🔍 Read the full analysis: Is On-Prem AI Worth It? IBM’s CEO Points To Volume 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
IBM CEO Arvind Krishna argued on the company’s October 21 earnings call that open-weight models can deliver about 90% of frontier-model performance for 10% to 25% of the token cost. Independent benchmark data reviewed October 26 put the best currently downloadable model at 80% of the top proprietary model; owning hardware can also be more expensive than renting when utilization is low.
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 per token as frontier models while being “90% as effective,” making them a possible fit for companies running AI on their own infrastructure. Independent benchmark data reviewed October 26 supports the case that downloadable models have improved, but puts the strongest currently available one at 80% of the leading proprietary model—and the economics of buying servers depend heavily on how often they are used.
Krishna made the comparison during the earnings-call Q&A, describing how the performance gap between open and proprietary models had narrowed. He said that in late 2024 a leading model might score 70 on a broad benchmark while an open model scored 20 or 30; by late 2025, he said, the comparison was 70 to 50. He framed the trade-off 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 figures were his examples, not results tied to a specific benchmark.
Artificial Analysis’s Intelligence Index offers a separate comparison. On the version checked October 26, Xiaomi’s downloadable MiMo-V2.6-Pro scored 46.3, against 57.6 for proprietary Claude Opus 5.5. That makes MiMo about 80% of the leader’s score on that index. The source’s historical comparison puts the open model at 51% of the proprietary leader at the end of 2024 and 74% at the end of 2025. These scores reflect the index’s evaluations, not a guarantee that a model will perform similarly on every company’s workload.
One model could narrow the measured gap further, but it is not yet available with downloadable weights. Kimi K3.1 Pro scored 55.1 on the index, about 96% of Claude Opus 5.5’s score, but Moonshot had only promised to release its weights “by Nov 4.” Until that happens, it remains a closed model. The source also reports that no licence had been published for it at the time of writing.
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 Owning AI Hardware Pays
Krishna’s case matters to companies deciding whether to rent AI capacity from cloud providers or buy servers. The potential savings are not just a question of model prices. On-premises deployment can give an organization more control over where data and logs are held, while making the organization responsible for the infrastructure and people needed to run the system.
The source gives a price illustration for an eight-GPU H200 server listed from $172,643 and AWS rental of the same GPU class at $63.30 an hour. Using those list prices, it estimates the purchase cost would match rental after about 114 days if the server ran at full load, but about 15 months at 25% utilization. That estimate excludes staff, cooling and networking, so it is not a complete comparison of total cost. It does show why the business case depends on steady demand: idle hardware keeps costing money without replacing much cloud spending.
Krishna also pointed to data sovereignty as a reason some banks and trading firms may prefer private infrastructure. But the operational burden is real. He acknowledged that companies need people who understand model kernels and networking, and said that gap remains outside the most sophisticated enterprises. For readers weighing the decision, the takeaway is not that on-prem AI is automatically cheaper: it is that volume, utilization, staffing and data requirements can change the answer.
Popular right now

Ninja Luxe Café Mini Plus Espresso and Drip Coffee Machine with AutoFroth, Stainless Steel, ES351

Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation

Bose QuietComfort Ultra Wireless Noise Cancelling Earbuds, 2nd Gen, with 20W 2-Port Wall Charger
As an affiliate, we earn on qualifying purchases.
The Claim Behind IBM’s Pitch
Krishna’s remarks were made by the head of a company that sells enterprise infrastructure, making IBM an interested party in the on-premises argument. He said open-weight models give organizations permission and intellectual-property rights to run them on their own infrastructure or private clouds, subject to each model’s licence. His examples of routine work included summarizing email, examining contracts, answering customer calls and composing messages—tasks he suggested may not require the most capable available model.
IBM also said in prepared remarks that its llm-d project can run large models at five to 10 times lower cost. That is the company’s claim; the source material does not provide an independently checked measurement for it. Krishna said IBM expects on-premises demand to support sales of its Spyre cards. The source also cites The Register’s report that IBM mainframe revenue fell 20% in the quarter, 18 months into the z17 cycle, a separate business development that may be relevant to IBM’s commercial incentives.
Model access and pricing vary. The source says hosted DeepSeek V4.1 Flash 1017, whose weights are MIT-licensed, costs 3% to 6% of Claude Opus 5.5’s list price. It also cites closed GPT-6 Luna at $0.10 per million input tokens and $0.50 per million output tokens. Those examples do not establish a universal price advantage: hosting arrangements, token use, licence terms and the models being compared all matter. A downloadable model’s licence may also impose conditions, so a high benchmark score alone does not establish that it suits a particular business.
Costs, Weights and Workloads
The benchmark comparison does not settle whether open-weight models are “90% as effective” for a particular enterprise. Krishna’s figures were illustrative, while Artificial Analysis measures performance across a defined set of evaluations. Results can differ by task, deployment and model version. The available figures also compare benchmark scores, not the complete costs or outcomes of operating a business application.
Kimi K3.1 Pro’s weights had not shipped as of the source’s October 26 check, so its near-leader score cannot yet be treated as an open-model result available for deployment. Its release timing and licence remain unresolved in the supplied material. Server payback estimates are also incomplete because they exclude staffing, cooling and networking, and depend on the assumed utilization rate. Actual cloud quotes, hardware configurations and workload volumes could change the calculation.
The Next Tests for Buyers
The next near-term marker is Moonshot’s promised November 4 release date for Kimi K3.1 Pro’s weights. If the weights are released, their licence and practical hardware requirements will help determine whether the model can be used as an open alternative, rather than judged only by its index score.
For enterprises, the immediate next step is to compare models on their own tasks and calculate costs using expected token demand and realistic server utilization. That assessment should include licensing, engineering capacity and infrastructure overhead, not just per-token prices or headline benchmark results. IBM has not provided an independently verified measure of its llm-d cost claim in the source material, so that claim remains separate from the benchmark comparison.
Key Questions
What did IBM’s CEO claim about open-weight AI models?
Arvind Krishna said on IBM’s October 21 earnings call that open-weight models can cost 10% to 25% of frontier-model token costs while being “90% as effective.” The performance figure was an illustrative claim, not a result he tied to a named benchmark.
How close are the best available open models to proprietary models?
On the Artificial Analysis Intelligence Index checked October 26, 2026, Xiaomi’s MiMo-V2.6-Pro scored 46.3, compared with 57.6 for Claude Opus 5.5. That is about 80% of the leading score on this index, not a universal measure of performance.
Does buying AI servers always cost less than renting them?
No. The supplied estimate says an eight-GPU H200 server listed from $172,643 reaches rental-cost parity with AWS’s $63.30 hourly price after about 114 days at full utilization, or about 15 months at 25%. It excludes staff, cooling and networking, and actual costs depend on a company’s workload and terms.
Is Kimi K3.1 Pro an open-weight model yet?
Not according to the source’s October 26 check. Moonshot had promised the weights “by Nov 4,” but they had not shipped, and the source said no licence had been published.
Source: ThorstenMeyerAI.com
