🔍 Read the full analysis: IBM’s CEO Shares His Outlook On Self-Run AI 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 said open-weight AI models can deliver about 90% of frontier-model effectiveness at 10% to 25% of the token cost, and that some enterprises can start with a few racks. An Artificial Analysis comparison read Oct. 26 put the leading model with downloadable weights at 80% of the top proprietary model; it could reach 96% if promised Kimi weights are released. The cost case for running models on-premises depends on usage, infrastructure and staffing.
IBM CEO Arvind Krishna said on the company’s Oct. 21 earnings call that open-weight AI models can cost 10% to 25% as much as frontier models while being “90% as effective,” making them a potential option for companies that want to run AI on their own infrastructure. An independent comparison published by Artificial Analysis and read Oct. 26 put the best model with downloadable weights at 80% of the top proprietary model, leaving the precise performance gap dependent on which models and measures are compared.
Krishna said a large enterprise could begin on-premises deployment with “two or three racks”. He argued that many routine tasks, such as summarizing email, reviewing contracts and answering customer calls, may not require the strongest available model. He also cited data control as a reason banks and trading firms might prefer private infrastructure, saying they want to run AI so that “nobody can take any of the underlying data or underlying logs.”
The cost and performance figures on the call are Krishna’s assessment, not a standardized benchmark result. He described the gap between open and closed models as narrowing from late 2024, when he said leading models might score 70 against 20 or 30 for open models, to late 2025, when he described a 70-to-50 gap. His “90%” comparison was a broad illustration rather than a result tied to a named evaluation.
Artificial Analysis’s Intelligence Index offers one independent comparison. In its data read Oct. 26, Xiaomi’s MiMo-V2.6-Pro scored 46.3, compared with 57.6 for the leading proprietary model, Claude Opus 5.5. That makes the open model about 80% of the leader on that index. Moonshot’s Kimi K3.1 Pro scored 55.1, or about 96% of the leader, but its weights had not yet been released. Moonshot had promised them by Nov. 4, so it was not yet downloadable as of the comparison date.
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 How Much AI Runs
Krishna’s pitch matters because enterprises weighing AI deployment face a choice between paying a provider for model access and investing in hardware and people to operate models themselves. Open weights can provide more control over where a model runs, but they do not make deployment free: companies need suitable accelerators, power, cooling, networking and staff to maintain the system.
The source’s list-price comparison illustrates why utilization matters. An eight-GPU H200 server was listed from $172,643, while renting the same GPU class on AWS was priced at $63.30 an hour. The calculation estimated that a purchase could match rental costs in about 114 days at full use, but take roughly 15 months at 25% use, before operating expenses. Those are illustrative figures, not a universal break-even point; actual prices and workloads vary.
There is also a strategic interest behind IBM’s argument. IBM sells on-premises hardware and expects demand for private AI infrastructure to support sales of its Spyre accelerators. Krishna acknowledged a skills gap, asking where enterprises would find kernel engineers and people to manage the underlying network. He said that issue remained for all but the most sophisticated companies, while cloud services were the easier route despite being more expensive.
Products Worth Considering
Samsung 27 Odyssey OLED G5 (G50SF) QHD QD-OLED Gaming Monitor
Amazon Kindle Scribe Colorsoft 64GB (1st Generation)
Samsung 27-inch Odyssey OLED G6 QD-OLED QHD Gaming Monitor
As an affiliate, we earn on qualifying purchases.
A Narrowing Gap, Not a Fixed Score
The Artificial Analysis comparison shows the open-to-proprietary score ratio rising from 51% at the end of 2024 to 74% at the end of 2025 and 80% in the Oct. 26, 2026 reading. Those figures indicate the leading downloadable models have gained ground relative to the index leader. They do not validate Krishna’s exact 90% claim: the comparison uses a particular index, and the score gap in points was 7.5 at the end of 2024 and 11.3 in the latest reading because closed models also improved.
“Open” also describes access to model weights, not necessarily unrestricted use. The source says MiMo-V2.6-Pro and DeepSeek V4.1 Flash 1017 use MIT licences, while some other models have custom or commercial terms. Licensing conditions differ, so companies need to check the relevant licence rather than infer commercial rights from a model’s presence on a leaderboard.
Cost comparisons also depend on how a model is accessed. The source reports that 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 examples show that a lower-cost model is not automatically cheaper to run on a company’s own hardware; workload, pricing and infrastructure determine the economics.
Performance and Payback Remain Workload-Specific
It is not yet clear whether Kimi K3.1 Pro’s weights will be released on the promised schedule, or what licence will govern them. Until they are available, its near-leader score does not establish that enterprises can download and run the model. The source also says no licence had been published at the time of its report.
There is no single score that establishes how well a model will perform across every company’s tasks. The index offers a consistent comparison across its evaluations, but real-world results can differ by use case, model version and deployment setup. The source’s calculation for IBM’s llm-d project, which IBM said can run large models at five to 10 times lower cost, was not independently verified there.
The hardware payback examples use list prices and simplified utilization assumptions. Staffing, cooling and networking were excluded, and costs may change with negotiated cloud rates, electricity prices and system load. IBM has a commercial interest in on-premises deployments, and the source notes that the company expects them to support sales of its Spyre cards. That interest does not disprove Krishna’s case, but it is relevant when weighing the company’s claims.
Watch for Weights and Real-World Costs
The next near-term test is whether Moonshot releases Kimi K3.1 Pro’s weights by Nov. 4, as promised, and under what licence. If that happens, the model could be compared as a downloadable option rather than as a closed service, though its index score alone would not establish how practical it is for a particular company to deploy.
For enterprises, the next step is to compare model performance and total operating costs against their own workloads. That includes checking licence terms, estimating usage, testing output quality and accounting for infrastructure and staffing. IBM’s earnings-call remarks set out the company’s case for this approach; they do not settle whether buying hardware beats cloud access for a given customer.
Key Questions
What did IBM CEO Arvind Krishna say about open-weight AI?
Krishna said open-weight models cost 10% to 25% of frontier-model token costs while being “90% as effective.” He also said a large enterprise could begin with two or three racks. These are his claims, not independent measurements.
How did the leading downloadable model compare with the top proprietary model?
In the Artificial Analysis Intelligence Index reading dated Oct. 26, 2026, MiMo-V2.6-Pro scored 46.3 against 57.6 for Claude Opus 5.5, about 80% of the proprietary leader’s score.
Could Kimi K3.1 Pro close the gap?
Its reported score of 55.1 is about 96% of the index leader’s, but Moonshot had not released its weights as of Oct. 26. It had promised them by Nov. 4; the release and licence remained unconfirmed in the source material.
Does running an open model on-premises always cost less?
No. The economics depend on utilization, hardware prices, power, cooling, networking and staffing. The source’s example estimated a much longer purchase payback at 25% use than at full load, and excluded several operating costs.
Source: ThorstenMeyerAI.com
