🔍 Read the full analysis: Could On-Prem AI Work For Your Company? IBM’s CEO On The Volume Bet on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
IBM CEO Arvind Krishna said open-weight AI models can deliver about 90% of frontier-model performance at 10% to 25% of the token cost, and argued that large enterprises can run them on their own infrastructure. An October 26 comparison from Artificial Analysis puts the leading downloadable model at 80% of its proprietary counterpart, while 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 cost 10% to 25% as much per token as frontier models while being “90% as effective,” making it practical for some large companies to run AI on their own infrastructure. An independent benchmark comparison available October 26 shows the best downloadable model at about 80% of the leading proprietary model’s score, underscoring that the savings case depends on the task, hardware use and how performance is measured.
In the earnings-call Q&A, Krishna described open models as catching up with proprietary systems and said large enterprises could begin with “two or three racks.” He pointed to routine work such as summarizing email, reviewing contracts and answering customer calls as tasks that may not require the strongest available model. He also said banks and trading firms value the ability to run AI “in a way that is sovereign,” keeping data and logs under their control. Open weights, he argued, give companies permission and intellectual-property rights to deploy models on premises or in private clouds.
The comparison cited in the source material is from the Artificial Analysis Intelligence Index, which applies the same ten evaluations across models. As read October 26, Xiaomi’s downloadable MiMo-V2.6-Pro scored 46.3, against 57.6 for proprietary Claude Opus 5.5 — roughly 80% of the latter score. The source says the open-model share was 51% at the end of 2024, indicating a narrowing relative gap, though its cited point difference grew from 7.5 to 11.3. Scores depend on the index and do not by themselves establish performance on a company’s specific work.
Krishna’s 90% figure is not tied to a named benchmark in the call. The source notes that Kimi K3.1 Pro scored 55.1, about 96% of Claude Opus 5.5, but Moonshot had only promised to release its weights by November 4. Until weights are available, it remains a closed model for deployment purposes. IBM also said its llm-d project can run large models at five to ten times lower cost; that is a company claim, not an independently verified measurement in the material provided.
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
The case for on-premises AI is not simply that open models are cheaper. Companies must compare model quality and token prices with the cost of buying and operating servers, as well as the value of data control and predictable access. The numbers supplied illustrate how much utilization changes the calculation: an eight-GPU H200 server lists from $172,643, while renting the same GPU class on AWS is listed at $63.30 an hour. At full use, the purchase price equals about 114 days of rental; at 25% utilization, the source estimates about 15 months, before staffing, cooling and networking.
Those estimates are not a complete ownership-cost analysis, and they compare listed prices rather than a negotiated quote or a specific enterprise deployment. Still, they show why a company with steady, high-volume workloads may evaluate purchasing differently from one whose demand is intermittent. Krishna himself identified shortages of kernel engineers and networking specialists as a barrier for all but the most sophisticated enterprises. A lower token price does not remove the operational work of running a model.
The price advantage also varies by provider and model. The source reports that hosted DeepSeek V4.1 Flash 1017, whose weights are MIT-licensed, costs 3% to 6% of Claude Opus 5.5’s list price. It also cites OpenAI’s closed GPT-6 Luna at $0.10 per million input tokens and $0.50 per million output tokens. That comparison cautions against treating “open” as a guarantee of the lowest cost: companies need to compare the actual model, hosting arrangement, licensing terms and workload.
Popular right now
KitchenAid Artisan 5-Quart Tilt-Head Stand Mixer, KSM150PS, Almond Cream
Philips 4400 Series Fully Automatic Espresso Machine with LatteGo Milk System
Vitamix Propel Series 510 Professional-Grade Blender, 3 Programs, Black
As an affiliate, we earn on qualifying purchases.
IBM’s Case for Private AI
Krishna’s argument connects two developments: improving open-weight models and enterprise interest in keeping sensitive data within controlled environments. Downloadable weights can allow organizations to choose their own hosting setup, but they do not automatically make a model free to use for every purpose. The source lists models with MIT licences as well as systems released under custom or commercial terms, making licence review part of deployment planning.
IBM has a commercial interest in the shift it is advocating. It sells on-premises infrastructure and expects demand to support sales of its Spyre cards. The source also cites The Register’s report that IBM mainframe revenue fell 20% in the quarter, 18 months into the z17 cycle. That figure does not establish why Krishna made the argument, but it is relevant to readers assessing the company’s incentives. His call also included a qualification: the frontier cloud remains the easier route for organizations lacking specialized engineering staff.
A separate enthusiast development in the source involves gpu-unlock, an open-source Linux module that reportedly raises a tensor-core speed limit in some GeForce cards. Its maintainers and forks report gains of up to 80% in matrix benchmarks and 14% to 36% in prompt processing, with little change in text generation. The source says the software exploits a boot-firmware flaw and that Nvidia’s GeForce licence bars circumvention of technical limits; Nvidia had not commented as of the article’s stated date. These reports do not establish a general enterprise hardware option or a safe, supported deployment path.
Performance and Cost Still Vary
It is not yet clear whether Kimi K3.1 Pro’s weights will be released by the promised November 4 date, under what licence they would ship, or what infrastructure the 2.8-trillion-parameter model would require. The source says no licence had been published. Its high score therefore should not be treated as proof that enterprises can currently download and deploy the model.
The benchmark comparison also does not settle how open models will perform on particular business tasks. Artificial Analysis provides a common index, but the source does not supply the task-level results, independent verification of IBM’s token-cost range or a full calculation of ownership costs. Hardware prices, usage rates, staffing, energy, cooling and networking can all change the financial result. Nvidia had not commented on the reported GeForce firmware workaround as of October 26, and its legal or technical consequences remain unsettled in the material provided.
Watch Model Releases and Deployment Costs
The next near-term test is whether Moonshot releases Kimi K3.1 Pro’s weights as promised by November 4 and publishes terms that permit enterprise use. Companies considering on-prem AI will also need to benchmark candidate models on their own workloads, check each model’s licence and estimate realistic utilization alongside staffing and facility costs.
Further earnings disclosures, product availability and independently measured deployment costs may clarify how much demand IBM’s pitch can generate. For now, Krishna’s argument is best read as a case for evaluating on-premises AI in selected, high-volume or sensitive applications — not as evidence that every enterprise can achieve the stated savings or performance.
Key Questions
What did IBM CEO Arvind Krishna say about open-weight AI?
On IBM’s October 21 earnings call, Krishna said open-weight models could cost 10% to 25% of frontier-model token costs while being “90% as effective.” The 90% figure was not tied to a named benchmark in the source material.
How close are the leading downloadable models to proprietary models?
In the Artificial Analysis Index data cited as read October 26, Xiaomi MiMo-V2.6-Pro scored 46.3, compared with 57.6 for Claude Opus 5.5 — about 80% of its score. The result reflects that index, not every task or deployment.
Does open-weight AI always cost less?
No. Costs depend on the model, provider, token rates, licence and hosting. On-premises use also adds hardware, staffing, cooling and networking expenses; buying servers is more attractive when they are used consistently.
What remains unknown about Kimi K3.1 Pro?
The source says Moonshot had promised to release the weights by November 4, but they were not yet available as of October 26. The licence and practical deployment requirements were also unclear.
Source: ThorstenMeyerAI.com
