🔍 Read the full analysis: IBM’s CEO Sees A Volume Bet In On-Prem AI on ThorstenMeyerAI.com
Get hardware and tech essentials 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 argued on the company’s Oct. 21 earnings call that open-weight AI models can deliver about 90% of frontier-model performance at a fraction of the token cost, and said large enterprises could begin with two or three racks. An independent model index reviewed Oct. 26 puts the best available open-weight model at 80% of its top proprietary competitor; the cost case for buying hardware also depends on utilization, staffing and infrastructure.
IBM CEO Arvind Krishna told investors on the company’s Oct. 21 earnings call that open-weight AI models can cost 10% to 25% as much per token as frontier models while being “90% as effective,” making on-premises deployment worth considering for some enterprises. An independent benchmark reviewed Oct. 26 put the strongest downloadable model at 80% of the leading proprietary model, while the economics of owning servers depend heavily on how fully they are used.
Krishna said a large enterprise could start an on-premises deployment with “two or three racks”. He pointed to routine tasks such as summarizing email, reviewing contracts, answering customer calls and composing messages as work that may not require the most capable available model. He also argued that keeping models and data within an organization’s infrastructure can help address data control and sovereignty concerns, particularly for banks and trading firms.
The source report compared those statements with the Artificial Analysis Intelligence Index, which applies the same ten evaluations to models and identifies whether their weights can be downloaded. On the index version read Oct. 26, proprietary Claude Opus 5.5 scored 57.6, while the leading model with available weights, Xiaomi’s MiMo-V2.6-Pro, scored 46.3. That is about 80% of the proprietary score. The report says the comparable share was 51% at the end of 2024, though the absolute score gap grew from 7.5 points to 11.3 as both categories improved.
Krishna’s cost comparison is not a guarantee that self-hosting will be cheaper overall. The source report lists an eight-GPU H200 server from $172,643 and AWS rental of the same GPU class at $63.30 per hour. Its calculation estimates a purchase would match rental costs after about 114 days at full utilization, or about 15 months at 25% utilization, before accounting for staff, cooling and networking. Those are illustrative list-price comparisons, not a complete operating-cost analysis.
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
When On-Prem AI Can Pay Off
The argument matters because enterprises are weighing recurring cloud bills against the upfront cost and operational work of running AI themselves. If a model is capable enough for a high-volume, well-defined task, using less expensive open weights on owned infrastructure could reduce token charges and give an organization greater control over its data. That is the business case Krishna is making, and it aligns with IBM’s interest in selling enterprise hardware and services.
But a lower token price is only one part of the calculation. Server purchases require capital, and low utilization can stretch the time needed to offset rental costs. Krishna himself identified a shortage of kernel and networking engineers as a constraint for many organizations, calling frontier cloud services the easier option. The practical decision depends on workload volume, model quality, deployment skills, licence terms and the cost of operating the equipment—not just a comparison of per-token rates.
The benchmark comparison also puts a boundary around the “90%” claim. The strongest currently downloadable model in the cited index reached 80% of the top proprietary score. A model that has announced a future weights release could narrow the difference, but it does not count as open-weight until those weights are available. Scores from one index are useful evidence, not proof that models perform equally on a particular company’s tasks.
Products Worth Considering
Apple MacBook Pro 16-inch Laptop with M5 Pro, 18-Core CPU, 20-Core GPU, 48GB Memory, 2TB SSD, Space Black
Apple MacBook Pro 14.2-inch with M5 Max, 18-Core CPU, 20-Core GPU, 48GB Memory, 2TB SSD, Space Black
Logitech MX Keys S Wireless Keyboard, Graphite
As an affiliate, we earn on qualifying purchases.
IBM’s Pitch and the Model Gap
Krishna described a rapid narrowing of the performance difference between proprietary and open-weight systems. In his account, broad benchmark scores were once around 70 for a leading model and 20 or 30 for an open model; by late 2025, he said, the comparison was closer to 70 and 50. He then framed the purchasing choice as paying a fraction of the cost for a model he characterized as nearly as capable. The figures were examples in his earnings-call answer, rather than results tied there to a named benchmark.
