🔍 Read the full analysis: IBM’s CEO On Why AI May Belong In Your Own Environment on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
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 said enterprises can start with a few racks. An Artificial Analysis comparison read October 26 put the best available open-weight model at 80% of the leading proprietary model; a higher-scoring model’s weights had not yet been released.
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 as leading proprietary models while being “90% as effective,” making them a candidate for companies’ own data centres. A comparison by Artificial Analysis, read October 26, placed the best model with downloadable weights at 80% of the leading proprietary model’s score, leaving both the performance gap and the economics dependent on which models and infrastructure companies choose.
Krishna said enterprises could begin running models on their own infrastructure with “two or three racks.” He argued that many routine tasks—including summarizing email, reviewing contracts and answering customer calls—may not require the most capable available model. He also said open-weight models give businesses permission and intellectual-property rights to run them on-premises or in private clouds. Those are Krishna’s characterizations; the costs and performance depend on the model, licence, workload and deployment.
The independent comparison cited in the source material ranked Xiaomi’s MiMo-V2.6-Pro at 46.3 and Claude Opus 5.5 at 57.6 on Artificial Analysis’s Intelligence Index. That puts the downloadable model at about 80% of the proprietary leader’s score on the index. Moonshot’s Kimi K3.1 Pro scored 55.1, close to the leader, but its weights had not been released as of October 26; the company had promised them by November 4. Until release, it is not an available open-weight alternative.
Krishna also acknowledged that companies need staff who can work on model kernels and networking. He said that capability gap remains outside the most sophisticated enterprises, while using a frontier model through the cloud is the easier option. IBM has a commercial interest in the shift: it sells hardware and services for enterprise computing. The source reports IBM’s prepared remarks also claimed its llm-d project can run large models at five to 10 times lower cost, but provides no independently verified measurement of that claim.
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 Infrastructure
The choice between hosted AI and on-premises deployment is not simply a comparison of model prices. Companies weighing a purchase must account for GPU utilization, staffing, networking, cooling and the work needed to operate models. The source’s illustrative figures list an eight-GPU H200 server from $172,643 and AWS rental of the same GPU class at $63.30 an hour. Its estimate says the purchase could match rental costs after about 114 days at full use, but about 15 months at 25% utilization—before other operating costs. These figures are list-price comparisons, not a forecast for every enterprise.
For organizations handling sensitive data, keeping models and logs on controlled infrastructure may offer a governance or data-sovereignty benefit, as Krishna argued. But that potential benefit must be weighed against the technical burden. A business with sporadic workloads or limited AI operations staff may find hosted services simpler, even if token prices are higher. The performance comparison also cautions against treating Krishna’s 90% figure as a universal benchmark: the measured result varies by model and evaluation, and the closest cited contender had not yet released its weights.
Popular right now
Sony WH-1000XM4C Wireless Noise-Cancelling Headphones, Black
Vitamix Propel Series 510 Professional-Grade Blender, 3 Programs, Black
Ninja Foodi Air Fryer 10 Qt DualZone 6-in-1 Black DZ302
As an affiliate, we earn on qualifying purchases.
How Open Models Have Narrowed the Gap
Krishna described the progress of open models using a broad benchmark example: a top model scoring 70 against 20 or 30 for an open model in late 2024, then 70 against 50 in late 2025. He presented those figures as illustrations on the call, not as results from a named benchmark. The Artificial Analysis series supplied in the source gives a separate comparison: its open model reached 51% of the proprietary leader at the end of 2024, 74% at the end of 2025, and 80% on October 26, 2026. Scores are from the index version current on that date.
That trend supports the narrower point that some open-weight models have become more competitive. It does not establish that all open models are close to the frontier or that the gap has shrunk on every measure. On the cited index, the absolute score difference was 7.5 points at the end of 2024 and 11.3 points on October 26, even as the open model’s score became a larger share of the leader’s. Costs vary too: model access prices, hardware expenses and licensing terms differ. A high leaderboard score alone does not establish that a model can be used commercially or run economically on a company’s equipment.
Performance and Costs Still Vary
Krishna’s 90% performance estimate was not tied to a named evaluation in the cited call, and Artificial Analysis’s index is one comparison rather than a measure of every task an enterprise might run. It remains unclear whether Kimi K3.1 Pro’s promised weights will be released on schedule, what licence will govern them, and whether businesses will be able to run the large model efficiently. The source says no licence had been published as of its October 26 review.
The cost examples also do not settle whether buying hardware is cheaper for a particular company. The stated break-even estimates exclude staffing, cooling and networking and assume particular utilization rates and list prices. IBM’s own lower-cost claim for llm-d was not independently verified in the supplied material. No specific deployment, customer results or independent measurement of IBM’s on-premises savings was provided.
Model Releases and Deployment Economics
The next near-term test is whether Kimi K3.1 Pro’s weights arrive by November 4, as promised, and under what licence. A release would make the model’s score more relevant to businesses considering downloadable models, though its size and hardware demands would still matter. Enterprises assessing on-premises AI will also need to compare actual workload volume and utilization against the full cost of purchasing and operating infrastructure.
IBM’s comments set out a sales case, not a commitment by enterprises to move workloads out of the cloud. Further evidence would include independently checked costs for IBM’s llm-d project, operational results from customers, and model evaluations tied to specific business tasks. Until those details are available, Krishna’s argument is best understood as a case for evaluating on-premises options—not proof that they are cheaper or suitable for every organization.
Key Questions
What did IBM CEO Arvind Krishna say about open-weight AI?
On IBM’s October 21 earnings call, Krishna said open-weight models can cost 10% to 25% of frontier-model token costs while being “90% as effective.” He also said an enterprise could start with two or three racks. These were his claims, not independently established results for every model or workload.
How close are available open-weight models to the leading proprietary model?
In the Artificial Analysis comparison read October 26, Xiaomi’s MiMo-V2.6-Pro scored 46.3, compared with 57.6 for Claude Opus 5.5—about 80% of the proprietary model’s score on that index.
Why might a company run AI on its own infrastructure?
Krishna pointed to control over data and logs, as well as potential token-cost savings. On-premises systems also require hardware, technical staff and ongoing operating costs, so the case depends on workload and utilization.
Is running AI on-premises always cheaper than renting GPUs?
No. The source’s example estimates an H200 server could match rental costs after about 114 days at full utilization, or about 15 months at 25% utilization. Those illustrative figures exclude staffing, cooling and networking and do not predict costs for every organization.
Source: ThorstenMeyerAI.com
