🔍 Read the full analysis: Why IBM’s CEO Says Companies Should Run AI Themselves on ThorstenMeyerAI.com
Get movie nights 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 told the company’s Oct. 21 earnings call that open-weight AI models can deliver about 90% of frontier-model performance at 10% to 25% of the token cost, and can be run on company infrastructure. An independent benchmark puts the leading downloadable model at 80% of the leading proprietary model; actual savings depend on workload, hardware use, staffing and licensing.
IBM CEO Arvind Krishna told the company’s Oct. 21 earnings call that open-weight AI models can be about 90% as effective as leading proprietary models at a fraction of their token cost, making them worth running on company-owned or private-cloud infrastructure. An independent model index puts the best downloadable model at 80% of the top proprietary model as of Oct. 26, while the cost case for buying and operating hardware depends on how heavily it is used.
Krishna said open-weight models can cost 10% to 25% of the token cost of frontier models. He suggested a large enterprise could begin with two or three racks of equipment. The comments came during the question-and-answer portion of IBM’s earnings call, rather than in its prepared remarks. IBM has a commercial interest in on-premises AI infrastructure, including its Spyre accelerator cards.
Krishna framed the approach as a fit for routine tasks such as summarizing email, reviewing contracts and answering customer calls, where a company may not need the strongest available model. He also cited data control: banks and traders, he said, want AI to run in a way that keeps underlying data and logs from being taken by others. Open-weight models can give users permission to run them on-premises or in private clouds, subject to each model’s licence.
The Artificial Analysis Intelligence Index, as read on Oct. 26, scored Claude Opus 5.5 at 57.6 and the leading downloadable model, Xiaomi’s MiMo-V2.6-Pro, at 46.3. That makes the open model about 80% of the proprietary leader’s score on the index. The figures support a narrowing relative gap compared with the end of 2024, when the top open model scored 51% of the proprietary leader, but they do not establish that open models perform equally well across every task.
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
Krishna’s pitch speaks to two enterprise concerns: AI operating costs and control over sensitive information. Running a model on company infrastructure may give an organization more control over where data and logs are processed. It can also reduce reliance on a cloud provider’s per-token pricing, particularly for workloads that generate a high and steady volume of requests.
But owning hardware does not automatically make AI cheaper. The source’s list-price comparison gives an eight-GPU H200 server at $172,643, while renting the same GPU class on AWS is listed at $63.30 an hour. At full utilization, the purchase price would equal roughly 114 days of rental; at 25% utilization, about 15 months. Those calculations exclude staffing, cooling and networking, and do not by themselves predict an organization’s total cost.
That distinction matters because enterprises must weigh the model’s price against the people and systems needed to run it. Krishna acknowledged that companies may lack kernel engineers and networking specialists. For organizations without those skills or enough sustained demand, a cloud service may remain simpler even if its list price is higher.
Popular right now
Philips 4400 Series Fully Automatic Espresso Machine with LatteGo Milk System
De'Longhi Dinamica Plus Fully Automatic Espresso Machine
TCL 85 Designer Series QD-Mini LED TV (Model 85A400U) with Google Smart TV
As an affiliate, we earn on qualifying purchases.
A Benchmark Gap That Has Narrowed
Krishna described the relative performance gap as closing: he recalled open models scoring roughly 20 or 30 against a leading model’s 70 in late 2024, then about 50 against 70 in late 2025. His “90% as good” comparison was an illustration, not a result tied to a named benchmark in the call. Artificial Analysis provides one separate reference point, using the same ten evaluations across models and identifying whether weights are downloadable.
On that index, the top open model’s share of the proprietary leader’s score rose from 51% at the end of 2024 to 74% at the end of 2025 and 80% in the Oct. 26, 2026 reading. The source’s comparison also shows that the point gap increased from 7.5 to 11.3 over that period, because proprietary models improved too. A possible higher score depends on Moonshot releasing Kimi K3.1 Pro’s weights: it scored 55.1, but the company had only promised to release them by Nov. 4. Until that happens, it is not a downloadable model.
Licensing is another part of the choice. The source says MiMo-V2.6-Pro and DeepSeek V4.1 Flash 1017 use MIT licences, while other models carry custom or commercial terms. A strong benchmark score alone does not establish that a model can be used for every business purpose. Krishna also said IBM’s llm-d project can run large models at five to 10 times lower cost; that is IBM’s claim, not an independently verified measurement in the source material.
Open Models Still Have Caveats
The independent index is one benchmark, not a universal measure of usefulness, and its scores do not show how a model performs on each company’s own tasks. Krishna’s 90% figure was not tied to a named test during the call. It is also not clear from the source material how the token-cost comparison was calculated across specific models, workloads or service providers.
The hardware comparison uses listed prices and simplified utilization assumptions. It does not include the full cost of staff, cooling, networking, maintenance or financing. Actual costs will vary with hardware availability, request volume and the proportion of time equipment sits idle. Model licences also differ, so companies need to check permitted uses rather than relying on the label “open-weight.”
Kimi K3.1 Pro’s weights had not shipped as of the Oct. 26 benchmark reading, and no licence had been published in the source account. Nvidia had not commented on the reported GeForce firmware limitation exploit described elsewhere in the source material; that enthusiast issue does not establish anything about enterprise AI economics.
Watch Model Releases and Workloads
The next stated milestone for the benchmark comparison is Nov. 4, the date by which Moonshot had promised Kimi K3.1 Pro’s weights. If released, its score could change the comparison, but its licence and computing requirements will also matter. The source says the model has 2.8 trillion parameters, making practical deployment a data-centre-scale undertaking.
For companies weighing an in-house deployment, the immediate next step is to compare models on their own tasks and costs, check the licence, and estimate expected hardware utilization alongside staffing and infrastructure needs. IBM’s comments establish the company’s case for on-premises AI; they do not settle whether owning capacity is cheaper or better suited to any particular organization.
Key Questions
What did IBM CEO Arvind Krishna say about open-weight AI?
Krishna said open-weight models can be 90% as effective as frontier models at about 10% to 25% of their token cost. The 90% figure was an illustrative claim on IBM’s earnings call, not a result tied there to a named benchmark.
How close are open models to proprietary models on the cited index?
As of Oct. 26, 2026, Artificial Analysis scored the leading downloadable model at 46.3 and the leading proprietary model at 57.6. The open model’s score was about 80% of the proprietary leader’s.
Does running AI on company hardware guarantee lower costs?
No. The economics depend on utilization, staffing and infrastructure, among other factors. The cited server-versus-rental comparison excludes costs such as cooling, networking and technical staff.
Why might a company run a model on-premises?
Krishna pointed to data control and sovereignty, as well as potential savings for high-volume use. Open-weight models may be run on company infrastructure or private clouds, subject to their licences.
What should companies check before choosing an open-weight model?
They should test the model on their own workloads, estimate the full cost of operating it, and review its licence and deployment requirements. A benchmark score alone does not determine whether a model is suitable for a specific business use.
Source: ThorstenMeyerAI.com
