🔍 Read the full analysis: A Business Guide To IBM’s Self-Run AI And On-Prem Volume Bet on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- 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. An October 26 Artificial Analysis comparison put the best downloadable model at 80% of the top proprietary model; hardware purchase economics also depend heavily on utilization, staffing and operating costs.
IBM CEO Arvind Krishna told investors on the company’s October 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 fit for businesses running AI on their own systems. An independent benchmark comparison read October 26 put the leading downloadable model at 80% of the top proprietary model, underscoring that the case for on-premises AI rests on more than performance alone.
Krishna made the cost and performance comparison during the call’s question-and-answer session. He described open models as catching up with closed systems and said organizations could begin on-premises deployments with “two or three racks.” He pointed to routine tasks such as summarizing email, reviewing contracts and answering customer calls as work that may not require the most capable model. IBM also said its llm-d project can run large models at “5 to 10 times lower cost”; that is the company’s claim, and the source material does not provide an independent measurement of it.
The Artificial Analysis Intelligence Index offers a separate comparison, though it is not the benchmark Krishna cited. In figures read October 26, the leading downloadable model, Xiaomi’s MiMo-V2.6-Pro, scored 46.3, against 57.6 for proprietary Claude Opus 5.5. That makes the open model about 80% of the proprietary leader’s score on the index. Kimi K3.1 Pro scored 55.1, but its weights had not shipped: Moonshot had promised them by November 4. Until release, the model cannot be counted as downloadable.
The cost comparison also depends on how a business gets access to a model and how much it uses its hardware. The source cites an 8-GPU H200 server listed from $172,643 and an AWS rental price of $63.30 an hour for the same GPU class. A simple comparison puts purchase payback at about 114 days under continuous full utilization, or about 15 months at 25% utilization, before staffing, cooling and networking costs. Those assumptions make the figures a rough illustration, not a general guarantee that buying is cheaper.
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 AI On-Site
For companies processing sensitive information or large volumes of routine requests, the decision is not simply whether an open model matches the best closed model. It is whether the quality is sufficient for a particular task, whether the licence permits the intended use, and whether the organization can operate the infrastructure effectively. Krishna said some financial firms want AI to run “in a way that is sovereign,” so data and logs remain under their control. Open weights can allow deployment on company-owned systems or private clouds, but access to weights does not by itself supply the hardware, engineering staff or security arrangements.
The utilization math makes scale central to the purchase decision. A server that is busy around the clock can spread its cost across many workloads; a lightly used server ties up capital while continuing to incur operating expenses. Krishna himself acknowledged a skills bottleneck, asking where enterprises would find “kernel engineers” and people to manage the underlying network. That qualification matters for businesses weighing a rack purchase against hosted services, which may cost more per token but require less in-house infrastructure work.
Popular right now
De'Longhi Stilosa Compact Manual Espresso Machine with Milk Frother
Bose QuietComfort Headphones (2nd Gen), Wireless Noise-Cancelling Over-Ear Headphones, Eucalyptus Green
KitchenAid Artisan 5-Quart Tilt-Head Stand Mixer, KSM150PS, Almond Cream
As an affiliate, we earn on qualifying purchases.
How the Model Gap Has Shifted
The index figures in the source show the best downloadable model rising from 51% of the leading proprietary score at the end of 2024 to 74% at the end of 2025 and 80% on October 26, 2026. These are comparisons within Artificial Analysis’s index, not a verification of Krishna’s illustrative 90% figure. The source says the absolute score gap on that index was 7.5 points at the end of 2024 and 11.3 points in October 2026, because the leading proprietary model also improved.
Prices and permissions vary by model. The source lists DeepSeek V4.1 Flash 1017, whose weights are MIT-licensed, at 3% to 6% of Claude Opus 5.5’s list price when hosted by its maker. It also gives a contrasting example: closed model GPT-6 Luna is listed at $0.10 per million input tokens and $0.50 per million output tokens. Those examples do not establish a universal price advantage for open models; hosting, usage, model capability and licensing all affect the comparison. A model’s licence should be checked separately from its benchmark score.
IBM has a commercial interest in the on-premises case: Krishna said the company expects demand to support sales of its Spyre cards. The source also reports that IBM mainframe revenue fell 20% in the quarter, 18 months into the z17 cycle, citing The Register. That figure is separate from the AI claims and does not, by itself, establish why the company is emphasizing on-premises AI.
Open Weights and Cost Assumptions
Several parts of the comparison remain unsettled. Kimi K3.1 Pro’s weights had not been released as of the October 26 benchmark read, and the source says no licence had been published. Its score therefore cannot yet establish that a downloadable model has reached 96% of Claude Opus 5.5 on the index. The source also notes that the model’s reported 2.8-trillion-parameter size would itself require substantial data-centre resources.
The index provides one common scoring framework, but it does not establish how well each model performs on a specific company’s tasks or what its full operating cost would be. The hardware payback estimates exclude staff, cooling and networking, while the source does not detail all assumptions behind the prices. IBM’s llm-d cost claim is also not independently measured in the material provided. Nvidia has not commented on the reported GeForce firmware exploit described in the source, and the source says Nvidia’s GeForce licence forbids circumvention of technical limitations; that separate enthusiast development is not evidence about enterprise AI economics.
Model Releases and Enterprise Decisions
The next stated milestone for the model comparison is November 4, the date by which Moonshot had promised to release Kimi K3.1 Pro’s weights. If they are released, their licence and practical hardware requirements will affect whether the model qualifies as an enterprise-ready open-weight option; the promise alone does not settle either point.
For businesses, the next step is to test candidate models against their own workloads and measure total costs at realistic utilization. That means checking model quality, token prices, licensing, data-control requirements, hardware use and the availability of staff to maintain systems. IBM’s earnings call makes the on-premises case, but the benchmark and cost figures available on October 26 do not resolve whether it is the better choice for any particular organization.
Key Questions
What did IBM’s CEO say about open-weight AI models?
Arvind Krishna said on IBM’s October 21 earnings call that open-weight models can cost 10% to 25% as much as frontier models while being “90% as effective.” That is his stated comparison, not a finding independently established by the benchmark figures in the source.
How close are downloadable models to the leading proprietary model?
In the Artificial Analysis figures read October 26, 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 on that index.
Does buying an AI server always cost less than renting one?
No. The source’s estimate says an 8-GPU H200 server listed from $172,643 could match AWS rental costs in about 114 days at full utilization, or about 15 months at 25% utilization. Those estimates exclude staffing, cooling and networking, so actual costs can differ.
Why might an enterprise run AI on its own infrastructure?
Some organizations may want greater control over sensitive data and logs, or may find on-premises operation economical at high and steady usage. The choice also depends on model quality, licensing, hardware needs and whether the company has staff to operate the systems.
Source: ThorstenMeyerAI.com
