🔍 Read the full analysis: What’s Driving IBM’s Bet On On-Prem AI? on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
IBM CEO Arvind Krishna is pitching open-weight AI models running on company infrastructure as a lower-cost, more private option for some enterprise work. Independent benchmark data shows leading downloadable models have narrowed the gap with proprietary systems, but the 90% comparison depends on a model whose weights have not yet been released. The economics of buying hardware also depend heavily on keeping it busy and having the staff to operate it.
IBM CEO Arvind Krishna used the company’s October 21 earnings call to make the case that enterprises can run open-weight AI models on their own infrastructure at a fraction of frontier-model costs. He said such models can be “90% as effective” at 10% to 25% of the token cost, but independent benchmark figures cited by ThorstenMeyerAI.com put the leading downloadable model at about 80% of the top proprietary model as of October 26. The pitch matters to companies weighing AI spending, data control and the cost of buying or renting specialized computing hardware.
Krishna said open-weight models give companies permission to run models on premises or in private clouds, and described use cases such as summarizing email, reviewing contracts and answering customer calls as tasks that may not need the most capable model. He also said banks and traders want AI systems that keep underlying data and logs under their control. These were Krishna’s arguments on the call, not a guarantee that every enterprise can obtain the same costs or performance.
The source’s check of the Artificial Analysis Intelligence Index found MiMo-V2.6-Pro at 46.3 and proprietary Claude Opus 5.5 at 57.6, putting the downloadable model at about 80% of the leader. The source says the open-to-proprietary score ratio was 51% at the end of 2024. Kimi K3.1 Pro scores 55.1, or about 96% of Claude’s score, but Moonshot had only promised its weights “by Nov. 4.” Until the weights are released, it is not an open-weight model.
Krishna also said a large enterprise could begin with two or three racks and cited open models costing “25% to 10%” of frontier-model token prices. The source notes that hosted pricing varies: DeepSeek V4.1 Flash 1017 was listed at 3% to 6% of Claude Opus 5.5’s list price, while closed GPT-6 Luna was listed at $0.10 per million input tokens and $0.50 per million output tokens. Those examples show why model choice and workload matter; they do not establish a universal cost advantage for on-prem systems.
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 Case for Owning AI Capacity
IBM’s argument brings together two enterprise concerns: operating costs and control over sensitive information. Downloadable weights can allow organizations to choose where a model runs, while avoiding some per-token charges associated with hosted services. For businesses with large, steady workloads or strict data-handling rules, owning infrastructure may be worth evaluating alongside cloud services.
But the purchase price is only part of the calculation. The source compares an 8-GPU H200 server listed from $172,643 with renting the same GPU class on AWS for $63.30 an hour. Its estimate says the hardware purchase would take about 114 days to match rental costs at full utilization, versus about 15 months at 25% utilization. The calculation excludes staff, cooling and networking, and is not a full operating-cost comparison. A lightly used system could take much longer to justify than a busy one.
IBM has a commercial interest in encouraging on-prem AI: Krishna expects demand to support sales of IBM’s Spyre cards. The source also reports that IBM’s mainframe revenue fell 20% in the quarter, 18 months into the z17 cycle, citing The Register. Those details make it important to distinguish the business case Krishna is advocating from independently measured performance and costs.
Products Worth Considering
Apple iPad Air (4th Generation, 10.9-inch, Wi-Fi, 256GB) – Space Gray
Apple iPad Air (2025) 13‑inch Wi‑Fi 256 GB – Space Gray (M3 chip)
Samsung 27-Inch Odyssey QD-OLED G8 4K 240Hz Gaming Monitor
As an affiliate, we earn on qualifying purchases.
Benchmark Gains and Remaining Gaps
Krishna described the performance gap as having narrowed over time. On the call, he recalled broad benchmark examples of a top model scoring 70 against an open model at 20 or 30 in late 2024, then 70 against 50 in late 2025. He framed the current trade-off as choosing a model that costs less if it is close enough for the task. These figures were presented as illustrations, rather than results from a named benchmark.
Artificial Analysis provides a separate comparison, using the same ten evaluations across models and identifying whether weights can be downloaded. Its figures in the source show the open model’s share of the proprietary leader’s score rising from 51% at the end of 2024 to 80% in October 2026. Yet the absolute score gap in that comparison grew from 7.5 points to 11.3 points, because proprietary models also improved. A percentage of the leader’s score is not the same thing as equivalent performance on every enterprise task.
Licences also affect whether a model can be used as a company expects. The source identifies MiMo-V2.6-Pro and DeepSeek V4.1 Flash 1017 as MIT-licensed, while several other models have their own commercial or community terms. A high leaderboard score alone does not settle whether a model’s licence permits a particular deployment or commercial use.
Costs, Staffing and Model Access
The available comparison does not show how well each model performs across every business task, or whether the benchmark results translate into a particular company’s workflow. The source gives list prices and a simplified rent-versus-buy calculation, but does not provide a complete cost model covering staffing, electricity, cooling, networking, maintenance and utilization. Krishna himself identified a shortage of kernel and networking engineers as a barrier for all but the most sophisticated enterprises.
It also remains uncertain whether Kimi K3.1 Pro’s weights will be released by the promised November 4 date, what licence would accompany them, and whether organizations could run a model of that scale economically. IBM’s claim that its llm-d project can run large models at five to 10 times lower cost is also not independently verified in the source. IBM’s commercial interest in on-prem hardware should be considered when evaluating its claims.
The Next Tests for On-Prem AI
The immediate milestone in the source’s account is Moonshot’s promised release of Kimi K3.1 Pro’s weights by November 4. If released, the model’s licence and the practical requirements for running it will matter as much as its benchmark score. Enterprises will also need to compare models on their own tasks and calculate expected utilization before committing to hardware.
IBM’s cost case will be tested as customers decide whether to buy capacity, rent it or use a mix of both. The source does not report a verified deployment result showing that IBM’s rack estimate or llm-d cost claim applies broadly. Readers should look for reported customer workloads, operating costs and model-specific performance data as those decisions develop.
Key Questions
What is IBM’s argument for running AI on premises?
IBM CEO Arvind Krishna says open-weight models can cost less to use and can run on a company’s own infrastructure or private cloud. He also points to data control as a reason some organizations may prefer this approach.
Do open-weight models currently score 90% of the best proprietary models?
Not in the benchmark comparison cited by the source. MiMo-V2.6-Pro scored 46.3 against Claude Opus 5.5 at 57.6, about 80%. Kimi K3.1 Pro scored about 96% of the leader, but its weights had not yet been released.
Is on-premise AI always cheaper than renting cloud GPUs?
No. The source’s comparison estimates that buying an H200 server could match rental costs in about 114 days at full utilization, but about 15 months at 25% utilization, before operating expenses. Actual economics depend on use and costs not included in that estimate.
What is still uncertain about IBM’s cost claims?
The source does not independently verify IBM’s claim that its llm-d project can lower large-model costs by five to 10 times. It also does not establish a complete on-prem operating-cost comparison or show that the claimed savings apply to every workload.
Source: ThorstenMeyerAI.com
