🔍 Read the full analysis: IBM’s View Of AI: In-House Runs And A Bet On Volume on ThorstenMeyerAI.com
Get hardware and tech essentials 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 argued on the company’s October 21 earnings call that open-weight models can deliver about 90% of frontier-model performance at 10% to 25% of the token cost. An Artificial Analysis comparison puts the strongest currently downloadable model at 80% of the leading proprietary score, while the economics of on-premises hardware depend heavily on utilization and operating costs.
IBM CEO Arvind Krishna used the company’s October 21 earnings call to argue that enterprises can run open-weight AI models on their own infrastructure for much less than using frontier models, while retaining about 90% of their effectiveness. An independent benchmark comparison supports the direction of his argument but puts the strongest currently downloadable model at 80% of the leading proprietary model’s score, highlighting that the cost case depends on the task, model availability and the expense of keeping hardware busy.
In the earnings-call Q&A, Krishna said open-weight models cost “25% to 10% of the token cost” of frontier models while being “90% as effective.” He described common enterprise uses such as summarizing email, examining contracts and answering customer calls—tasks where, in his view, companies may not need the most capable available model. He also said a large enterprise could begin with “two or three racks” and run models on premises or in private clouds.
Krishna’s figures are an executive’s characterization, not the result of a named benchmark he cited. Artificial Analysis’s Intelligence Index, read October 26, scored downloadable MiMo-V2.6-Pro at 46.3 and proprietary Claude Opus 5.5 at 57.6. That makes the open model’s score about 80% of the leader’s in this comparison. Kimi K3.1 Pro scored 55.1, but its weights had not been released; its developer had promised them by November 4. Until they are available, it is not a downloadable open-weight model.
The cost comparison also turns on how an enterprise obtains computing capacity. The source’s listed example gives an eight-GPU H200 server a starting price of $172,643, compared with $63.30 an hour to rent the same GPU class on AWS. At full use, the purchase price alone would equal about 114 days of rental; at 25% utilization, about 15 months. Those calculations exclude staffing, cooling and networking, so they are not a complete ownership-cost comparison. IBM also said its llm-d project can run large models at five to 10 times lower cost, a company claim for which the source did not provide an independently checked measurement.
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 Economics Depend on Utilization
Krishna’s case matters to companies deciding whether to rent AI computing from cloud providers or invest in their own data-center capacity. Lower token prices can make open-weight models attractive for routine, high-volume workloads, particularly when organizations can host them privately and have suitable hardware already available. Krishna also said banks and trading firms value control over data and logs, describing this as a question of sovereignty and privacy.
But buying hardware is not automatically cheaper. The server-versus-rental arithmetic changes sharply with utilization, and the listed payback estimates omit operating expenses and technical staffing. Krishna himself identified a shortage of kernel and networking engineers as a constraint for all but the most sophisticated enterprises. That makes in-house deployment a bet not only on cheaper models, but also on sustained demand, effective use of equipment and the ability to operate it.
There is also a commercial interest behind IBM’s pitch: IBM sells on-premises systems and expects demand to support its Spyre accelerator cards. The source notes that IBM’s mainframe revenue fell 20% in the quarter, 18 months into the z17 cycle. That does not establish why Krishna made the argument, but it is relevant context when weighing the company’s claims about the advantages of enterprise-owned infrastructure.
Products Worth Considering
Apple MacBook Pro 14.2-inch with M5 Max, 18-Core CPU, 20-Core GPU, 48GB Memory, 2TB SSD, Space Black
Apple MacBook Pro 16.2-inch with M5 Pro, 18-Core CPU, 20-Core GPU, 48GB Memory, 1TB SSD, Nano-Texture Display, Space Black
Apple MacBook Pro 14‑inch (2021) with M1 Pro, 16 GB RAM, 512 GB SSD – Space Gray
As an affiliate, we earn on qualifying purchases.
