🔍 Read the full analysis: IBM’s CEO On The Case For Running AI Yourself 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 argued on the company’s October 21 earnings call that open-weight AI models can deliver about 90% of frontier-model performance at a fraction of the token cost, and said enterprises can run them on their own infrastructure. An independent benchmark cited in the source puts the strongest currently downloadable model at about 80% of the leading proprietary model; the cost and practicality of on-premises deployment depend on workload, utilization, staffing and licensing.
IBM CEO Arvind Krishna told investors on the company’s October 21 earnings call that open-weight AI models can cost a fraction of frontier models while delivering roughly 90% of their effectiveness, making them worth running on company-owned infrastructure. That is Krishna’s characterization, not a result from a named benchmark: an independent model index cited in the source puts its best currently downloadable model at 80% of the leading proprietary model.
Krishna said open-weight models could cost 10% to 25% of the token cost of frontier models. He argued that many routine business tasks—such as summarizing email, reviewing contracts and answering customer calls—may not require a top-end proprietary model. For larger organizations, he said, an initial deployment could involve two or three racks of equipment.
The argument is not only about cost. Krishna said banks and trading firms may want to run AI in a way that keeps data and logs under their control. Open weights, he said, give organizations permission and intellectual-property rights to run models on their own systems or private clouds. Those rights vary by model, however, and a downloadable model is not automatically unrestricted: companies need to check the applicable licence.
The source’s comparison of Artificial Analysis’s model index, read October 26, puts Xiaomi’s downloadable MiMo-V2.6-Pro at 46.3 and proprietary Claude Opus 5.5 at 57.6. That makes the open model about 80% of the leader on that index. Moonshot’s Kimi K3.1 Pro scores 55.1, or about 96% of the leader, but the source says its weights had not yet shipped and were promised by November 4. Until they are available, it is not an open-weight option for companies to download and operate.
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 pitch could matter to companies weighing cloud spending against capital investment, especially those processing large volumes of routine requests or handling sensitive information. Running a model locally can give a company more control over where data is processed, while avoiding some charges tied to hosted-model usage. But lower token prices alone do not settle the economics.
The source’s hardware example illustrates the trade-off. An eight-GPU H200 server is listed at $172,643, while renting the same GPU class on AWS costs $63.30 an hour. The source estimates that purchase price would be recouped in about 114 days at full utilization, compared with about 15 months at 25% utilization. Those estimates exclude staff, cooling and networking, so they are not a full cost comparison. A lightly used server can take much longer to justify than a heavily used one.
There is also an operational hurdle. Krishna himself identified a shortage of kernel engineers and networking expertise as a constraint, saying that the gap remains for enterprises outside the most sophisticated group. Companies considering on-premises systems need to account for the people and infrastructure required to deploy, secure and maintain them, not just the model and GPU price.
Products Worth Considering
LG UltraGear 27-Inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDMI 2.1
Sonos Era 100 Wireless Smart Speaker, Black
Amazon Kindle Paperwhite (12th Generation, 2024), 16 GB, 7-inch Display, Jade
As an affiliate, we earn on qualifying purchases.
What Krishna Said on the Call
The comparison that drew attention came during the question-and-answer portion of IBM’s earnings call, rather than its prepared remarks. Krishna described the performance gap as narrowing over time: he recalled a broad benchmark in late 2024 where a leading model might score 70 while an open model scored 20 or 30, followed by a 70-to-50 comparison in late 2025. He then posed the choice as paying a tenth as much for something “90% as good” or paying ten times more.
Those figures were illustrative examples, not results Krishna tied to a specific evaluation. The source checks them against Artificial Analysis, which applies ten evaluations across its index. In that comparison, the share of the leading proprietary score held by the best downloadable model rose from 51% at the end of 2024 to 80% on October 26, 2026. The source also notes that the raw point gap increased, from 7.5 points to 11.3, because the proprietary leader’s score rose as well.
IBM has a commercial interest in the shift it is advocating: it sells on-premises hardware, and Krishna said he expects demand for local AI to support sales of IBM’s Spyre cards. The source also reports that IBM claimed its llm-d project can run large models at five to ten times lower cost, but says it could not independently check that figure. These points are reasons to distinguish IBM’s business case from independently measured model performance.
Performance, Licensing and Payback
The benchmark comparison does not show that open models are 90% as effective across every task. It measures performance on one index, and the source does not identify Krishna’s broad benchmark or the methods behind his 70, 50 and 90 figures. Scores also do not establish whether a model will meet a particular company’s accuracy, speed, privacy or reliability requirements.
It is also unclear whether Kimi K3.1 Pro’s promised weights will arrive on schedule or what licence will govern them. The source says no licence had been published at the time of its check. Model size can be another barrier: it describes Kimi as a 2.8-trillion-parameter model, which it characterizes as a data-centre-scale project. Pricing comparisons also depend on provider, workload and usage; an open model is not automatically cheaper to host than every closed alternative.
The hardware payback figures are estimates based on listed prices and assumed utilization. They do not include operating costs or establish what any particular enterprise will pay. IBM’s claim about llm-d’s cost savings was not independently verified in the source, and Nvidia had not commented on the reported GeForce modification described there.
Evidence as Models and Costs Change
One near-term test is whether Moonshot releases Kimi K3.1 Pro’s weights as promised by November 4, and under what licence. If they become available, the model’s score would place it closer to the proprietary leader on the cited index, but availability and licensing are needed before organizations can assess it as a deployable open-weight option.
For enterprise buyers, the next step is a workload-specific comparison: test relevant models on the company’s own tasks, review licence terms, and calculate infrastructure costs at realistic utilization. The source’s figures indicate why a cloud-versus-purchase decision can change with usage, but they do not supply a universal break-even point.
Artificial Analysis’s scores and model prices can change as providers release new systems and update rates. The reported gap between downloadable and proprietary models should therefore be treated as a snapshot from October 26, 2026, not a lasting ranking. IBM’s argument will be tested by those changing benchmarks and by whether enterprises can staff and operate the systems economically.
Key Questions
What did IBM CEO Arvind Krishna argue?
Krishna said open-weight models can cost 10% to 25% as much per token as frontier models while being about 90% as effective. He presented that as a case for some enterprises to run models on their own infrastructure.
Does the cited benchmark show open models are 90% as good?
Not for the best model whose weights were available in the source’s October 26 comparison. Artificial Analysis listed MiMo-V2.6-Pro at 46.3, versus 57.6 for Claude Opus 5.5—about 80% of the proprietary model’s score on that index.
Why might a company run AI on its own servers?
Potential reasons include control over where data and logs are handled, and the possibility of lower costs for heavily used workloads. Local deployment also brings hardware, staffing, cooling, networking and maintenance costs.
Are open-weight models always cheaper?
No. Total cost depends on the model’s licence, hardware, utilization and operating needs, as well as the price of hosted alternatives. The source’s server-payback estimates exclude several operating costs and are not a universal comparison.
Source: ThorstenMeyerAI.com
