🔍 Read the full analysis: On-Prem AI Has A Volume Question, IBM’s CEO Says on ThorstenMeyerAI.com
Get smart everyday buys 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 said open-weight models can deliver about 90% of frontier-model performance at 10% to 25% of the token cost, making on-premise AI attractive for some enterprise workloads. An Artificial Analysis index puts the leading downloadable model at 80% of the top proprietary model as of Oct. 26, while the economics of buying hardware depend heavily on utilization and operating costs.
IBM CEO Arvind Krishna said on Oct. 21 that open-weight AI models can cost 10% to 25% as much per token as frontier models while being “90% as effective,” a pitch for enterprises to run AI on their own infrastructure. An Artificial Analysis comparison reviewed Oct. 26 puts the leading downloadable model at 80% of the top proprietary model, showing that the value of on-premise AI depends on workload, hardware use and the performance gap a company can accept.
Krishna made the comparison during the question-and-answer portion of IBM’s third-quarter earnings call. He said the performance gap between open and closed models had narrowed, and argued that many routine tasks — including summarizing email, reviewing contracts and answering customer calls — do not require the most capable model. He also said a large enterprise could begin with “two or three racks.”
The figures are Krishna’s characterization, not results from a specific benchmark he named. Artificial Analysis, which evaluates models on the same set of ten assessments, listed MiMo-V2.6-Pro at 46.3 as the leading downloadable model on Oct. 26, compared with Claude Opus 5.5 at 57.6. That is about 80% of the proprietary model’s score. The source’s historical comparison puts the open model at 51% of the leading closed model at the end of 2024.
A model’s price and availability also depend on how it is accessed. The source reports that hosted DeepSeek V4.1 Flash 1017 costs 3% to 6% of Claude Opus 5.5’s list price, but notes that closed OpenAI GPT-6 Luna is listed at $0.10 per million input tokens and $0.50 per million output tokens. Those prices do not, by themselves, establish the full cost of running a system on premises.
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
When On-Prem AI Can Pay Off
The argument matters to businesses weighing recurring cloud charges against the upfront and ongoing costs of operating their own AI hardware. Buying equipment can reduce per-use costs at high utilization, but idle capacity weakens that case. The source compares an eight-GPU H200 server listed from $172,643 with AWS rental of the same GPU class at $63.30 an hour. Its calculation estimates a purchase payback of about 114 days at full use, versus about 15 months at 25% use, before staff, cooling and networking.
Those figures are illustrative list-price comparisons, not a full enterprise cost model. Actual results depend on negotiated hardware and cloud prices, utilization, power, facilities, maintenance and the engineering needed to run models. For organizations with steady workloads, in-house infrastructure may also offer more control over data and logs. For smaller or irregular workloads, cloud access may be easier to scale and operate.
Krishna acknowledged that operational expertise is part of the trade-off. He pointed to shortages of kernel and networking engineers, saying that gap remains for all but the most sophisticated enterprises. IBM sells on-premise hardware and software, so its cost and deployment claims come from a company with a commercial interest in greater on-prem demand.
Products Worth Considering

Dell XPS 13 13.4-Inch Touchscreen Laptop, Intel Core 5 320, 16GB RAM, 512GB SSD

Vari Ergo Electric Standing Desk, 54×26, Light Wood/White

Logitech MX Master 4 Ergonomic Wireless Mouse with Haptics – Graphite
As an affiliate, we earn on qualifying purchases.
Benchmark Gaps and Model Access
Krishna described model progress using an illustrative sequence: in late 2024, he said, a top model might score 70 while an open one scored 20 or 30; by late 2025, he described scores of 70 and 50. He then argued that customers may prefer a model that is “90% as good” at one-tenth the cost. Because he did not identify that sequence as a particular standardized benchmark, it should not be treated as directly equivalent to the Artificial Analysis index.
On that index, the source reports a widening raw score difference between its top proprietary and open models: 7.5 points at the end of 2024 and 11.3 points on Oct. 26, 2026. The open model’s relative share of the leader’s score nonetheless rose from 51% to 80% over the period. The comparison depends on which models are available and the index version used.
One possible exception was not yet downloadable at the time of the report. Kimi K3.1 Pro scored 55.1, close to Claude Opus 5.5’s 57.6, but its maker, Moonshot, had only promised to release the weights by Nov. 4. Until the weights are actually released, it is not an open-weight option for companies to download and run. Model licences also differ, so a strong score does not automatically mean unrestricted commercial use.
Costs, Scores and Release Dates
The 90% performance claim remains approximate: Krishna’s examples were not linked to a named benchmark, while the independent index uses its own evaluations and model set. Scores also cannot capture every company’s task mix, accuracy requirements or deployment constraints. It is not clear how the models would compare on any particular enterprise’s production workload.
The server payback figures omit staff, cooling and networking, and use stated list prices and utilization assumptions. They do not establish what a specific company would save. IBM also said its llm-d project can run large models at “5 to 10 times lower cost,” but the source says it could not independently check that claim.
It remains unclear whether Moonshot will release Kimi K3.1 Pro’s weights by its promised Nov. 4 date, and what licence will apply. A separate enthusiast project, gpu-unlock, reportedly raises some RTX 30-card matrix benchmark results by as much as 80%, with smaller gains in prompt processing and almost no change in text generation. The report says Nvidia has not commented; it also notes that Nvidia’s GeForce licence prohibits circumventing technical limitations. That project is distinct from IBM’s enterprise on-premise case.
Watch Model Releases and Deployment Costs
The next near-term marker is whether Moonshot makes Kimi K3.1 Pro’s weights available by Nov. 4, as promised, and what terms accompany any release. The Artificial Analysis comparison could change if new models or revised index results shift the leading open and proprietary scores.
For enterprise buyers, the practical next step is testing candidate models on their own workloads and comparing the full cost of cloud access with hardware, staffing and facility costs at realistic utilization levels. IBM’s on-premise pitch is an argument to evaluate, not proof that owning infrastructure will be cheaper for every organization. No deployment decision or company-specific savings figure is established in the source material.
Key Questions
What did IBM’s CEO say about open-weight AI?
Arvind Krishna said open-weight models can cost 10% to 25% of frontier-model token costs while being “90% as effective.” He presented this as an argument for using them on enterprise infrastructure, not as a result from a benchmark he identified.
Does the independent benchmark support the 90% figure?
Not for the leading downloadable model listed on Oct. 26. Artificial Analysis showed MiMo-V2.6-Pro at 46.3 against Claude Opus 5.5 at 57.6, or about 80% of the proprietary model’s score. A different model, Kimi K3.1 Pro, scored 55.1, but its weights had not yet been released.
Is running AI on-premise always cheaper?
No. The source’s hardware comparison estimates that an eight-GPU H200 server could pay off in about 114 days at full utilization, but about 15 months at 25% utilization, before operating costs. The outcome depends on utilization and expenses such as staff, cooling and networking.
Why might a company prefer on-premise AI?
Krishna cited control over data and logs, as well as the lower token costs he attributes to open-weight models. On-premise deployment can offer more control, but it also requires hardware and technical expertise.
What remains to be confirmed about Kimi K3.1 Pro?
Moonshot had promised the model’s weights by Nov. 4, but the source says they were not available on Oct. 26. Whether the release happens on time and what licence applies remain unclear.
Source: ThorstenMeyerAI.com
