🔍 Read the full analysis: Self-Run AI, Open Weights, And The On-Prem Bet At IBM 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
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 enterprises can run them on their own infrastructure. An Artificial Analysis comparison read on October 26, 2026, puts the strongest downloadable model at about 80% of the leading proprietary model; costs and operational needs also depend on model choice and how intensively hardware is used.
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 them an option for companies that want to run AI on their own systems. A benchmark snapshot checked October 26 puts the best downloadable model at 80% of the top proprietary model’s score, while the economics of buying hardware depend heavily on how much it is used.
Krishna said companies could begin with “two or three racks” and use open-weight models for routine work such as summarizing email, reviewing contracts and answering customer calls. He also framed private deployment as a data-control choice: banks and trading firms, he said, may want AI systems that keep underlying data and logs within their control. IBM sells infrastructure for on-premises AI, giving the company a commercial interest in that approach.
Artificial Analysis’s Intelligence Index, as read October 26, listed MiMo-V2.6-Pro at 46.3 as the highest-scoring downloadable model and Claude Opus 5.5 at 57.6 as the leading proprietary model. That makes the downloadable model’s score about 80% of the proprietary leader’s. The source’s historical comparison shows the share rising from 51% at the end of 2024, although the absolute point gap grew from 7.5 to 11.3 as both groups’ scores changed.
The cost comparison is not a single fixed saving. The source says hosted DeepSeek V4.1 Flash 1017 costs 3% to 6% of Claude Opus 5.5’s list price, while also citing OpenAI’s closed GPT-6 Luna at $0.10 per million input tokens and $0.50 per million output tokens. Those examples show that model pricing varies; they do not establish that self-hosting is always cheaper. The source estimates an eight-GPU H200 server at a list price starting at $172,643, against AWS rental for the same GPU class at $63.30 an hour.
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 Cost Depends on Workload
The decision facing enterprises is not simply whether an open model can match a proprietary model on a benchmark. It is whether a particular model is capable enough for a task, legally usable under its licence and affordable to operate at the company’s expected volume. A lower token price can matter for high-volume workloads, but a model that needs more hardware, engineering time or human review can change the calculation.
Buying equipment also shifts the cost profile. The source’s list-price arithmetic estimates that an H200 server could match rental costs after about 114 days at full utilization, but about 15 months at 25% utilization. Those estimates exclude staff, cooling and networking. They are not a total-cost-of-ownership comparison, and the result could vary with negotiated prices, financing, electricity and how much capacity sits idle.
IBM’s own remarks point to an operational constraint as well as a sales opportunity. Krishna said enterprises may lack “kernel engineers” and people responsible for the underlying network. That expertise can be needed to deploy and maintain AI systems at scale. On-premises operation may offer more control over data, but it places infrastructure and staffing responsibilities on the customer.
Products Worth Considering
Jackery Explorer 2000 v2 Portable Power Station with Two 200W Solar Panels
Logitech MX Keys S Wireless Keyboard, Graphite
Sony WH-1000XM6 Wireless Noise-Canceling Over-Ear Headphones, Blue, Bundle
As an affiliate, we earn on qualifying purchases.
A Narrowing but Uneven Model Gap
Krishna’s “90%” figure was an illustrative comparison on the earnings-call Q&A, not a result tied in the source to a named benchmark. The independent index provides one way to compare models across evaluations, but its score is not a direct measure of performance for every enterprise task. The source’s October snapshot shows a narrower relative gap than at the end of 2024, while also showing that proprietary models continued to improve.
One possible exception is Kimi K3.1 Pro, which the source lists at 55.1, or about 96% of Claude Opus 5.5’s score. But the model’s weights had not been released as of the source’s October 26 check; Moonshot had promised them “by Nov 4.” Until the weights are available, it is not a downloadable open-weight option. The source also says no licence had been published and notes the model’s 2.8-trillion-parameter scale, which would make deployment a substantial infrastructure undertaking.
Licensing differs across models. The source identifies MiMo-V2.6-Pro and DeepSeek V4.1 Flash 1017 as MIT-licensed, while other models have custom or commercial terms. A high index score alone does not establish that a model can be used for a company’s intended commercial purpose. IBM also claimed in prepared remarks that its llm-d project can run large models at five to 10 times lower cost; the source does not provide an independent measurement to verify that claim.
Costs, Licences and Scores
The figures do not settle whether on-premises AI will save a particular company money. The hardware comparison uses list prices and utilization assumptions, and excludes staffing, power, cooling and networking. The source does not provide a matched, independently verified comparison of the full operating cost for the same model and workload across self-hosted and hosted setups.
It is also unclear whether Krishna’s 90% performance estimate applies broadly beyond the routine tasks he named. Index rankings summarize a defined set of evaluations and may not predict results on a company’s own data or requirements. Kimi K3.1 Pro’s reported score cannot be treated as an available open-weight result unless its weights are released, and its licence and deployment requirements need clarification.
IBM’s cost claim for llm-d remains a company statement in the source material rather than an independently checked result. The source also reports that Nvidia had not commented on the GeForce firmware issue described below.
Model Releases and Deployment Tests
The next near-term test identified in the source is whether Moonshot releases Kimi K3.1 Pro’s weights by its stated November 4 target and publishes terms for their use. That could change the open-versus-proprietary benchmark comparison, though it would not by itself establish that the model is practical or economical for every organization to run.
For enterprise buyers, the next step is a workload-specific comparison: test candidate models on the tasks they will actually handle, review licence conditions, and estimate total costs at realistic utilization. IBM’s pitch will also be judged by whether customers can staff and operate deployments successfully, not just by token prices or benchmark rankings. The source provides no customer deployment data or subsequent IBM results that would settle those questions.
Key Questions
What did IBM’s CEO say about open-weight AI?
Arvind Krishna said open-weight models could cost 10% to 25% of frontier-model token costs while being “90% as effective.” The source presents this as his earnings-call estimate, not a result from a named benchmark.
How close are downloadable models to the leading proprietary model?
In the Artificial Analysis snapshot read October 26, 2026, MiMo-V2.6-Pro scored 46.3 against Claude Opus 5.5’s 57.6, or about 80% of the leading proprietary score on that index.
Does running a model on-premises guarantee lower costs?
No. Hardware purchase costs may compare favorably with rental at high utilization, but the source’s estimate excludes staff, cooling and networking. Actual costs also depend on workload, power use and the chosen model.
Is Kimi K3.1 Pro available as an open-weight model?
Not according to the source’s October 26 check. Moonshot had promised the weights by November 4, but they had not yet shipped, and the source said a licence had not been published.
Why might a company choose to run AI on its own infrastructure?
Krishna cited control over data and logs, as well as the ability to run models on-premises or in private clouds. That approach also requires companies to provide and maintain the necessary hardware and technical staff.
Source: ThorstenMeyerAI.com
