🔍 Read the full analysis: IBM’s CEO Says Run AI Yourself. Open-Weight Models Are Close, But On-Prem Is A Volume Bet on ThorstenMeyerAI.com
Get the latest gadgets 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 models can deliver about 90% of leading-model performance at a fraction of the token cost, supporting on-premises and private-cloud deployments. An Artificial Analysis comparison puts the leading downloadable model at 80% of the best proprietary model as of Oct. 26; a higher-scoring model’s weights have not yet shipped. On-prem savings also depend on hardware use, staffing and other costs.
IBM CEO Arvind Krishna said on Oct. 21 that open-weight AI models can cost 10% to 25% as much per token as leading proprietary models while being “90% as effective,” making them a potential fit for enterprise systems run on-premises or in private clouds. A comparison from Artificial Analysis, read Oct. 26, puts the best currently downloadable model at 80% of the top proprietary model’s score, underscoring both the progress and the remaining performance gap behind IBM’s pitch.
Krishna made the comparison during the question-and-answer portion of IBM’s Q3 earnings call, describing model performance as having narrowed over time. He said that in late 2024 a top model might score 70 on a broad benchmark while an open model scored 20 or 30; by late 2025, he said, the comparison was 70 to 50. His cost-and-performance summary was: “Would you like something that’s 10% of the cost if it’s 90% as good as the best, or do you want to pay that 10 times higher cost?” These figures were Krishna’s illustrations, not results from a benchmark he identified on the call.
Artificial Analysis scores models across ten evaluations and indicates whether their weights are downloadable. Its Oct. 26 comparison lists Claude Opus 5.5 at 57.6 as the leading proprietary model and Xiaomi’s MiMo-V2.6-Pro at 46.3 as the highest-scoring model with downloadable weights. That makes the latter about 80% of the proprietary leader’s score on this index. The open-to-proprietary share was 51% at the end of 2024 and 74% at the end of 2025, according to the same comparison.
A model called Kimi K3.1 Pro scored 55.1, or 96% of Opus 5.5’s score, but Moonshot had only promised to release its weights “by Nov 4.” Until those weights are available, it is not an open-weight option for enterprises to download and run. Krishna also said open models could support work such as summarizing email, reviewing contracts and answering customer calls, and described data sovereignty as a reason banks and traders might prefer their own infrastructure.
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 Running AI In-House
Krishna’s argument matters because the decision is not only which model performs best: it is also whether an organization can meet its needs at a lower total cost and retain greater control over data. A model with downloadable weights can be deployed on company infrastructure or a private cloud, depending on its licence and the organization’s technical capacity. That may appeal to firms handling sensitive information or seeking more control over where data and logs are stored.
But a lower token price does not by itself prove that owning hardware is cheaper. The source’s example lists an eight-GPU H200 server from $172,643 and AWS rental of the same GPU class at $63.30 an hour. On those list prices, the purchase cost is equivalent to about 114 days of continuous rental at full use, or about 15 months at 25% use. Those estimates exclude staff, cooling and networking, and do not account for other procurement or operating costs. The economics hinge on keeping expensive hardware busy.
IBM also has a commercial interest in greater on-premises adoption: it sells infrastructure, and Krishna said demand could support sales of IBM Spyre cards. That does not make his claims false, but it means the performance and cost case should be checked against independent comparisons and a company’s own workload rather than treated as neutral guidance.
Products Worth Considering
Samsung 27-inch Odyssey OLED G6 QD-OLED QHD Gaming Monitor
Sony WH-1000XM4C Wireless Noise-Cancelling Headphones, Black
Bose QuietComfort Headphones (2nd Gen), Wireless Noise-Cancelling Over-Ear Headphones, Eucalyptus Green
As an affiliate, we earn on qualifying purchases.
What the Earnings Call Established
Krishna’s remarks set out both a case for running models locally and a warning about the skills required. He said open weights give companies permission and intellectual-property rights to run models on-premises or in private clouds, subject to the particular model’s licence. He also described “sovereign” operation as attractive to banks and traders concerned about access to underlying data and logs.
At the same time, Krishna identified shortages of “kernel engineers” and people who manage underlying networks as a bottleneck outside the most technically sophisticated enterprises. He characterized frontier cloud services as more expensive but “the easy answer.” That qualification points to a practical constraint: buying servers does not automatically provide the expertise needed to deploy and operate large models reliably.
IBM’s prepared remarks also said its llm-d project helps run large models at “5 to 10 times lower cost.” That is an IBM claim; the source material does not provide an independent measurement supporting it. Likewise, model scores are not a complete guide to a deployment decision. Model licences differ, and some may impose conditions. Companies need to check the licence as well as benchmark results before choosing a model.
Open Weights and Real-World Costs
It remains unclear whether Kimi K3.1 Pro’s weights will arrive by Moonshot’s promised Nov. 4 date, what licence will govern them, or whether the model’s reported score will translate to enterprise workloads. The source says no licence had been published. Its reported 2.8-trillion-parameter scale also raises practical infrastructure questions that a leaderboard score cannot answer.
The Artificial Analysis figures measure performance on ten evaluations, not the full cost or usefulness of a model in a particular company’s work. The comparison also does not establish Krishna’s “90% as effective” claim: the highest-scoring model with downloadable weights was at 80% of the proprietary leader on Oct. 26. The result could differ with another benchmark, model version or task.
Nor do the server and rental figures settle the buy-versus-rent decision. They are list-price comparisons and omit staffing, cooling, networking and other operating costs. Actual utilization, hardware availability, service requirements and model licensing remain key variables. IBM’s lower-cost claim for llm-d has not been independently verified in the supplied material.
Watch for Model Releases and Deployment Data
The next near-term milestone cited in the source is Moonshot’s promised release of Kimi K3.1 Pro’s weights by Nov. 4, 2026. Its availability, licence and deployability will determine whether it becomes a genuine open-weight alternative, rather than a high-scoring model that remains inaccessible for self-hosting.
For IBM’s broader argument, the relevant test will be evidence from real deployments: whether companies can sustain high hardware utilization, staff the required engineering work and achieve savings after operating expenses. Future independent comparisons may also show whether open-weight models continue to narrow the performance gap. Until then, Krishna’s remarks are a case for evaluating on-prem AI, not proof that it is the cheaper choice for every enterprise.
Key Questions
What did IBM CEO Arvind Krishna say about open-weight AI?
On IBM’s Oct. 21, 2026, earnings call, Krishna said open-weight models cost 10% to 25% as much per token as frontier models while being “90% as effective.” He presented the figures as part of an argument for running models on-premises or in private clouds.
Do independent scores show that open models are already 90% as capable?
Not in the comparison supplied here. Artificial Analysis listed the leading downloadable model at 46.3 and the leading proprietary model at 57.6 on Oct. 26, 2026, a ratio of about 80%. Kimi K3.1 Pro scored 55.1, but its weights had not yet been released.
Does running AI on-premises automatically cost less?
No. The example’s server purchase price equates to about 114 days of AWS rental at continuous full use, or about 15 months at 25% use. Those estimates exclude staff, cooling and networking, so actual costs depend on utilization and operating needs.
Why might a company choose an open-weight model?
A company may want more control over deployment and data handling, or to avoid paying proprietary-model token prices for tasks that do not require the leading model. It must still check the model’s licence, performance and infrastructure requirements before deployment.
Source: ThorstenMeyerAI.com
