🔍 Read the full analysis: Why IBM Links On-Prem AI To Scale on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
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 10% to 25% of the token cost, making on-premises AI attractive to some enterprises. Artificial Analysis data read October 26 puts the leading downloadable model at 80% of the top proprietary model; buying hardware can also be economical only when it is used heavily, before operating costs.
IBM CEO Arvind Krishna said on the company’s October 21 earnings call that open-weight AI models can cost 10% to 25% as much per token as frontier models while being “90% as effective,” strengthening IBM’s case for running AI on company-owned infrastructure. The pitch comes with qualifications: an independent benchmark puts the leading downloadable model at 80% of the leading proprietary model, and the economics of buying servers depend on how consistently they are used.
In the earnings-call Q&A, Krishna described open models as a lower-cost option for routine work such as summarizing email, reviewing contracts and answering customer calls. He said an enterprise could begin with two or three racks. He also argued that private infrastructure can help organizations, including banks and trading firms, keep data and logs under their control. Open weights, he said, give customers permission and intellectual-property rights to run models on premises or in private clouds.
The Artificial Analysis Intelligence Index, checked on October 26, gives Xiaomi’s MiMo-V2.6-Pro a score of 46.3, compared with 57.6 for proprietary Claude Opus 5.5. That makes the best downloadable model about 80% of the proprietary leader on the index. The gap has narrowed as open models improved, but Krishna’s 90% comparison is not a direct reading from a benchmark he named on the call.
Hardware costs can change the calculation. The source’s comparison lists an eight-GPU H200 server from $172,643 and AWS rental of the same GPU class at $63.30 an hour. At full use, the purchase price alone would be recovered in about 114 days against that rental rate; at 25% use, the estimate is about 15 months. Both figures exclude staff, cooling and networking, and are illustrative comparisons rather than a complete cost assessment.
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 Depend on Utilization
Krishna’s argument matters because it frames AI infrastructure as a choice between paying a provider for each use and investing in capacity a company controls. For organizations processing large, predictable volumes, lower token prices and data control may support a case for private deployment. But buying servers transfers responsibility for infrastructure and operations to the customer.
That burden is not limited to the purchase price. Krishna himself identified a shortage of kernel engineers and network specialists as a constraint, saying the skills gap remains outside the most sophisticated enterprises. A company that cannot keep its systems busy or staff them effectively may not realize the savings implied by a model’s token price. IBM also has a commercial interest: the source says it expects on-premises demand to support sales of its Spyre cards.
The comparison is not simply open versus closed or owned versus rented. A model’s licence, quality, workload, utilization and operating requirements all affect the result. The source notes that some open models use MIT licences while others have custom or commercial terms. A strong leaderboard score alone does not establish that a model can be used freely for every business purpose.
Popular right now
Vitamix 5200 Professional-Grade Blender, 64 oz. Container, Black
Shark Rocket Pet Plus Corded Stick Vacuum, Bordeaux, HV322
De'Longhi Linea Classic Espresso Machine with Adjustable Milk Frother
As an affiliate, we earn on qualifying purchases.
Model Scores and IBM’s Pitch
Krishna’s “90% as effective” phrasing was an illustrative comparison, not a figure tied to a named test in his remarks. He described open models as gaining ground: in his account, the broad gap moved from roughly 70 versus 20 or 30 in late 2024 to 70 versus 50 in late 2025. Those examples should not be treated as benchmark results.
Artificial Analysis provides a separate comparison using ten evaluations. Its figures in the supplied material show the leading downloadable model at 51% of the proprietary leader at the end of 2024 and 80% on October 26, 2026. The absolute score gap, however, went from 7.5 points to 11.3 points as both categories advanced. The source also says Kimi K3.1 Pro scores 55.1, close to Claude Opus 5.5, but its weights had only been promised by November 4. Until they are released, it is not an available open-weight option.
IBM also said in prepared remarks that its llm-d project can run large models at five to ten times lower cost. That is an IBM claim; the source does not provide an independent measurement verifying it. The company’s interest in selling on-premises hardware is relevant when weighing its broader cost argument.
Costs, Licences and Weights
The figures do not settle whether on-premises AI will save money for a particular buyer. The server comparison excludes staffing, cooling and networking, and it assumes a specific rental rate and utilization level. Actual workloads, hardware configurations and cloud prices may differ. The source gives no full cost model for the examples.
It is also unclear how broadly Krishna’s 90% comparison applies: he did not identify a benchmark, model pair or set of tasks for that claim. Artificial Analysis offers one comparison, but its index scores do not by themselves establish performance on each company’s work. Kimi K3.1 Pro’s weights and licence remained unconfirmed in the supplied report, despite a promised release date of November 4.
Finally, licensing varies across models, and the right to download weights does not automatically mean unrestricted commercial use. Buyers must check each model’s terms as well as its score. IBM’s claims about its own software and the financial case for its hardware have not been independently established by the material provided.
The November Weights Test
The next near-term check is whether Kimi K3.1 Pro’s weights arrive by November 4, as promised in the source, and what licence accompanies them. That would affect whether its 55.1 index score can be counted as an available open-weight alternative, rather than a score for a model whose weights have not shipped.
For enterprise buyers, the practical next step is to test candidate models against their own workloads and calculate total costs at realistic levels of utilization. That means accounting for licenses, hardware, cloud alternatives, power, networking and specialist staff—not just comparing per-token prices. IBM’s pitch sets out the case for on-premises AI; customer-specific evidence will determine whether it pays.
Key Questions
What did IBM’s CEO say about open-weight AI?
Arvind Krishna said open-weight models can cost 10% to 25% as much per token as frontier models while being “90% as effective.” He presented this as an argument for enterprise use, not as a result from a named benchmark.
Do independent benchmark figures support the 90% claim?
Not for the leading downloadable model listed in the source. Artificial Analysis data read October 26 puts MiMo-V2.6-Pro at 46.3 versus 57.6 for Claude Opus 5.5, or about 80% of its score. The source says Kimi K3.1 Pro scores closer to the leader, but its weights had not yet shipped.
When can owning AI hardware beat renting it?
It depends heavily on use. The supplied comparison estimates that an $172,643 eight-GPU H200 server would match the listed AWS rental cost in about 114 days at full utilization or about 15 months at 25% utilization. Those estimates omit staff, cooling and networking.
Why might a company run models on premises?
Krishna pointed to data control and sovereignty, as well as potential savings for workloads that use AI heavily. Private deployment can also give an organization control over where its model runs, but it requires suitable hardware and technical staff.
Are all open-weight models free for commercial use?
No. The source identifies models with MIT licences as well as models released under custom or commercial terms. Companies need to review the specific licence for each model rather than assume that downloadable weights permit unrestricted use.
Source: ThorstenMeyerAI.com
