🔍 Read the full analysis: Why IBM’s CEO Wants Companies To Run AI Themselves 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 says companies can use open-weight AI models on their own infrastructure at a fraction of the cost of leading proprietary models, with added control over data. Independent benchmark figures show the best downloadable model at about 80% of the top proprietary score on October 26, 2026—not 90%—while the economics of buying hardware depend on how heavily it is used.
IBM CEO Arvind Krishna said on the company’s October 21 earnings call that open-weight AI models can deliver about 90% of a leading model’s effectiveness at 10% to 25% of its token cost, making them a candidate for companies running AI on their own systems. Independent benchmark data available October 26 put the best downloadable model at 80% of the top proprietary score, while the costs of buying and operating hardware vary with usage.
Krishna said a large enterprise could begin with two or three racks and use open-weight models for routine work such as summarizing email, reviewing contracts, answering customer calls and drafting messages. He argued that companies can run these models on premises or in private clouds, retaining more control over their data and logs. IBM sells hardware and services for enterprise computing, giving the company a commercial interest in greater on-premises AI use.
The comparison cited by the source uses Artificial Analysis’s Intelligence Index, which applies the same set of ten evaluations across models. As of October 26, Xiaomi’s downloadable MiMo-V2.6-Pro scored 46.3, compared with 57.6 for proprietary Claude Opus 5.5. That makes the open model about 80% of the proprietary leader’s score on this index. At the end of 2024, the corresponding figures were 7.7 for Qwen2.5 72B and 15.2 for o1, or 51%.
A possible exception is Kimi K3.1 Pro, which scored 55.1 on the index, about 96% of Claude Opus 5.5’s score. But Moonshot had only promised to release its weights “by Nov 4.” Until downloadable weights are available, the model does not qualify as open-weight under the comparison. The source also says no licence had been published, leaving the terms of use unresolved.
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 Trade-Off Behind On-Prem AI
The argument affects more than the price of individual model calls. Running a model in a company’s own data centre can give it greater control over data and logs, a concern Krishna specifically linked to banks and traders. Open-weight models may also give businesses more flexibility about where and how they deploy a system, depending on the model’s licence.
But lower token prices do not by themselves establish that owning hardware is cheaper. The source compares an 8-GPU H200 server listed from $172,643 with AWS rental of the same GPU class at $63.30 an hour. Its estimate says the purchase could match rental costs after about 114 days at full use, or about 15 months at 25% use, before staffing, cooling and networking. Those assumptions matter: utilization, operating costs and technical capacity can change the calculation.
Krishna acknowledged that companies may lack the staff to operate this infrastructure. He pointed to shortages of kernel engineers and networking expertise, describing the frontier cloud as the easier option despite higher cost. That makes the choice a question of operational readiness and control as well as model performance and price.
Products Worth Considering
FLEXISPOT 36-Inch Electric Standing Desk Converter, Black
Dell UltraSharp 27 Thunderbolt Hub Monitor (U2726DER)
Apple iPad Air 13-inch (M4), 128GB, Wi-Fi 7, Space Gray
As an affiliate, we earn on qualifying purchases.
How the Open-Model Gap Changed
Krishna described model performance as closing in stages: in late 2024, he said, a top model might score 70 while an open model scored 20 or 30; by late 2025, he characterized the comparison as 70 to 50. These were illustrations in his earnings-call answer, not results tied to a named benchmark. The independent index offers a separate way to track the relative scores, and its figures show the open model’s share of the leader rising from 51% at the end of 2024 to 80% in October 2026.
The ratio improved, but the absolute score gap did not shrink in the cited index: it was 7.5 points at the end of 2024 and 11.3 points in October 2026. Both proprietary and downloadable models improved over that period, so the relative gain does not mean open models caught up in every measure or task.
Price comparisons also depend on the product and route to access. The source says hosted DeepSeek V4.1 Flash 1017 costs 3% to 6% of Claude Opus 5.5’s list price, while OpenAI’s closed GPT-6 Luna is listed at $0.10 per million input tokens and $0.50 per million output tokens. Those are different providers’ listed prices, not a universal cost for open versus closed models. IBM also 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 supplied material.
Performance, Licences and Costs
The benchmark comparison is a snapshot from October 26, 2026, using Artificial Analysis’s index; it does not establish how every model performs on every company’s tasks. The source does not provide a named benchmark for Krishna’s “90% as effective” figure, so that claim cannot be directly matched to the index result. Kimi K3.1 Pro’s promised weight release and licensing terms also remain unsettled in the supplied material.
The cost examples use list prices and simplified utilization assumptions. The server estimate excludes staffing, cooling and networking, and the source does not specify how those expenses would affect the payback period in a particular enterprise. Model licences also differ: some are permissive, while others impose their own conditions. A high benchmark score or low token price alone does not establish that a model is suitable for a given use or can be used under a company’s intended terms.
IBM’s commercial position is relevant to its recommendation. The company sells on-premises computing products and Krishna said the strategy could support demand for IBM’s Spyre cards. The supplied material also reports that IBM’s mainframe revenue fell 20% in the quarter, but does not establish that this result caused Krishna’s comments or that the AI strategy will change that business.
The Next Tests for Enterprise Adoption
The immediate marker for the open-model comparison is whether Moonshot releases Kimi K3.1 Pro’s weights by its promised November 4 date, and what licence accompanies them. If the weights arrive on those terms, its index score would put it close to Claude Opus 5.5 on that particular measure. The release and licence were not confirmed in the source material.
For companies weighing deployment, the next step is to test candidate models against their own workloads and calculate total operating costs at realistic utilization—not rely solely on a general benchmark or token-price comparison. They will also need to examine licensing and whether their teams can manage the hardware, networking and model operations. The source gives no forecast for how many companies will adopt on-prem AI or how quickly IBM’s hardware strategy may affect sales.
Key Questions
What did IBM CEO Arvind Krishna say about open-weight AI?
On IBM’s October 21, 2026, earnings call, Krishna said open-weight models can cost 10% to 25% as much in tokens as frontier models while being “90% as effective.” That performance figure was his claim, not a result tied to a named benchmark in the supplied material.
Do independent benchmark figures support the 90% claim?
Not for the best downloadable model listed on October 26. Artificial Analysis scored MiMo-V2.6-Pro at 46.3 and Claude Opus 5.5 at 57.6, putting the open model at about 80% of the proprietary leader on that index. Kimi K3.1 Pro scored closer, but its weights had not yet been released according to the source.
Why would a company run AI on its own infrastructure?
Krishna cited data control, privacy and sovereignty, alongside potentially lower costs. On-premises or private-cloud deployment can give a company more control over where data and logs are handled, but operating the systems requires appropriate hardware and technical staff.
Is buying AI hardware always cheaper than renting it?
No. The cited comparison estimates that an $172,643 H200 server could match AWS rental costs after about 114 days at full utilization, or about 15 months at 25% utilization. Those estimates exclude staffing, cooling and networking, so they are not a complete cost calculation for any specific company.
What is still unknown about the leading open-model comparison?
The source says Moonshot had promised Kimi K3.1 Pro’s weights by November 4, but they were not yet available in the cited material and no licence had been published. Whether the release happens on time and what use its licence permits remain unresolved.
Source: ThorstenMeyerAI.com
