🔍 Read the full analysis: What Businesses Should Know About IBM’s On-Prem AI Bet on ThorstenMeyerAI.com
Get the latest gadgets 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 argued on the company’s Oct. 21 earnings call that open-weight AI can deliver about 90% of frontier-model effectiveness at 10% to 25% of the token cost, and that large enterprises can start with two or three racks. An independent benchmark cited in the source puts the best downloadable model at 80% of the leading proprietary model as of Oct. 26, 2026; the business case also depends on utilization, licensing, staffing and data-control needs.
IBM CEO Arvind Krishna said on the company’s Oct. 21 earnings call that open-weight AI models can be about 90% as effective as leading frontier models at a fraction of their token cost, making on-premises or private-cloud deployment worth considering for some enterprises. But an independent benchmark cited in the source puts the best downloadable model at 80% of the top proprietary model as of Oct. 26, while the economics of buying hardware depend on keeping it busy and having the staff to run it.
In the call’s question-and-answer session, Krishna said open-weight models cost “25% to 10% of the token cost” of frontier models while being “90% as effective.” He said a large enterprise could begin with “two or three racks.” He pointed to routine tasks such as summarizing emails, reviewing contracts and answering customer calls as work that may not need the most capable available model.
Krishna also framed private deployment as a data-control and sovereignty choice for businesses such as banks and trading firms. Open weights, he said, provide rights to run a model on a company’s own infrastructure or in a private cloud. That does not mean every model has the same usage rights: the source says some carry custom or commercial licences, so companies need to review the terms rather than rely on a benchmark score alone.
The source’s comparison of the Artificial Analysis Intelligence Index, read Oct. 26, lists Xiaomi’s downloadable MiMo-V2.6-Pro at 46.3 and proprietary Claude Opus 5.5 at 57.6. That is about 80% of the leader on this index. Moonshot’s Kimi K3.1 Pro scores 55.1, but the source says its weights were promised by Nov. 4 and had not yet shipped. Its score cannot yet be treated as an available open-weight option.
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
For business leaders, the proposal is not simply “open models are cheaper.” It is a capital investment versus usage-cost decision, with privacy and control potentially part of the return. The source gives a list price of $172,643 for an eight-GPU H200 server and an AWS rental rate of $63.30 an hour for the same GPU class. Its calculation estimates that a purchase pays off in about 114 days at full use, but about 15 months at 25% utilization, before staffing, cooling and networking costs.
Those figures illustrate why an on-premises deployment can look attractive for a business with steady, high-volume workloads, but less so for uneven demand. The calculation is not a complete cost-of-ownership comparison: it excludes several operating expenses and uses listed prices. Actual terms, discounts, workload patterns and deployment requirements could change the result.
Model capability is another constraint. The benchmark comparison shows progress in downloadable models, but the leading available model still trails the top proprietary one on the cited index. A company should test its own tasks, latency needs and error tolerance rather than assume a general score translates directly into business performance. As Krishna acknowledged on the call, organizations may also lack kernel and networking engineers; the source describes the frontier cloud as the easier option, albeit one he called four times more expensive.
Products Worth Considering
Jackery HomePower 1000 Plus v2 Portable Power Station
DJI Osmo Pocket 3 Filter Bundle – Vlogging Camera with 1 CMOS Sensor and 4K/120fps
Logitech MX Master 4 for Mac, Ergonomic Bluetooth Mouse with Haptic Feedback, Ultra-Fast Scrolling, USB-C Charging, for MacOS, iPadOS, Windo
As an affiliate, we earn on qualifying purchases.
How the Benchmark Gap Has Shifted
The source compares Artificial Analysis scores at several dates. At the end of 2024, the best open-weight score was 7.7 against a proprietary leader at 15.2, or 51%. At the end of 2025, the comparison was 22.4 against 30.4, or 74%. The Oct. 26, 2026 comparison was 46.3 against 57.6, or 80%. These are scores on one index, not a universal measure of how well a model performs on every company’s tasks.
The source notes that the absolute score gap in that comparison was 7.5 points at the end of 2024 and 11.3 points on Oct. 26, 2026, even as the open model’s score rose relative to the leader. Proprietary models also improved during that period. Kimi K3.1 Pro would put the open-to-proprietary ratio near 96% if its weights are released, but the source says they were not yet available and that a model of its reported size would require substantial data-centre resources.
There is also a business interest behind IBM’s argument. IBM sells infrastructure, and the source says Krishna expects on-premises demand to support sales of the company’s Spyre cards. IBM separately said its llm-d project can run large models at five to 10 times lower cost; that is an IBM claim, not an independently verified result in the material provided. A cost comparison also varies by provider: the source says hosted DeepSeek V4.1 Flash 1017 costs 3% to 6% of Claude Opus 5.5’s listed price, while closed GPT-6 Luna is listed at $0.10 per million input tokens and $0.50 per million output tokens.
Open Weights and Costs Still Need Checking
The source’s 90% effectiveness figure is Krishna’s characterization, not the result of a named benchmark in his remarks. The cited Artificial Analysis comparison provides one independent reference point, but it does not establish how a model will perform on a particular company’s workloads. Kimi K3.1 Pro’s status also remained conditional in the source: its weights had been promised for Nov. 4, but had not shipped by Oct. 26.
The full cost of owning and operating AI hardware is not established by the listed server and rental prices. The payback estimate excludes staff, cooling and networking, and does not account for negotiated prices or changing utilization. The source also does not provide an independently checked measurement for IBM’s llm-d cost claim. Licensing terms, deployment costs and real-world performance therefore remain key questions for any prospective buyer.
What Enterprise Buyers Should Verify
Companies weighing deployment options will need to compare a representative workload on candidate models, check whether model weights are actually available, and review each licence for their intended commercial use. They should also estimate utilization across the year and include hardware, energy, networking, maintenance and staff in the cost model—not just token prices or server purchase costs.
The next stated milestone in the source is Moonshot’s promised Nov. 4 release of Kimi K3.1 Pro’s weights; whether those weights arrive, and under what licence, was not confirmed as of Oct. 26. More broadly, buyers will need to see how model scores, provider prices and hardware options change, and whether IBM’s cost claims can be assessed against comparable deployments.
Key Questions
What did IBM’s CEO say about open-weight AI costs?
Arvind Krishna said on IBM’s Oct. 21, 2026 earnings call that open-weight models cost 10% to 25% as much as frontier models on a token basis while being “90% as effective.” That is his characterization, not a result from a named benchmark in the remarks.
How close are the best available open-weight models to proprietary leaders?
In the Artificial Analysis comparison cited by the source and read on Oct. 26, 2026, MiMo-V2.6-Pro scored 46.3, compared with Claude Opus 5.5 at 57.6—about 80% of the proprietary leader’s score on that index.
Does buying an AI server always cost less than renting GPUs?
No. The source estimates that an eight-GPU H200 server listed at $172,643 could reach purchase-price payback in about 114 days at full use, compared with about 15 months at 25% use. The estimate excludes staffing, cooling and networking, so it is not a complete ownership-cost comparison.
What should a company check before using an open-weight model?
It should test performance on its own workloads, confirm that the weights are available, review the model’s licence for commercial use, and account for infrastructure, utilization and staff requirements. A high benchmark score does not by itself establish suitability or legal permission for a particular use.
Source: ThorstenMeyerAI.com
