🔍 Read the full analysis: IBM’s CEO Connects Self-Run AI With On-Premises Scale on ThorstenMeyerAI.com
Get everyday essentials delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
IBM CEO Arvind Krishna said open-weight AI models can deliver about 90% of leading-model performance at 10% to 25% of the token cost, and that large enterprises could start with two or three racks. A comparison cited by ThorstenMeyerAI.com puts the best downloadable model at 80% of the top proprietary model on its index as of October 26, while on-prem economics depend on usage, infrastructure and staffing.
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 as frontier models while being about 90% as effective, making it possible for large enterprises to run AI on their own infrastructure. A comparison published by ThorstenMeyerAI.com puts the leading downloadable model at 80% of the top proprietary model on the Artificial Analysis Intelligence Index as of October 26, leaving a gap between Krishna’s estimate and that snapshot.
Krishna said enterprises could begin with two or three racks of equipment. He framed the choice as a trade-off between model performance and cost, arguing that many routine tasks—such as summarizing emails, examining contracts and answering customer calls—may not require the most capable model. He also said open-weight models give organizations permission and intellectual-property rights to run them on-premises or in private clouds, depending on the model’s licence.
The cited index comparison lists MiMo-V2.6-Pro as the highest-scoring downloadable model at 46.3, against Claude Opus 5.5 at 57.6. That makes the downloadable model about 80% of the proprietary leader on this index, not 90%. The source’s historical figures show the open-to-proprietary score ratio rising from 51% at the end of 2024 to 74% at the end of 2025, though the raw score gap increased from 7.5 points to 11.3 points over the period.
A possible exception depends on a future release: Kimi K3.1 Pro scored 55.1, or 96% of the index leader, but its weights had not been released as of October 26. Moonshot had promised them by November 4, according to the source. Until weights are available, the model does not qualify as downloadable under the comparison’s distinction. Krishna also acknowledged a practical constraint: enterprises need people with expertise in kernels and networking, skills he said may be lacking outside the most sophisticated organizations.
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 Behind On-Prem AI
The argument matters because companies weighing AI deployment face more than a model-price comparison. Self-hosting can offer control over data and logs, an important consideration for organizations such as banks and trading firms, Krishna said. It can also shift spending from per-token cloud charges to hardware, facilities and staff. Those costs make the economics sensitive to how often expensive equipment is used.
ThorstenMeyerAI.com compared an eight-GPU H200 server, listed from $172,643, with renting the same GPU class on AWS at $63.30 per hour. Using those list prices, the source estimates a purchase could match rental costs in about 114 days at full utilization, but take about 15 months at 25% utilization. The estimate excludes staff, cooling and networking, so it is not a complete cost comparison. A lightly used system may not deliver the savings implied by per-token price comparisons.
The choice also has implications for vendors. IBM sells infrastructure that could benefit if more companies bring AI workloads into their own data centers. That commercial interest does not invalidate Krishna’s argument, but it is relevant when weighing his estimate. The source also reports that IBM expected its llm-d project to run large models at five to ten times lower cost; that figure is an IBM claim, not an independently verified measurement in the report.
Products Worth Considering
Jackery Explorer 300 Portable Power Station, 292Wh
Jackery Explorer 2000 v2 Portable Power Station with Two 200W Solar Panels
LG UltraGear 27-Inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDMI 2.1
As an affiliate, we earn on qualifying purchases.
From Model Scores to Enterprise Racks
Krishna’s comparison of model capability was presented as a broad illustration, not as the result of a named benchmark. ThorstenMeyerAI.com tested the claim against the Artificial Analysis Intelligence Index, which evaluates models across ten assessments and identifies whether their weights can be downloaded. The index offers one way to compare models, but its score is not a direct measure of performance on every company’s workload.
The source says open models have narrowed their relative score gap with proprietary leaders since late 2024. That movement does not mean every open model is inexpensive to run or suitable for commercial use. Licensing terms vary: MiMo-V2.6-Pro and DeepSeek V4.1 Flash 1017 are listed under MIT licences, while other models use custom or commercial licences. Organizations need to check both the licence and the hardware requirements before treating a model as a deployable alternative.
Price comparisons also depend on how a model is accessed. The source says hosted DeepSeek V4.1 Flash 1017 costs 3% to 6% of Claude Opus 5.5’s list price, while OpenAI’s closed GPT-6 Luna is listed at $0.10 per million input tokens and $0.50 per million output tokens. Those examples show why “open” does not automatically mean cheaper than every closed option; prices, workload and deployment method all affect the result.
Performance, Licensing and Utilization Gaps
The index figures do not establish how open and proprietary models compare on a particular company’s tasks, nor do they confirm Krishna’s 90% estimate across workloads. The comparison is a single index snapshot, and models, index versions and rankings can change. Kimi K3.1 Pro’s reported score could narrow the gap if its weights are released, but the source says the weights and licence were not available as of October 26.
The total cost of an on-prem deployment is also unresolved by the examples provided. The server-versus-cloud calculation excludes staff, cooling and networking, and actual utilization may differ substantially from the 100% and 25% scenarios. The source does not provide a full IBM-specific deployment cost or independently verify the cost reductions claimed for llm-d.
Krishna’s statements come from an earnings call, and IBM has a business interest in infrastructure sales. The available material does not establish how many enterprises will adopt on-prem AI, whether their savings will match the cited estimates, or how the economics compare across different workloads and licences.
The Next Tests for On-Prem AI
The next near-term check identified by the source is whether Moonshot releases Kimi K3.1 Pro’s weights by its promised November 4 date, and what licence accompanies them. If they arrive, the model’s score can be assessed as a downloadable option rather than a hosted one, although its reported 2.8 trillion parameters would still make deployment a substantial data-center undertaking.
For enterprise buyers, the practical next step is to compare candidate models on their own tasks and calculate the full cost of deployment, including hardware utilization, staffing, networking, power and licensing. IBM’s earnings-call claims, the public index and rental prices each answer different parts of that decision; none alone settles whether a company should buy racks or use cloud services.
Key Questions
What did IBM CEO Arvind Krishna say about open-weight AI?
Krishna said open-weight models can cost 10% to 25% of frontier-model token costs while being about 90% as effective. That was his estimate on IBM’s October 21 earnings call, not a result tied to a named benchmark.
Does the cited benchmark show open models are 90% as capable?
Not in the October 26 snapshot described by ThorstenMeyerAI.com. The top downloadable model, MiMo-V2.6-Pro, scored 46.3 against 57.6 for Claude Opus 5.5, or about 80% of the proprietary leader on that index.
Why might a company run AI on its own infrastructure?
Krishna cited control over data and logs, as well as potential cost savings, as reasons to use on-premises or private-cloud systems. The financial case depends on hardware utilization and the costs of staff, facilities and networking.
Does buying an AI server always cost less than renting GPUs?
No. The source’s comparison estimates that an eight-GPU H200 server could match rental costs in about 114 days at full load, but about 15 months at 25% utilization. Those estimates exclude staff, cooling and networking.
Source: ThorstenMeyerAI.com
