🔍 Read the full analysis: Open-Weight Models And On-Prem AI: IBM’s CEO Explains The Bet 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 argued on the company’s October 21 earnings call that open-weight models can deliver about 90% of frontier-model performance at a fraction of the token cost, making on-premises AI attractive to some enterprises. An October 26 Artificial Analysis comparison put the leading downloadable model at 80% of the top proprietary model; a model that could narrow the gap further had not yet released its weights. The economics of buying hardware depend heavily on how much it is used.
IBM CEO Arvind Krishna said on October 21 that open-weight AI models can cost 10% to 25% as much per token as frontier models while being “90% as effective,” making a case for enterprises to run AI on their own infrastructure. But an October 26 comparison from Artificial Analysis put the best model with downloadable weights at 80% of the top proprietary model’s score, and buying servers only beats renting when equipment is used enough to justify the upfront cost.
Krishna made the comparison during the Q&A on IBM’s third-quarter earnings call, rather than in his prepared remarks. He described open models as catching up with closed systems and argued that many routine business tasks do not require the most capable model. His examples included summarizing email, examining contracts and answering customer calls. He said a large enterprise could start with “two or three racks,” though he also acknowledged that operating the infrastructure requires specialist engineers and networking expertise.
The independent comparison cited in the source uses Artificial Analysis’s Intelligence Index, which applies the same ten evaluations across models and identifies whether weights can be downloaded. As of October 26, the leading downloadable model listed was Xiaomi’s MiMo-V2.6-Pro, at 46.3; the leading proprietary model, Claude Opus 5.5, scored 57.6. That puts the open model at about 80% of the proprietary leader’s score. The source’s historical comparison gives open models 51% of the leader’s score at the end of 2024 and 74% at the end of 2025.
One possible exception had not yet become an open model by the stated cutoff. Kimi K3.1 Pro scored 55.1, about 96% of Claude Opus 5.5, but its maker, Moonshot, had only promised to release the weights “by Nov 4.” Until weights are available, it cannot be counted as downloadable. The source also cautions that model scores alone do not establish whether a licence permits a company’s intended use.
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 Economics of Running Models
For businesses, Krishna’s pitch joins three decisions: which model is good enough, where its data will be processed, and whether to rent computing or buy it. Open weights may let customers choose their own hosting and avoid sending sensitive data to an outside provider. Krishna said banks and traders want AI to run “in a way that is sovereign,” with underlying data and logs kept from outside access. Those are his account of customer needs, not a guarantee that every on-prem deployment will meet a particular security or regulatory standard.
The cost comparison is also conditional. The source gives an example of an eight-GPU H200 server listed from $172,643 and AWS rental of the same GPU class at $63.30 an hour. On those figures, the purchase cost is recovered in about 114 days at full utilization, but roughly 15 months at 25% utilization, before staffing, cooling and networking. Low usage can make renting less costly; buying shifts the burden to operating and maintaining the equipment.
Nor does a lower model price automatically mean lower total costs. The source says hosted DeepSeek V4.1 Flash 1017 is priced at 3% to 6% of Claude Opus 5.5’s listed price, while OpenAI’s closed GPT-6 Luna is listed at $0.10 per million input tokens and $0.50 per million output tokens. The comparison shows why token prices must be checked model by model: openness is not, on its own, a reliable proxy for the cheapest service.
Products Worth Considering
Therabody Theragun Relief Massage Gun, 3-Speed, Black
Oral-B iO Series 5 Rechargeable Electric Toothbrush with Travel Case and 3 Brush Heads, Black
Theragun Mini 3rd Generation Massage Gun – Travel Essential (Black)
As an affiliate, we earn on qualifying purchases.
IBM’s Pitch and Its Incentives
IBM has a commercial interest in on-premises AI: it sells enterprise hardware and services, and Krishna expects increased on-prem demand to support sales of IBM’s Spyre accelerator cards. That interest does not disprove his claims, but it is relevant when evaluating them. The source also reports that IBM’s mainframe revenue fell 20% in the quarter, 18 months into the z17 cycle, citing The Register.