From Benchmark Gap to Rack Costs
Krishna described open models as catching up rapidly. In his account, a broad benchmark comparison in late 2024 put a top model at 70 and an open model at 20 or 30; by late 2025, he said, the comparison was 70 to 50. Those were illustrative figures, not a benchmark series identified on the call. The source’s separate Artificial Analysis comparison shows open models rising from 51% of the proprietary leader’s score at the end of 2024 to 74% at the end of 2025 and 80% on October 26, 2026.
The present-day comparison does not mean the absolute score gap has narrowed in that dataset: it was 7.5 points at the end of 2024 and 11.3 points on October 26, 2026, as proprietary scores also rose. The index applies the same 10 evaluations to each model and identifies whether weights can be downloaded. Scores and rankings can change as models and index versions change, and one aggregate score cannot establish that a model performs equally well on every business task.
Price claims vary by model and provider. The source says hosted DeepSeek V4.1 Flash 1017, whose weights carry an MIT licence, costs 3% to 6% of Claude Opus 5.5’s listed price. It also cites closed GPT-6 Luna at $0.10 per million input tokens and $0.50 per million output tokens. These are provider prices, not a universal measure of the total expense of running open models. Licensing matters too: models may carry permissive, custom or commercial terms, so a high benchmark score alone does not determine whether an enterprise can use a model as intended.
Costs and Model Availability Remain Open
The available figures do not establish that open-weight models are 90% as effective across all enterprise tasks. The benchmark cited in the source puts the top downloadable model at 80% of the proprietary leader’s score, while Krishna did not identify the benchmark behind his comparison. Kimi K3.1 Pro could narrow the measured gap if its promised weights are released, but its availability and licensing terms were not confirmed in the source material.
The cost estimates are also incomplete. The hardware comparison excludes staffing, cooling and networking; it does not show how much an enterprise would spend on installation, maintenance or financing. The source gives a utilization-based calculation, not a verified payback period for a particular company. IBM’s llm-d cost claim likewise remains a company claim in the material provided, without an independent measurement described there. Nvidia had not commented on the reported GeForce firmware limitation in the source material.
Watch for Weights and Real-World Costs
The next concrete point to watch is whether Kimi K3.1 Pro’s weights arrive by November 4, as its developer promised, and what licence accompanies them. A release would make it possible to assess the model as downloadable weights rather than compare its score as if it were already available for self-hosting. Enterprises weighing deployment options will also need to compare their own workload volumes and utilization with cloud rental costs, including staffing and facility expenses not captured by the listed hardware figures.
Further evidence could come from disclosed, independently checkable measurements of IBM’s llm-d cost claim and from companies reporting actual operating costs and model performance for defined tasks. Until then, Krishna’s remarks are best read as an argument for considering open models and on-premises capacity—not proof that every enterprise will get a 90%-as-effective model at a tenth of the total cost.
Key Questions
What did Arvind Krishna say about open-weight AI models?
On IBM’s October 21 earnings call, Krishna said open-weight models can cost 10% to 25% as much per token as frontier models while being “90% as effective.” He did not identify a named benchmark for that 90% comparison in the source material.
How close are the best downloadable models to the leading proprietary model?
In the Artificial Analysis comparison read October 26, 2026, MiMo-V2.6-Pro scored 46.3 against Claude Opus 5.5’s 57.6, or about 80% of the leader’s score. Kimi K3.1 Pro scored 55.1, but its weights had not yet been released.
Does buying AI servers automatically save money?
No. The source’s example estimates that an eight-GPU H200 server listed from $172,643 equals about 114 days of AWS rental at full utilization, or about 15 months at 25% utilization. Those calculations exclude staffing, cooling and networking, so they do not represent total ownership costs.
Why might an enterprise run a model on its own infrastructure?
Possible reasons include lower costs for sustained, high-volume workloads and greater control over data and logs. Krishna cited data sovereignty as a concern for banks and trading firms, while also acknowledging that specialist engineering skills can be a barrier.
Are all open-weight models free to use commercially?
No. Model licences vary, and a downloadable model may carry commercial conditions or other limits. The source notes different licensing terms across models; organizations need to check the applicable licence rather than relying on a benchmark score alone.
Source: ThorstenMeyerAI.com
