🔍 Read the full analysis: How IBM’s CEO Thinks About Running AI In-House on ThorstenMeyerAI.com
Get hardware and tech 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 told the company’s Oct. 21 earnings call that open-weight models can deliver about 90% of frontier-model performance at 10% to 25% of the token cost, and said enterprises may start with two or three racks. An independent Artificial Analysis comparison puts the strongest downloadable model at 80% of its top proprietary model as of Oct. 26; the cost case also depends on usage, hardware, staffing and the model licence.
IBM CEO Arvind Krishna told investors on the company’s Oct. 21 earnings call that open-weight AI models can cost 10% to 25% as much as leading proprietary models while being “90% as effective,” making it practical for some enterprises to run AI on their own infrastructure. An independent model scoreboard puts the best downloadable model at 80% of the leading proprietary model as of Oct. 26, and the economics of buying hardware depend on how heavily it is used.
Krishna said a large enterprise could begin with two or three racks of equipment. He described open-weight models as offering the “permission and the IP rights” to run them on premises or in private clouds. His examples of tasks that may not require the most capable model included summarizing email, reviewing contracts and answering customer calls. He also cited data control as a reason banks and traders might prefer private deployments, saying they want to keep underlying data and logs from being taken by others.
The cost and performance figures are Krishna’s characterization, not a result from a named benchmark in the earnings call. Artificial Analysis’s index, which applies the same ten evaluations across models, listed MiMo-V2.6-Pro at 46.3 as the top model with downloadable weights and Claude Opus 5.5 at 57.6 as the leading proprietary model. That puts the downloadable model at about 80% of the leader on that index. IBM also said its llm-d project can run large models at five to 10 times lower cost; that is IBM’s claim, and the supplied material provides no independent measurement of it.
Krishna acknowledged that operating models in-house requires specialist skills. He pointed to a shortage of “kernel engineers” and people who handle underlying networks, saying the gap remained except at the most sophisticated enterprises. The source article also notes IBM has a commercial interest in on-premises AI hardware, including its Spyre cards, so its cost and deployment arguments come from a company that may benefit if customers buy more infrastructure.
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 of Owning AI Infrastructure
Krishna’s argument puts a practical question in front of enterprise buyers: whether the lower cost per token of an open-weight model can offset the expense and work of running it. A company that owns its equipment has more direct control over where data and logs are processed, but it also takes responsibility for hardware, staffing, cooling and networking. The performance trade-off matters too: the index comparison shows a substantial improvement in open-model scores over time, but not parity with the top proprietary model in the current snapshot.
The utilization calculation in the source illustrates why a token-price comparison alone is incomplete. An eight-GPU H200 server is listed from $172,643, while renting the same GPU class on AWS is listed at $63.30 an hour. The source estimates that buying could match rental costs after about 114 days at full use, but about 15 months at 25% use, before staffing, cooling and networking. Those are estimates based on the stated prices and utilization assumptions, not a guarantee of savings. For companies with uneven demand, rented capacity may compare differently than it does for a system running near full load.
Products Worth Considering
Apple MacBook Pro 16-inch Laptop with M5 Pro, 18-Core CPU, 20-Core GPU, 48GB Memory, 2TB SSD, Space Black
Apple MacBook Pro 16.2-inch with M5 Pro, 18-Core CPU, 20-Core GPU, 48GB Memory, 1TB SSD, Nano-Texture Display, Space Black
Apple MacBook Pro 14.2-inch with M5 Max, 18-Core CPU, 20-Core GPU, 48GB Memory, 2TB SSD, Space Black
As an affiliate, we earn on qualifying purchases.
Open Models Close the Benchmark Gap
The Artificial Analysis figures cited in the source show the leading open-weight model moving from 51% of the top proprietary score at the end of 2024 to 74% at the end of 2025, then 68% at the end of June 2026 and 80% on Oct. 26. The comparison uses the index’s current version and the model entries identified in its open-versus-proprietary chart. Scores across changing model generations do not establish that any one open model is suitable for a particular company’s tasks.
One possible exception to the 80% snapshot was Kimi K3.1 Pro, which scored 55.1, or about 96% of Claude Opus 5.5’s 57.6. But the source says Moonshot had promised to release the weights by Nov. 4; they had not yet shipped as of Oct. 26. Until weights are available, the model does not meet the source’s definition of open-weight. Licensing also differs: the source identifies MIT licences for MiMo-V2.6-Pro and DeepSeek V4.1 Flash 1017, while other models have custom or commercial terms. A model’s score alone does not establish what a business is allowed to do with it.
Costs and Model Access Still Vary
The index score is one measure of model performance, not proof that a model will perform equally well on a particular company’s tasks. The source does not provide a task-by-task comparison supporting Krishna’s “90%” figure. It also does not establish that the token-cost range applies across all providers, model sizes or workloads. Hosted prices, list prices and self-hosting expenses are different measures, and actual costs will depend on usage and operational requirements.
Several details remain unsettled. Kimi K3.1 Pro’s weights had not been released by the date of the cited comparison, and the source says no licence had been published. The purchase-versus-rental calculation excludes staff, cooling and networking, so it is not a full cost-of-ownership analysis. IBM’s own commercial interest in on-premises hardware is relevant when evaluating its claims, while the supplied material does not include an independent test of IBM’s llm-d cost statement. Nvidia had not commented on a separate report in the source concerning an open-source module that modifies performance limits on some GeForce cards; that issue is distinct from IBM’s enterprise deployment pitch.
Watch Weights, Utilization and Licences
The next concrete check in the source’s timeline is whether Moonshot releases Kimi K3.1 Pro’s weights by its promised Nov. 4 date, and what licence accompanies them. For enterprise buyers, the relevant next steps are to compare models on their own workloads, confirm licence terms and estimate total infrastructure costs using realistic utilization rather than token prices alone. IBM’s pitch is likely to be judged against those deployment results, but the supplied material identifies no announced IBM customer test or independent validation of its llm-d savings claim.
Key Questions
What did IBM CEO Arvind Krishna say about open-weight AI?
On IBM’s Oct. 21 earnings call, Krishna said open-weight models can cost 10% to 25% as much as frontier models while being “90% as effective.” Those figures are his characterization, not a comparison tied to a named benchmark in the call.
How close are the leading downloadable models to proprietary models?
In the Artificial Analysis comparison read on Oct. 26, 2026, MiMo-V2.6-Pro scored 46.3, against 57.6 for Claude Opus 5.5. That makes the downloadable model about 80% of the proprietary leader on that index.
Does running AI in-house always cost less?
No. The economics depend on equipment prices, workload and utilization, as well as staffing, cooling and networking. The source’s estimate says an eight-GPU H200 server could match rental costs in about 114 days at full use or about 15 months at 25% use, before several operating expenses.
Why would a company want to run an AI model on its own systems?
Krishna cited control over data and logs, as well as the ability to run open-weight models on company infrastructure or private clouds. An in-house deployment also requires the company to manage hardware and specialist operations.
Are all open-weight models licensed for the same uses?
No. The source identifies some models with MIT licences and others with custom or commercial terms. Businesses need to check the specific model licence; a high benchmark score alone does not establish permitted uses.
Source: ThorstenMeyerAI.com
