🔍 Read the full analysis: Can Volume Make On-Prem AI Pay Off? IBM’s CEO Shares His View on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
IBM CEO Arvind Krishna said open-weight models can deliver about 90% of frontier-model performance at 10% to 25% of the token cost, making on-premises AI attractive for some enterprise workloads. An independent benchmark puts the best currently downloadable model at 80% of the leading proprietary model; the economics of buying hardware depend heavily on utilization and operating costs.
IBM CEO Arvind Krishna said on the company’s October 21 earnings call that open-weight AI models can be about 90% as effective as leading proprietary systems while costing 10% to 25% as much per token, renewing the case for enterprises to run AI on their own infrastructure. An independent benchmark puts the best model with downloadable weights at 80% of the top proprietary model as of October 26, while the practical savings from buying servers depend on how steadily they are used.
Krishna made the comparison during the question-and-answer portion of the call. He said the gap between leading and open models had narrowed, describing an earlier period when a top model might score 70 while an open model scored 20 or 30, followed by a later 70-to-50 comparison. He framed the current choice as paying a fraction of the price for a model that is close enough for many routine tasks. His examples included summarizing emails, reviewing contracts, answering customer calls and composing messages.
On-premises deployment is also a matter of control, Krishna said. Banks and trading firms may want AI to operate “in a way that is sovereign,” with sensitive data and logs remaining under their control. Open-weight models, he said, give organizations permission and intellectual-property rights to run models on their own infrastructure or private clouds. IBM has a commercial interest in that shift: it sells enterprise hardware, and Krishna said large companies could begin with two or three racks. IBM also said its llm-d project can run large models at five to 10 times lower cost, a company claim that the source material does not independently verify.
The Artificial Analysis Intelligence Index offers one comparison, though it is not the benchmark Krishna cited. On the index version read October 26, MiMo-V2.6-Pro was the highest-scoring model with downloadable weights, at 46.3, compared with 57.6 for proprietary Claude Opus 5.5. That puts the open model at about 80% of the proprietary model’s score. The source’s historical figures show the share at 51% at the end of 2024 and 74% at the end of 2025. These are index scores, not direct measures of how well every model performs for every business 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
When On-Prem AI Can Save Money
The argument matters to organizations weighing recurring cloud charges against upfront infrastructure costs. Open weights can give businesses more options: they may choose a lower-cost model, run it privately, and tailor deployment to workloads where the most capable frontier system is unnecessary. That may be particularly appealing where data handling and control are major concerns.
But lower token prices do not, by themselves, prove that owning hardware is cheaper. The source compares an eight-GPU H200 server listed from $172,643 with renting the same GPU class on AWS for $63.30 an hour. On those figures, the purchase price equals about 114 days of rental at continuous full use, or about 15 months at 25% utilization. Those are simplified comparisons before staffing, cooling and networking costs; they are not a complete cost-of-ownership calculation. Krishna also acknowledged that many enterprises lack the kernel and networking engineers needed to operate the systems. For buyers, capacity utilization and staffing could determine whether the promised savings materialize.
Products Worth Considering
Amazon Kindle Paperwhite Signature Edition (12th Generation, 2024), 32 GB, Metallic Black
Apple iPad Air (2025) 13‑inch Wi‑Fi 256 GB – Space Gray (M3 chip)
Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
As an affiliate, we earn on qualifying purchases.
The Model Gap and Its Caveats
Krishna’s percentages were illustrative comparisons, not results from a named evaluation. Artificial Analysis provides a separate, consistent index of models, but a single score cannot establish that open models are 90% as capable across all tasks. Its October 26 figures show the best downloadable model at 46.3 and the best proprietary model at 57.6. The relative gap has narrowed from the historical comparison supplied by the source, though the absolute score difference is now 11.3 points, compared with 7.5 points at the end of 2024.
One potential exception is Kimi K3.1 Pro, which scores 55.1 on the index, close to Claude Opus 5.5. But its weights had not been released as of the reported date; Moonshot had promised them by November 4. Until they are available, it should not be counted as a downloadable open-weight model. Scores are only part of the decision: model licences vary, and businesses need to check whether a model’s terms permit their intended commercial use.
Costs, Weights and Real-World Performance
Several parts of the business case remain unsettled. The available material does not establish what proportion of enterprise workloads can use open-weight models without unacceptable performance trade-offs, or whether the benchmark scores predict results on each company’s own tasks. Nor does the hardware comparison include the full costs of operating and maintaining a deployment. Utilization, staffing, electricity, cooling and networking can change the calculation.
The stated token-cost advantage also depends on the model and how it is accessed. The source says hosted DeepSeek V4.1 Flash 1017 costs 3% to 6% of Claude Opus 5.5’s list price, but gives no complete, like-for-like accounting for on-prem deployment. It also notes that a closed model, OpenAI’s GPT-6 Luna, lists at $0.10 per million input tokens and $0.50 per million output tokens. The source does not provide enough information to compare total costs across providers and owned hardware on a common basis.
Finally, the benchmark comparison could change if promised weights are released or new model versions appear. Kimi’s stated November 4 target had not yet been met at the time of the October 26 figures. Availability, licence terms and actual business performance remain separate questions from a model’s score.
What Enterprise Buyers Will Test
The next evidence will come from model releases and deployments that can be compared on real workloads. Buyers considering on-premises systems will need to test whether an open-weight model meets their accuracy, latency and privacy requirements, then estimate costs using realistic utilization and staffing assumptions. They will also need to review each model’s licence rather than treating all downloadable weights as interchangeable.
Moonshot’s stated November 4 deadline for Kimi K3.1 Pro weights is one near-term point to watch; their release and licence status were still pending in the supplied October 26 account. IBM has also positioned its hardware and llm-d software around lower-cost inference, but the source provides no independent measurement confirming IBM’s five-to-10-times cost claim. The core question for enterprises is not only whether an open model approaches a frontier score, but whether it can do so reliably and at lower total cost in their own environment.
Key Questions
What did IBM CEO Arvind Krishna claim about open-weight models?
Krishna said open-weight models can be about 90% as effective as frontier models at 10% to 25% of their token cost. He presented this as a broad comparison, not the result of a named benchmark.
How do the independent benchmark figures compare?
On the Artificial Analysis index read October 26, 2026, MiMo-V2.6-Pro scored 46.3 and Claude Opus 5.5 scored 57.6. The open model’s score was about 80% of the proprietary leader’s score on that index.
Does buying AI servers guarantee lower costs?
No. The source’s simplified H200 comparison estimates a purchase-price break-even of about 114 days at full use or 15 months at 25% use, before staffing, cooling and networking. Actual economics depend on utilization and operating costs.
Why might a company choose on-premises AI?
Some organizations may want more control over sensitive data, logs and deployment. Krishna said banks and trading firms have sovereignty concerns, while open weights can allow models to run on private infrastructure. Companies still need appropriate technical staff and must check model licences.
Was Kimi K3.1 Pro available as an open-weight model?
Not in the October 26 account. It scored 55.1 on the index, but Moonshot had only promised to release its weights by November 4. Until the weights are released, it should not be treated as downloadable.
Source: ThorstenMeyerAI.com