Krishna’s examples of the performance gap were illustrative, not results from a named benchmark. He recalled a broad comparison in late 2024 as a score of about 70 for a leading model against 20 or 30 for an open model, followed by a 70-to-50 comparison in late 2025. Artificial Analysis’s index offers one way to test the direction of his argument, but its figures are not directly interchangeable with those examples. In the index series supplied by the source, the open-model share rose from 51% at the end of 2024 to 80% on October 26, 2026; the score gap in points rose from 7.5 to 11.3 over those same listed dates as the leading proprietary model also improved.
Licensing and technical capacity complicate the choice. Some models in the comparison use MIT licences, while others have custom or commercial terms. IBM also said in its prepared remarks that its llm-d project can run large models at five to ten times lower cost; that is IBM’s stated claim, and the source says it could not independently check the measurement. Krishna himself identified a shortage of kernel engineers and network specialists as a constraint for many enterprises.
Open Weights and Costs Still Vary
Several details remain unsettled. Moonshot’s promised November 4 release could change the comparison if Kimi K3.1 Pro’s weights arrive, but the source says no licence had been published by October 26. Its reported 2.8 trillion parameters would also make the model a data-centre-scale undertaking, according to the source; a high benchmark score does not show that an ordinary enterprise can run it economically.
The server payback figures are estimates based on list prices and stated utilization assumptions. They exclude staff, cooling and networking, and the source does not establish that a particular company would achieve continuous full load. Actual costs will depend on usage, hardware configuration, energy, support, model licensing and the performance required for a given task.
There is also no independent measurement in the supplied material confirming IBM’s claimed llm-d savings or a general 90%-performance figure across business workloads. The Artificial Analysis Index is a specific benchmark comparison, not proof that models will perform equally on each company’s tasks. Nvidia had not commented on the reported GeForce firmware workaround described in the source; it is separate from the enterprise case and does not establish a broader change in AI economics.
Watch for Weights and Utilization
The near-term test is whether Moonshot releases Kimi K3.1 Pro’s weights by its stated November 4 target, and what licence accompanies them. Until then, it remains a closed model for purposes of the downloadable-model comparison. Buyers considering open-weight systems will need to compare licences, task-specific results and hosting costs, rather than relying on a single index score.
For on-premises deployments, the practical next step is a workload and utilization calculation that includes staff, power, cooling, networking and the cost of idle capacity. IBM’s earnings-call pitch puts those decisions at the center of its AI strategy, but the source does not provide a confirmed enterprise deployment or measured customer savings that settles the case. Further benchmark updates and published deployment costs may clarify how often lower token prices translate into lower total costs.
Key Questions
What did IBM CEO Arvind Krishna say about open-weight models?
On IBM’s October 21 earnings call, Krishna said open-weight models could cost 10% to 25% as much per token as frontier models while being “90% as effective.” That was his assessment, not a result from the independent index cited in the source.
How close were the best downloadable models to the top proprietary model?
In Artificial Analysis’s comparison read October 26, 2026, Xiaomi’s MiMo-V2.6-Pro scored 46.3, compared with 57.6 for Claude Opus 5.5. The open model’s score was about 80% of the proprietary leader’s score.
Was Kimi K3.1 Pro open-weight on October 26?
No. It scored 55.1, but Moonshot had only promised to release its weights by November 4. The source says the model could not be counted as downloadable until the weights shipped.
Is buying AI servers cheaper than renting cloud GPUs?
It depends on utilization and operating costs. The source estimates that an eight-GPU H200 server priced from $172,643 would recoup its purchase price in about 114 days at full load or around 15 months at 25% load, before staffing, cooling and networking.
Source: ThorstenMeyerAI.com
