🔍 Read the full analysis: What It Takes To Run AI On-Prem, According To IBM’s CEO on ThorstenMeyerAI.com
Get the little things that make your day 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 on the company’s Oct. 21 earnings call that open-weight models can approach leading proprietary systems at a fraction of their token cost, and that large enterprises could begin on-prem deployments with two or three racks. An independent benchmark puts the strongest downloadable model at 80% of the top proprietary model’s score as of Oct. 26; the economics of buying hardware depend heavily on utilization, staffing and other costs.
IBM CEO Arvind Krishna said on the company’s Oct. 21 third-quarter earnings call that open-weight AI models can deliver about 90% of the performance of leading proprietary models for 10% to 25% of their token cost, making on-premises deployment worth considering for large companies. An independent benchmark reviewed Oct. 26 puts the strongest model with downloadable weights at 80% of the top proprietary model’s score, while the costs and staffing needs of running AI in company-owned infrastructure remain dependent on how heavily it is used.
Krishna said enterprises could begin with two or three racks and run open-weight models on their own infrastructure or in private clouds. He cited routine work such as summarizing email, examining contracts and answering customer calls as tasks that may not need the most capable available model. He also pointed to data control: banks and trading firms, he said, want to keep underlying information and logs within their own environment.
The comparison is Krishna’s characterization, not a result from a named benchmark. On Artificial Analysis’s Intelligence Index, the leading downloadable model listed on Oct. 26, Xiaomi’s MiMo-V2.6-Pro, scored 46.3, compared with 57.6 for proprietary Claude Opus 5.5. That puts the open model at about 80% of the proprietary leader’s score. The same index showed a 51% share at the end of 2024, indicating that the relative gap has narrowed, although the score difference rose from 7.5 points then to 11.3 points in October 2026.
A potentially closer comparison depends on a model that is not yet available with downloadable weights. Moonshot’s Kimi K3.1 Pro scored 55.1 on the index, about 96% of Claude Opus 5.5, but Moonshot had only promised to release its weights by Nov. 4. Until they are available, it remains a closed model for the purposes of comparing downloadable systems.
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 Usage
Krishna’s pitch puts a practical question in front of companies: whether recurring cloud charges justify the upfront cost of hardware and the work of operating it. The supplied price comparison lists an eight-GPU H200 server from $172,643 and AWS rental of the same GPU class at $63.30 an hour. At continuous full use, that comparison suggests a purchase could match the rental cost in about 114 days. At 25% use, the estimate stretches to about 15 months, before adding staff, cooling and networking.
That range makes utilization a central part of the business case. A company with steady, high-volume workloads may use owned hardware more intensively than one with intermittent demand. But lower hardware rental or model-token prices alone do not establish that an on-prem deployment will save money: installation, maintenance, power, network capacity and specialist staff also affect the total.
There is a control argument as well as a price argument. Krishna said open weights give companies permission and intellectual-property rights to run models on their own systems or private clouds, which may appeal to organizations seeking to limit where data and logs are held. The arrangement does not remove the need to review each model’s licence or decide how the system will be secured and maintained.
Products Worth Considering
Ring Alarm 14-Piece Wireless Security System Kit
Sonos Era 100 Wireless Smart Speaker, Black
Samsung 85 M80H Series Mini LED 4K Smart TV (2026)
As an affiliate, we earn on qualifying purchases.
IBM’s Case for Open Models
Krishna’s remarks came during an earnings call, where the CEO discussed a market that could create demand for both AI software and infrastructure. IBM sells on-premises systems, so it has a commercial interest in enterprises choosing to run models on their own equipment. The source material also says IBM expects on-prem demand to support sales of its Spyre cards. That interest does not settle whether the cost and performance comparisons are sound, but it is relevant when weighing the company’s argument.
He described model performance as having closed in stages: in late 2024, he said, a top model might score 70 while an open one scored 20 or 30; by late 2025, he put the comparison at 70 to 50. Artificial Analysis’s historical figures show the leading downloadable model at 51% of the proprietary leader’s score at the end of 2024, rising to 74% at the end of 2025 and 80% in October 2026. These are index results, not a direct validation of Krishna’s illustrative figures.
Model terms vary alongside scores. The source identifies MiMo-V2.6-Pro and DeepSeek V4.1 Flash 1017 as MIT-licensed, while other models use their own commercial or community licences. A model’s presence on a leaderboard does not by itself establish that a company can use it on the terms it needs; operators must check the applicable licence and deployment conditions.
Performance and Payback Limits
The benchmark figures do not establish how the models compare on every enterprise task, or whether the same model settings, quality requirements and operating conditions would apply to a particular company. Krishna’s 90% figure is not tied to a named evaluation in the supplied material; the independent index records 80% for the top downloadable model on its Oct. 26 snapshot. The 96% comparison depends on Kimi’s weights being released as promised, and its licence had not been published in the source material.
The hardware payback estimates are based on listed prices and stated utilization assumptions. They exclude staffing, cooling and networking, and do not establish the costs a specific company would face. IBM’s separate statement that its llm-d project can reduce the cost of running large models by five to ten times is also a company claim; the supplied material says it could not independently verify that measurement.
It remains unclear how many enterprises can maintain high enough utilization to make owned infrastructure less expensive than cloud services, or how large a pool of trained staff they will need. IBM’s commercial interest in on-prem systems is also relevant to its forecast. The earnings-call remarks set out the company’s case, not a guarantee of savings for individual customers.
Model Releases and Deployment Decisions
The next near-term test of the benchmark comparison is whether Moonshot releases Kimi K3.1 Pro’s weights by Nov. 4, as promised, and under what licence. The model’s score is already listed, but access to its weights and the conditions for use will determine whether it can be evaluated as a downloadable option for enterprise deployment.
Companies considering on-prem AI will need to compare their own workloads, expected utilization and operating costs with cloud pricing. That assessment includes more than the accelerator purchase: engineering staff, networking, cooling, security and licence terms all affect whether the setup is workable. IBM has made the case for the approach; the supplied material does not report a new customer deployment or independently verified savings from a named enterprise.
Key Questions
What did IBM CEO Arvind Krishna say about open-weight models?
On IBM’s Oct. 21 earnings call, Krishna said open-weight models could cost 10% to 25% as much in tokens as frontier models while being about 90% as effective. The 90% figure was his comparison, not a result from a named benchmark.
How close are downloadable models to the proprietary leader?
On Artificial Analysis’s Intelligence Index, MiMo-V2.6-Pro scored 46.3 against Claude Opus 5.5’s 57.6 on Oct. 26, or about 80% of the leader’s score. Kimi K3.1 Pro scored 55.1, but its weights had not yet been released.
Does buying AI hardware automatically save money?
No. The supplied comparison estimates a payback of about 114 days at full use for an eight-GPU H200 server versus AWS rental, but about 15 months at 25% use. Both estimates exclude staff, cooling and networking.
Why might a company run AI on-prem?
Krishna cited costs and the ability to keep data and logs under a company’s control. On-prem deployment also requires infrastructure and specialist staff, and the applicable model licence must allow the intended use.
Source: ThorstenMeyerAI.com
