🔍 Read the full analysis: The Trade-Off Behind IBM’s On-Prem AI Bet on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
IBM CEO Arvind Krishna argued on the company’s October 21 earnings call that open-weight AI models can offer about 90% of frontier-model performance at 10% to 25% of the token cost. A benchmark comparison available October 26 puts the best downloadable model at 80% of the leading proprietary model, while the cost case for running models on-prem depends on hardware utilization, staffing and licensing.
IBM CEO Arvind Krishna used the company’s October 21 earnings call to argue that enterprises can run open-weight AI models on their own infrastructure at a fraction of frontier-model token costs, with performance he described as about 90% as effective. An independent benchmark comparison available October 26 puts the leading downloadable model at 80% of the top proprietary model, leaving a meaningful gap and making the economics of an on-prem deployment dependent on workload, hardware use and staff capacity.
In the call’s question-and-answer session, Krishna said open-weight models cost “25% to 10% of the token cost” of frontier models while being “90% as effective.” He cited routine enterprise tasks such as summarizing email, reviewing contracts and answering customer calls as work that may not require the most capable model. He also said banks and trading firms may prefer to keep data and logs within infrastructure they control.
Krishna’s numbers are his characterization, not a comparison tied to a named benchmark. Artificial Analysis’s Intelligence Index, checked October 26, rated MiMo-V2.6-Pro at 46.3, the highest-scoring model with downloadable weights in its comparison. Proprietary model Claude Opus 5.5 scored 57.6, putting the open model at about 80% of that score. The index shows the relative score moving from 51% at the end of 2024 to 74% at the end of 2025 and 80% in October 2026; these figures describe that index, not every task or model.
A possible exception is Kimi K3.1 Pro, which scored 55.1, or 96% of the Claude score, in the supplied comparison. But Moonshot had only promised to release its weights “by Nov 4”; until they are available, it cannot be counted as a downloadable open-weight model. The source material also says no licence had been published and notes that the model’s reported 2.8-trillion-parameter scale could make deployment a substantial data-centre undertaking.
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 of Owning AI Capacity
IBM’s pitch is not simply that open models cost less per token. It is that enterprises may trade a cloud provider’s metered service for capital spending and operational responsibility. A company buying servers has to keep them productively occupied, manage networking and cooling, and recruit people able to run the models. Those costs can alter the comparison even when the model’s token price is low.
The source’s illustrative hardware figures show why utilization matters. An eight-GPU H200 server lists from $172,643, while renting the same GPU class on AWS is listed at $63.30 an hour. At continuous full use, the purchase price alone would equal about 114 days of rental; at 25% utilization, the comparison stretches to about 15 months. Those are simplified list-price calculations, not a full total-cost analysis: they exclude staffing, cooling, networking and other operating expenses.
For buyers, the practical choice is therefore workload-specific. A steady, high-volume use case may make owned capacity attractive, while unpredictable demand or a shortage of infrastructure expertise may favor rented services. Data control and licensing requirements can matter as much as benchmark rank.
Products Worth Considering
Sony WH-1000XM4C Wireless Noise-Cancelling Headphones, Black
DJI Osmo Mobile 8 Standard Combo Phone Gimbal Stabilizer
Samsung 27 Odyssey OLED G5 (G50SF) QHD QD-OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
What Krishna Said on the Call
Krishna described the open-versus-closed model gap as narrowing: in his account, a top model scoring 70 against an open model at 20 or 30 in late 2024 became 70 against 50 in late 2025. He then framed the decision as whether to accept a model “90% as good as the best” for a tenth of the cost, rather than pay ten times more. These score examples were illustrations in the call, not results from a benchmark he identified.
He also pointed to a capability gap inside some companies. “Where are my kernel engineers?” he asked, referring to the specialists and network expertise needed to run infrastructure efficiently. Krishna described frontier cloud services as the easy option, while arguing that open weights give customers permission and intellectual-property rights to run models on-premises or in private clouds. The exact rights depend on each model’s licence.
IBM has a commercial interest in the shift it is advocating because it sells on-premises infrastructure, including its Spyre cards. The source report also says IBM claimed its llm-d project can run large models at five to ten times lower cost; that figure is an IBM claim, and the report did not provide an independent measurement. The same report, citing The Register, says IBM mainframe revenue fell 20% in the quarter, 18 months into the z17 cycle, a separate business factor rather than proof of why IBM is promoting on-prem AI.
The Gaps in the Cost Case
The benchmark comparison does not settle how well open and proprietary models perform on a particular company’s tasks. Artificial Analysis provides a shared index, but the scores cited here do not establish that an open model will reach 80% of a proprietary model’s quality in every workflow. Krishna’s 90% figure remains a company executive’s broad claim, not a result independently verified in the supplied material.
Nor is the token-price comparison a complete cost comparison. A hosted model’s list price does not include the full expense of owning and operating a server, and a purchase-versus-rental calculation changes with utilization. The supplied material gives no staffing, cooling or networking estimates, and it does not show how frequently a typical enterprise would keep its hardware busy.
Model access and commercial rights also vary. Some models in the comparison use MIT licences, while others have custom or commercial terms. Kimi K3.1 Pro’s promised weight release and licence were unresolved as of the report’s October 26 benchmark check. IBM’s llm-d cost claim, the eventual performance of that system in customer deployments, and Nvidia’s response to the reported GeForce modification are also not independently established in the material.
Deployment Decisions and Model Releases
Enterprises weighing the approach will need to compare their own workloads, expected utilization, infrastructure budgets and data-control requirements against cloud pricing. They will also need to review each model’s licence and account for the engineering work Krishna said can be a bottleneck. The figures cited here are snapshots of prices and benchmark standings, not a forecast of a buyer’s eventual savings.
The next relevant model milestone in the supplied material is Moonshot’s promised release of Kimi K3.1 Pro weights by November 4, 2026. Whether the weights arrive on that schedule, under what licence, and with what practical hardware requirements remains to be seen. Further evidence would also be needed to test IBM’s lower-cost claims against independently measured deployments.
Key Questions
What did IBM’s CEO claim about open-weight AI models?
On the October 21 earnings call, Arvind Krishna said open-weight models cost 10% to 25% as much per token as frontier models while being about 90% as effective. The supplied material does not tie that estimate to a named benchmark.
How close is the best downloadable model to the leading proprietary model?
In the Artificial Analysis comparison read October 26, 2026, MiMo-V2.6-Pro scored 46.3 against Claude Opus 5.5’s 57.6, or about 80% of its score on that index. That ratio does not establish performance on every enterprise task.
Does running AI on-premises always cost less?
No. The economics depend on factors including hardware utilization, staffing, cooling and networking. The supplied server-versus-rental figures are simplified list-price comparisons and do not include those operating expenses.
Is Kimi K3.1 Pro an open-weight model now?
Not according to the source material’s October 26 status. Moonshot had promised to release the weights by November 4, but they had not yet shipped, and the report said a licence had not been published.
Why might a company choose on-prem AI?
Krishna pointed to data control and sovereignty, as well as potential token-cost savings. On-premises deployment also gives the company responsibility for the infrastructure and the specialist staff needed to operate it.
Source: ThorstenMeyerAI.com
