🔍 Read the full analysis: IBM’s CEO Says In-House AI Could Depend On Volume 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
On IBM’s Oct. 21 earnings call, CEO Arvind Krishna argued that open-weight AI models can deliver about 90% of frontier-model performance for 10% to 25% of the token cost, and said enterprises can begin with a few racks. An independent benchmark comparison in the supplied source puts the strongest downloadable model at 80% of the leading proprietary model; buying hardware instead of renting is most attractive when it stays heavily used.
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 per token as frontier models while being “90% as effective.” His argument is that enterprises with enough demand could run those models on their own infrastructure, but the source’s independent benchmark comparison puts the best downloadable model at 80% of the leading proprietary model as of Oct. 26.
Krishna made the comparison during the call’s question-and-answer session. He described model performance as having narrowed over time: in his account, a leading model scored around 70 while an open model scored 20 or 30 in late 2024, moving to 70 versus 50 by late 2025. He then framed the choice as paying a fraction of the cost for a model that is nearly as capable. Those scores were illustrative; he did not identify a specific benchmark for them.
Krishna named routine business tasks—including summarizing email, reviewing contracts and answering customer calls—as work that may not require the most capable model. He also argued that open-weight models allow companies to run systems on their own infrastructure or private clouds, a possible fit for organizations concerned about keeping data and logs under their control. Open weights do not, by themselves, establish that a model is suitable for every deployment or that its licence permits every intended use.
The source’s comparison of Artificial Analysis’s Intelligence Index, read Oct. 26, lists MiMo-V2.6-Pro at 46.3 as the highest-scoring model with downloadable weights, against Claude Opus 5.5 at 57.6 among proprietary models. That puts the downloadable model at about 80% of the proprietary leader’s score on that index. The comparison also says Kimi K3.1 Pro scored 55.1, but its weights had only been promised by Nov. 4; it was not downloadable at the time of the comparison.
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
Why Utilization Shapes AI Costs
The argument is not simply that open-weight models are cheaper. It is that a company may be able to trade some model performance for lower per-token spending, then choose where the system runs. For businesses with sustained, predictable workloads, owning the infrastructure could reduce costs and give them more control over data handling. For occasional or uneven workloads, expensive hardware sitting idle can erase that advantage.
The source’s illustrative comparison prices an eight-GPU H200 server from $172,643 and the same GPU class on AWS at $63.30 an hour. It estimates a purchase would match rental costs after about 114 days at full utilization, versus about 15 months at 25% utilization. These are estimates based on listed prices, not a full accounting: the calculation excludes staffing, cooling and networking, and does not establish the total cost of a working deployment.
That distinction matters for enterprise buyers deciding whether to rent capacity or invest in equipment. Krishna himself acknowledged that running systems on-premises requires people who understand model kernels and networking. A lower token price alone does not resolve infrastructure, operations, licensing or performance requirements.
Popular right now
De'Longhi Stilosa Compact Manual Espresso Machine with Milk Frother
Vitamix Propel Series 750 Professional-Grade Blender, 4 Programs, Black
Dyson Cyclone V10 Origin Cordless Vacuum
As an affiliate, we earn on qualifying purchases.
The Benchmark Gap Behind IBM’s Pitch
The figures in the source’s Artificial Analysis comparison show the open-to-proprietary score ratio rising from 51% at the end of 2024 to 74% at the end of 2025, then 80% on Oct. 26, 2026. The source reports the absolute score gap at 7.5 points in late 2024 and 11.3 points in October 2026. In other words, the open model’s score became a larger share of the leader’s, even as the point difference widened because the proprietary leader’s score also rose.
These results are specific to the Artificial Analysis index and its evaluations; they do not establish equivalent performance on every company’s tasks. The source also notes that models differ in licensing. MiMo-V2.6-Pro and DeepSeek V4.1 Flash 1017 are described there as MIT-licensed, while other models have their own commercial or community terms. A model’s score and token price therefore do not settle whether a company can legally or practically use it as planned.
IBM has a commercial interest in the shift Krishna described: the company sells infrastructure and expects on-premises AI demand to support sales of its Spyre cards. The source also reports that IBM said its llm-d project can run large models at five to 10 times lower cost, a company claim that the author says was not independently measured. That interest is relevant when weighing the pitch, but does not by itself disprove it.
Unsettled Costs and Model Availability
The source does not provide the exact benchmark Krishna used for his 90% comparison, so that statement cannot be directly checked against the Artificial Analysis figures. The index’s 80% result is one comparison on one evaluation set, not a universal measure of how well a model performs on a company’s particular work.
It is also unclear whether Kimi K3.1 Pro’s weights will become available on the promised date, under what licence they might be released, or what hardware and operating costs would apply. The source’s server-versus-rental figures are list-price estimates and leave out several major expenses. Actual break-even points will vary with negotiated prices, utilization, staffing, power, cooling and network needs.
Krishna’s cost and capability statements are IBM’s characterization, while the source’s benchmark and hardware calculations are separate comparisons. The available material does not establish that every enterprise can achieve the cited token costs or run models successfully with only two or three racks.
What Buyers Need to Verify
Enterprise buyers will need to compare candidate models on their own workloads, confirm that the licence covers their planned use, and estimate total operating costs alongside token prices. For an on-premises plan, the practical questions include expected utilization, hardware availability, staffing, networking and how the system will be maintained.
The next evidence point in the source’s account is whether Kimi K3.1 Pro’s weights arrive as promised by Nov. 4, 2026. Further releases and updated benchmark results could change the open-versus-proprietary comparison. IBM’s stated case will also depend on whether customers actually adopt its infrastructure and whether independent measurements support its cost claims; the supplied material gives no customer deployment results or verified savings.
Key Questions
What did IBM CEO Arvind Krishna claim?
On IBM’s Oct. 21, 2026, earnings call, Krishna said open-weight models could cost 10% to 25% as much per token as frontier models while being “90% as effective.” He presented this as a general comparison, not a result tied to a named benchmark.
How close are downloadable models to proprietary models in the cited benchmark?
In the Artificial Analysis Intelligence Index comparison read Oct. 26, the top downloadable model, MiMo-V2.6-Pro, scored 46.3, while Claude Opus 5.5 scored 57.6. The source calculates the downloadable model’s score as about 80% of the proprietary leader’s.
Does running an open-weight model on-premises always save money?
No. The source’s estimate suggests the purchase-versus-rental case changes sharply with utilization: its illustrative eight-GPU H200 comparison reaches cost parity after about 114 days at full load, but about 15 months at 25% load. The estimate excludes staffing, cooling and networking.
What did Krishna say about the challenges of on-premises AI?
He pointed to shortages of people with expertise in model kernels and networking. His remarks suggest that hardware cost is only one part of deploying and operating AI systems inside an enterprise.
Are the claims about IBM’s cost savings independently verified?
Not in the supplied material. The source attributes the cost comparison to Krishna and says IBM’s separate claim that its llm-d project lowers the cost of running large models by five to 10 times was not independently measured there.
Source: ThorstenMeyerAI.com
