Moonshot had only promised its weights “by Nov 4.” Its score should not be treated as a downloadable-model result until the weights are released.
What should an enterprise compare before buying servers?
It should compare model performance on its actual tasks, licence terms, expected hardware utilization, cloud rental rates and the full costs of staffing, power, cooling and networking. The source does not establish that one deployment approach is less expensive for every organization.
Source: ThorstenMeyerAI.com
