🔍 Read the full analysis: Running AI Yourself: IBM’s CEO On Models And On-Premises 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 said open-weight AI models can deliver about 90% of frontier-model performance at a fraction of the token cost, and that enterprises can run them on their own infrastructure. An Artificial Analysis comparison puts the leading downloadable model at 80% of the top proprietary model as of October 26, 2026; the cost case for buying hardware depends heavily on utilization, staffing and other expenses.
IBM CEO Arvind Krishna said on the company’s October 21 earnings call that open-weight AI models can cost 10% to 25% as much as frontier models while being “90% as effective,” and that large enterprises could begin running them on two or three racks. The pitch matters to companies weighing data control and AI spending, but an independent model comparison puts the best downloadable model at 80% of the leading proprietary model as of October 26, while the economics of buying servers depend on how intensively they are used.
Krishna made the comparison during the call’s question-and-answer session, describing the performance gap as having narrowed over time. He said that in late 2024 a top model might score 70 while an open model scored 20 or 30; by late 2025, he said, the comparison was 70 to 50. He then framed the purchase decision this way: “Would you like something that’s 10% of the cost if it’s 90% as good as the best, or do you want to pay that 10 times higher cost?” Those score examples were illustrations; Krishna did not name a benchmark for them.
Artificial Analysis, which compares models across ten evaluations and identifies whether weights can be downloaded, recorded a score of 46.3 for the top downloadable model, Xiaomi’s MiMo-V2.6-Pro. The leading proprietary model, Anthropic’s Claude Opus 5.5, scored 57.6, putting the downloadable model at about 80% of its score. The source’s historical comparison shows the open model’s share rising from 51% at the end of 2024 to 74% at the end of 2025, then 68% at the end of June 2026 and 80% in October.
The comparison could change if Moonshot releases the weights for Kimi K3.1 Pro. Artificial Analysis listed it at 55.1, or 96% of Claude Opus 5.5’s score, but Moonshot had only promised to release the weights “by Nov 4.” Until the weights are available, it remains a closed model. The source also says no licence had been published, leaving both availability and usage terms unresolved.
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 Capacity
Krishna’s case is not just about model scores. He said banks and trading firms want to run AI “in a way that is sovereign,” so that outside parties cannot access underlying data or logs. Open-weight models can give organizations permission to deploy a model on on-premises systems or private clouds, but data control and cost savings are separate considerations: a model’s weights being available does not, on its own, determine how it may be used or what deployment will cost.
The hardware figures in the source show why utilization matters. An eight-GPU H200 server was listed from $172,643, while renting the same GPU class on AWS was priced at $63.30 an hour. At continuous full use, the purchase price would equal about 114 days of rental charges; at 25% utilization, the comparison stretches to about 15 months. These are list-price comparisons, not a full cost-of-ownership calculation: staffing, power, cooling, networking and maintenance are excluded.
Krishna also acknowledged an operational constraint. He asked, “Where are my kernel engineers? Where are the people who worry about the underlying network?” and said that skills gap remains outside the most sophisticated enterprises. That caveat matters because buying accelerators does not automatically provide the engineering capacity to deploy and operate models efficiently.
Products Worth Considering
Amazon Kindle Paperwhite (12th Generation, 2024), 16 GB, 7-inch Display, Jade
Amazon Kindle Paperwhite Signature Edition (12th Generation, 2024), 32 GB, Metallic Black
Sonos Era 100 Wireless Smart Speaker, Black
As an affiliate, we earn on qualifying purchases.
IBM’s Open-Model Business Case
Krishna named routine business tasks—summarizing email, reviewing contracts, answering customer calls and composing messages—as examples where an organization may not need the best available model. His broader argument is that a less capable but cheaper model may be sufficient for many workloads, leaving the most demanding tasks to frontier services. That is a proposed deployment strategy, not evidence that every task can be shifted without affecting results.
IBM has a commercial interest in wider on-premises adoption: it sells hardware, including its Spyre accelerator cards. The source also reports that IBM said its llm-d project can run large models at “5 to 10 times lower cost”; that is an IBM claim, and the source says it could not independently verify the figure. It cites The Register as reporting that IBM mainframe revenue fell 20% in the quarter, 18 months into the z17 cycle. The combination makes Krishna’s remarks relevant both as an enterprise technology argument and as a sales pitch from a hardware provider.
Model prices and licences complicate the simple “open is cheaper” comparison. The source says hosted DeepSeek V4.1 Flash 1017, whose weights carry an MIT licence and whose Artificial Analysis score was 45.1, costs 3% to 6% of Claude Opus 5.5’s listed price. But it also lists closed GPT-6 Luna at $0.10 per million input tokens and $0.50 per million output tokens. Prices depend on the model, provider and usage, while licences vary: the source identifies MiMo-V2.6-Pro and DeepSeek V4.1 Flash 1017 as MIT-licensed, and says other models use custom or commercial terms.
Performance, Licences and True Costs
The available evidence does not establish that open-weight models generally perform at 90% of the best proprietary models. The Artificial Analysis comparison places the leading downloadable model at 80% of the leader’s score on October 26, 2026. The source also notes that the raw score gap was 7.5 points at the end of 2024 and 11.3 points in October, even as the open model’s score became a larger share of the proprietary leader’s score. Different benchmarks and tasks may produce different comparisons.
It is also unclear whether Kimi K3.1 Pro’s promised weights will arrive on schedule, what licence will accompany them, and what infrastructure would be needed to run a model the source describes as having 2.8 trillion parameters. The hardware payback estimates omit several expenses, and actual costs would depend on utilization and operating requirements. IBM’s cost claim for llm-d was not independently measured in the source material.
Watch for Weights and Deployment Data
The next near-term test is whether Moonshot releases Kimi K3.1 Pro’s weights by its stated November 4 target and publishes a licence. If the weights ship, the model can be assessed as downloadable; until then, its high score does not make it an open-weight option for deployment.
For enterprises, the practical next step is to compare models on their own tasks, check each model’s licence and estimate hardware use alongside staffing, power, cooling and networking costs. IBM’s remarks set out a case for on-premises AI, but the source material does not provide a complete cost comparison across workloads or evidence that the proposed savings will apply broadly.
Key Questions
What did IBM CEO Arvind Krishna say about open-weight models?
On IBM’s October 21, 2026, earnings call, Krishna said open-weight models could cost 10% to 25% as much as frontier models while being “90% as effective.” He also said a large enterprise could start with two or three racks.
How did the leading downloadable model compare with the top proprietary model?
In Artificial Analysis data read on October 26, 2026, MiMo-V2.6-Pro scored 46.3, compared with 57.6 for Claude Opus 5.5. That put MiMo-V2.6-Pro at about 80% of the proprietary model’s score on that index.
Is Kimi K3.1 Pro an open-weight model?
Not yet, according to the source. Moonshot had promised to release its weights “by Nov 4,” but until the weights are available it remains a closed model. Its licence was also not published in the source material.
Does buying an AI server always cost less than renting GPUs?
No. The source’s list-price comparison estimated that an eight-GPU H200 server would match AWS rental charges in about 114 days at full utilization, but about 15 months at 25% utilization. Those figures exclude staffing, power, cooling, networking and other operating costs.
Why might a company run a model on its own infrastructure?
Krishna said some banks and trading firms want control over where AI runs and who can access underlying data and logs. On-premises or private-cloud deployment may support that goal, but organizations still need to review each model’s licence and account for the expertise and expense needed to operate it.
Source: ThorstenMeyerAI.com
