🔍 Read the full analysis: How Businesses Can Weigh An 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 told an October 21 earnings call that open-weight AI models can be 90% as effective as frontier systems at 10% to 25% of their token cost, making on-prem deployment worth considering for some enterprises. An Artificial Analysis comparison puts the leading downloadable model at 80% of the proprietary leader on October 26, while hardware economics depend heavily on utilization, staffing and other operating costs.
IBM CEO Arvind Krishna said on October 21 that open-weight AI models can be 90% as effective at 10% to 25% of frontier-model token costs, renewing the case for businesses to run AI on their own infrastructure. But an Artificial Analysis comparison read October 26 put the best downloadable model at 80% of the leading proprietary model, while the financial case for buying servers depends on how steadily a company can use them.
Krishna made the cost comparison during the Q&A on IBM’s third-quarter earnings call. He said open-weight models could cost “25% to 10% of the token cost of the frontier models” while being “90% as effective.” He also said a large enterprise could begin with two or three racks. The examples he gave included summarizing email, examining contracts and answering customer calls—tasks that may not require a frontier system.
The comparison needs a qualification: Krishna’s score examples were illustrations, not results from a named benchmark. Artificial Analysis’s Intelligence Index, which applies the same ten evaluations across models, listed MiMo-V2.6-Pro at 46.3 as the highest-scoring model with downloadable weights on October 26. Its leading proprietary model, Claude Opus 5.5, scored 57.6, putting the downloadable model at about 80% of the leader on that index.
Hardware prices show why a low token price alone does not settle the decision. An eight-GPU H200 server lists from $172,643, while renting the same GPU class on AWS is listed at $63.30 an hour. A simple comparison suggests a purchase cost equal to about 114 days of rental at full utilization, or roughly 15 months at 25% utilization. Those estimates exclude staff, cooling and networking, and are not a complete cost-of-ownership calculation.
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 Model Use
For enterprises with high, steady workloads, owning hardware may offer control over infrastructure and potentially lower costs than renting equivalent capacity. But the server calculation becomes less attractive when demand is intermittent or utilization falls: idle capacity continues to cost money, while rented compute can be scaled with demand. The listed break-even estimates are useful starting points, not forecasts of a company’s actual savings.
Krishna also pitched on-prem deployment as a way for organizations such as banks and trading firms to keep data and logs under their control. That is a potential governance benefit, not a guarantee that a system will be secure or compliant. Deployments still require appropriate access controls, operations and oversight. Krishna acknowledged that companies need people with expertise in AI kernels and networking, and said that skills gap remains outside the most sophisticated enterprises.
IBM has a commercial interest in the case it is making: it sells on-premise infrastructure and expects AI demand to support sales of its Spyre cards. That does not invalidate Krishna’s argument, but it means the cost and performance claims should be weighed against independent benchmarks and each company’s own deployment estimates.
Products Worth Considering
Samsung 27 Odyssey OLED G5 (G50SF) QHD QD-OLED Gaming Monitor
Samsung 27-inch Odyssey OLED G6 QD-OLED QHD Gaming Monitor
Jackery Explorer 300 Portable Power Station, 292Wh
As an affiliate, we earn on qualifying purchases.
How Open Models Narrowed the Gap
The Artificial Analysis figures show a substantial change in relative performance, though they do not establish that every open model is close to every proprietary one. At the end of 2024, the index listed proprietary o1 at 15.2 and downloadable Qwen2.5 72B at 7.7, or 51% of the leader. At the end of 2025, the corresponding figures were GPT-5.2 at 30.4 and MiMo-V2-Flash at 22.4, or 74%. On October 26, MiMo-V2.6-Pro reached 46.3 against Claude Opus 5.5 at 57.6, or 80%.
The point difference on that index did not shrink across those endpoints: it was 7.5 points at the end of 2024 and 11.3 points on October 26. Relative scores rose partly because the open models improved, but proprietary models also advanced. Krishna’s “90%” figure is not an index result. A higher comparison is possible if Moonshot releases the weights for Kimi K3.1 Pro, which scored 55.1, but the source says the company had promised them by November 4; until the weights are available, it is not a downloadable model.
Token prices also vary by provider and model. The source lists hosted DeepSeek V4.1 Flash 1017 at 3% to 6% of Claude Opus 5.5’s list price, but cites OpenAI’s closed GPT-6 Luna at $0.10 per million input tokens and $0.50 per million output tokens. Those figures illustrate why “open” does not automatically mean cheaper: businesses need to compare specific models, usage patterns, hosting and licence terms. IBM separately said its llm-d project can run large models at five to ten times lower cost; that is an IBM claim, not an independently verified measurement in the source material.
Costs and Model Access Remain Open
The available figures do not establish the total cost for a particular business. The server purchase-versus-rental comparison excludes staff, cooling and networking; it also does not account for financing, maintenance, utilization patterns or whether a company needs the listed capacity. Actual costs will vary by workload and deployment.
It is also unclear whether the models compared perform equally well on the tasks a given organization needs. Artificial Analysis provides a common index, but a general score does not replace testing on a company’s own work. Licences differ: some models use permissive terms, while others have custom or commercial conditions. The source says Kimi K3.1 Pro’s weights were promised by November 4, but that release had not happened as of the October 26 comparison. IBM’s cost claims have not been independently verified in the supplied material.
Test Workloads Before Buying Racks
Businesses weighing deployment can compare a small set of candidate models against representative tasks, then estimate token use, peak demand and expected utilization. They also need to check each model’s licence and budget for the operations staff, networking and cooling required to run it. A pilot can help establish whether an open-weight model performs adequately before a company commits to multiple racks.
Upcoming evidence includes whether Moonshot publishes Kimi K3.1 Pro’s weights as promised by November 4 and how the model performs once it is downloadable. For IBM’s broader pitch, the remaining question is whether reported savings hold under independently measured workloads and full operating costs. Until those details are available, the on-prem case is strongest as a company-specific calculation—not a universal price rule.
Key Questions
What did IBM CEO Arvind Krishna claim about open-weight models?
On IBM’s October 21 earnings call, Krishna said they could cost 10% to 25% as much per token as frontier models while being “90% as effective.” The effectiveness figure was not tied to a named benchmark in his remarks.
How close was the leading downloadable model to the proprietary leader?
On Artificial Analysis’s Intelligence Index as read October 26, 2026, MiMo-V2.6-Pro scored 46.3 and Claude Opus 5.5 scored 57.6. The downloadable model was about 80% of the proprietary model’s score on that index.
Does buying an H200 server automatically cost less than renting?
No. The source’s simple estimate puts the purchase price of an eight-GPU server at about 114 days of AWS rental at full utilization, or roughly 15 months at 25% utilization. It excludes staff, cooling and networking, so it is not a complete comparison.
Why might a business run AI on its own infrastructure?
Krishna cited cost and control over data and logs, particularly for financial institutions. Whether those benefits outweigh infrastructure and staffing costs depends on the business’s workloads, governance requirements and ability to keep the hardware in use.
What should a company check before choosing an open-weight model?
Test the model on the company’s own tasks, compare expected token and hosting costs, estimate utilization, and review the model’s licence. A benchmark score alone does not show how well a model will perform on a specific business workflow.
Source: ThorstenMeyerAI.com
