🔍 Read the full analysis: How To Think About Running AI Yourself, According To IBM’s CEO 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
On IBM’s October 21 earnings call, CEO Arvind Krishna argued that open-weight AI models can deliver about 90% of frontier-model performance at 10% to 25% of the token cost, and that enterprises can begin with two or three racks. An independent benchmark comparison puts the best currently downloadable model at about 80% of the leading proprietary model; actual savings from on-premises deployment depend on hardware use, staffing, power and networking.
IBM CEO Arvind Krishna told investors 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 self-hosting worth considering for some enterprise workloads. An independent benchmark comparison available October 26 puts the best downloadable model at 80% of the leading proprietary model, while the cost of running AI on company-owned servers depends on how heavily those systems are used.
Krishna said enterprises could begin with two or three racks and use open-weight models for routine work such as summarizing email, reviewing contracts, answering customer calls and drafting messages. He argued that organizations can run such models on their own premises or private clouds, retaining more control over data and logs. The figures and advantages were part of Krishna’s case on an earnings call; IBM sells infrastructure that could benefit if customers build on-premises AI capacity.
The independent comparison cited in the source uses the Artificial Analysis Intelligence Index, which applies the same ten evaluations across models. On October 26, the leading downloadable model listed was Xiaomi’s MiMo-V2.6-Pro at 46.3 points, compared with 57.6 for proprietary Claude Opus 5.5. That makes the open model about 80% of the leader’s score on this index, not 90%. The source’s historical figures show the open-to-proprietary ratio rising from 51% at the end of 2024 to 80% in October 2026; they do not establish that every open model performs at that level or that the index represents every business task.
A potentially closer comparison is not yet available as an open model: Moonshot’s Kimi K3.1 Pro scored 55.1, about 96% of Claude Opus 5.5, but the source says Moonshot had promised to release its weights by November 4. Until those weights are actually published, it remains closed for purposes of the comparison. The source also notes that licensing varies: downloadable weights do not automatically mean unrestricted commercial use, so organizations need to review each model’s terms.
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 workload
For businesses running many predictable AI requests, lower per-token costs could make hosting a model locally financially attractive. It can also appeal to organizations with data-control or sovereignty requirements, including financial institutions, though self-hosting does not by itself settle every security, privacy or regulatory question. The practical choice is not simply “cheap open model” versus “expensive frontier model”: performance needs, licensing, data handling and operating costs all matter.
Hardware prices illustrate why utilization is central. The source lists an eight-GPU H200 server at $172,643 and AWS rental of the same GPU class at $63.30 per hour. On those figures, purchase cost alone would equal roughly 114 days of rental at full use, or about 15 months at 25% use. Those are simplified comparisons, not total-cost estimates: they exclude staff, cooling and networking. A company that cannot keep its equipment busy may not save money by owning it.
Krishna also identified a skills constraint: enterprises need people who understand model kernels and the underlying network. The source reports that he described this gap as persisting outside the most sophisticated enterprises. That qualification matters because buying hardware is only one part of operating an AI service reliably.
Products Worth Considering
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
DJI Osmo Mobile 8 Standard Combo Phone Gimbal Stabilizer
As an affiliate, we earn on qualifying purchases.
What Krishna said on the call
Krishna’s widely cited cost and performance comparison came during the question-and-answer portion of IBM’s Q3 earnings call, rather than the prepared remarks, according to the source. He described open-model performance as improving over time, using illustrative comparisons: a top model scoring 70 against an open model scoring 20 or 30 in late 2024, then 70 against 50 in late 2025. He framed the decision as whether to accept lower performance for a lower price on tasks that do not require the strongest available model.
The benchmark figures offer a separate way to check that broad argument, but not a direct verification of Krishna’s examples: the source says his scores were not tied to a named benchmark. Its index comparison shows open models gaining relative ground, while the absolute point gap in the listed comparison widened from 7.5 points at the end of 2024 to 11.3 points in October 2026. The two measures answer different questions: relative share of the leading score and the raw difference between scores.
Model access and price also differ by provider. The source says DeepSeek’s hosted V4.1 Flash 1017 costs 3% to 6% of Claude Opus 5.5’s list price, while OpenAI’s closed GPT-6 Luna lists at $0.10 per million input tokens and $0.50 per million output tokens. Those are provider list-price comparisons, not a like-for-like estimate of total enterprise costs. IBM also said in prepared remarks that its llm-d project can run large models at five to ten times lower cost; that is IBM’s claim, and the source says it could not independently verify the measurement.
Costs and model access remain unsettled
The source’s comparisons do not establish how much a particular company would save. Server purchase price and cloud rental rates omit important operating expenses, while the result depends on utilization, staffing, power, cooling and networking. Nor does a benchmark score show how well a model will perform on a specific company’s tasks, data or required accuracy level.
It is also unclear whether Moonshot will release Kimi K3.1 Pro’s weights on the promised timetable, and what licence terms would accompany them. A high score does not make a model open-weight if the weights are not available. Separately, the exact basis for IBM’s claimed five-to-tenfold cost reduction with llm-d is not independently assessed in the source. IBM’s financial interest in infrastructure sales should be kept in view when evaluating its recommendations.
Watch for weights and deployment data
The next concrete point to watch is whether Moonshot publishes Kimi K3.1 Pro’s weights by its stated November 4 target, along with a licence that permits the intended uses. If released, its status and performance can be compared with models whose weights are already available. Enterprises weighing deployment can also test candidate models on their own workloads and calculate total costs at realistic utilization levels rather than relying on token prices or hardware list prices alone.
Further evidence would be needed to judge IBM’s cost claims in practice, including comparable measurements that state the workloads, model quality, hardware and operating costs used. Until then, Krishna’s case is a strategic argument supported by broad performance and price comparisons, not a guarantee that self-hosting will be cheaper or suitable for every organization.
Key Questions
What did IBM CEO Arvind Krishna recommend?
Krishna argued that businesses should consider open-weight models on their own infrastructure or private clouds for tasks that do not require a frontier model. He said a large enterprise could begin with two or three racks, while acknowledging a need for specialist engineering and networking skills.
Do open-weight models currently match the leading proprietary models?
Not on the comparison cited here. As of October 26, 2026, Artificial Analysis listed Xiaomi’s MiMo-V2.6-Pro at 46.3, around 80% of Claude Opus 5.5’s 57.6 on that index. The result applies to that benchmark and those models, not every task.
Is running AI on a company’s own servers always cheaper?
No. The purchase-versus-rental calculation changes with how much the hardware is used and excludes costs such as staff, cooling and networking. Lower token prices do not alone determine the total cost of an on-premises system.
Does downloadable mean a model is free for commercial use?
No. Models have different licences, and some impose conditions. Organizations should check the specific model licence and confirm that it permits their planned commercial and deployment use.
Source: ThorstenMeyerAI.com