The independent index figures cited in the source show improvement in relative performance, but also show that the proprietary frontier moved. At the end of 2024, the listed leading proprietary model scored 15.2 and the leading open-weight model 7.7. On the Oct. 26, 2026 reading, the figures were 57.6 and 46.3. The report notes that model scores and rankings can vary with the index version and model configuration.
Open weights do not automatically mean unrestricted commercial use. The source report says MiMo-V2.6-Pro and DeepSeek V4.1 Flash 1017 use MIT licences, while other models have custom or commercial terms. It also reports that Moonshot had promised Kimi K3.1 Pro weights “by Nov. 4”; at the time of the Oct. 26 comparison, those weights had not shipped. Licence conditions and hardware requirements can shape whether a model is usable, even when its benchmark score is high.
IBM also promoted its llm-d project during prepared remarks, claiming it can run large models at “5 to 10 times lower cost.” The source report says it could not independently verify that figure. IBM has a commercial stake in on-premises AI: the report says Krishna expects demand to support sales of the company’s Spyre cards. It also cites The Register’s report that IBM mainframe revenue fell 20% in the quarter, 18 months into the z17 cycle. These factors are relevant when weighing the company’s forecast against independent comparisons.
Costs and Capability Still Vary
The index comparison does not establish how well each model performs on an individual company’s tasks, and Krishna’s “90%” figure was not linked to a specific benchmark in the call. The available-model comparison also depends on which systems are counted as open and on their licences. Kimi K3.1 Pro’s reported score would put it near the proprietary leader, but its weights had not been released by the Oct. 26 review.
The server-versus-rental payback estimates omit staffing, cooling and networking, and assume particular utilization levels and listed prices. Actual costs can differ by contract, location and workload. The source material also does not provide independently verified results for IBM’s claimed llm-d cost reduction or establish how many enterprises are already deploying systems at the scale Krishna described.
Watch for Weights and Deployments
The next concrete test for the model comparison is whether Moonshot makes Kimi K3.1 Pro’s weights available as promised, and under what licence. Any release would allow analysts and enterprises to examine the model’s access terms and deployment requirements; the reported score alone does not settle either question.
For the on-premises business case, the relevant evidence will come from deployment costs and performance in actual enterprise workloads, including how much the equipment is used and what technical staff are needed. IBM’s earnings-call statements set out its case for the approach, but the available source material does not provide a verified, full-cost comparison across customer deployments.
Key Questions
What did IBM’s CEO say about open-weight AI?
Arvind Krishna said on IBM’s Oct. 21 earnings call that open-weight models can cost 10% to 25% as much per token as frontier models while being “90% as effective.” That is his characterization, not a result tied to a named benchmark in the call.
How close are the best available open-weight models to proprietary models?
On the Artificial Analysis index reviewed Oct. 26, 2026, MiMo-V2.6-Pro scored 46.3, compared with 57.6 for Claude Opus 5.5, or about 80% of the proprietary score. Results depend on the benchmark and model configuration.
Does running AI on-premises always cost less?
No. The economics depend on equipment prices, utilization, staffing, cooling and networking. The source report’s server comparison estimated a much longer purchase payback at 25% utilization than at full load, and excluded several operating costs.
Why might a company choose an open-weight model?
Potential reasons include lower serving costs for suitable workloads and more control over where data and logs are handled. Krishna cited data sovereignty as a concern for financial firms. Licence terms and the organization’s ability to operate the system also matter.
What is still unknown about IBM’s cost claims?
The source material does not independently verify IBM’s claim that its llm-d project can run large models at 5 to 10 times lower cost. It also does not provide a complete, independently checked cost comparison for enterprise on-premises deployments.
Source: ThorstenMeyerAI.com
