🔍 Read the full analysis: IBM’s CEO On Choosing Between Cloud AI And Running It Yourself 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 models can deliver about 90% of frontier-model performance at a fraction of the token cost, and argued that enterprises can run them on their own infrastructure. An independent model index reviewed October 26 puts the leading downloadable model at 80% of the top proprietary score; buying hardware can also be costly when it is not kept busy.
IBM CEO Arvind Krishna used the company’s October 21 earnings call to argue that businesses can run open-weight AI models on their own infrastructure at much lower token costs than frontier services. His case matters to enterprises weighing data control and recurring AI bills against the expense and expertise needed to operate powerful systems themselves; an independent benchmark reviewed October 26 puts the best downloadable model at 80% of the leading proprietary model’s score, rather than 90%.
In the call’s question-and-answer session, Krishna said open-weight models could cost “25% to 10% of the token cost” of frontier models while being “90% as effective.” He said a large enterprise could begin with “two or three racks.” These are Krishna’s estimates, not a neutral cost comparison, and IBM has a commercial interest in on-premises infrastructure.
Krishna pointed to routine tasks such as summarizing email, examining contracts, and answering customer calls as work that may not require the most capable model. He also said banks and traders may want AI systems they can run “in a way that is sovereign,” keeping underlying data and logs under their control. Open weights, he said, provide permission and intellectual-property rights to run models on premises or in private clouds.
The independent Artificial Analysis Intelligence Index, read October 26, lists Xiaomi’s MiMo-V2.6-Pro at 46.3, against 57.6 for Claude Opus 5.5, the highest proprietary score in the comparison. That makes the downloadable model’s score about 80% of the proprietary leader’s. The same source lists Kimi K3.1 Pro at 55.1, but its weights had not shipped; its maker had promised them by November 4. Until release, it is not an available open-weight alternative.
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
The choice is not simply between an expensive cloud and a cheaper local model. On-premises deployment shifts costs from usage fees to hardware, electricity, networking, cooling, staffing, and the work of keeping equipment busy. It may also give organizations more direct control over sensitive data and logs, a priority Krishna associated with banks and traders.
The source’s illustrative hardware comparison shows how utilization changes the arithmetic: an eight-GPU H200 server was listed from $172,643, while renting the same GPU class on AWS was listed at $63.30 an hour. At continuous full use, the purchase price would equal roughly 114 days of rental; at 25% use, the comparison stretches to about 15 months. Those calculations exclude operating and personnel costs, so they are not a complete break-even estimate.
Krishna acknowledged a practical constraint: enterprises need people who can work on model kernels and underlying networks. He said that skills gap remains outside the most sophisticated companies. That means the economics and feasibility will vary by workload, utilization, model license, and an organization’s ability to run the infrastructure.
Products Worth Considering

Apple MacBook Air 13-inch with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and 12MP Center Stage Camera – Midnight

Apple MacBook Air 15-inch (2026) with M5 chip, 16GB memory, 512GB SSD — Midnight

Jackery Solar Generator 1000 V2 with 100W Solar Panel
As an affiliate, we earn on qualifying purchases.
What Krishna Said on the Call
Krishna described open and proprietary model performance as having converged over time. He offered illustrative comparisons from late 2024 and late 2025, then framed the decision as whether a business would accept a less capable model for a lower token price. His examples were general business tasks, not evidence that every workflow can use a smaller or open-weight model without trade-offs.
Artificial Analysis provides a separate comparison using the same ten evaluations across models. Its figures show the leading downloadable model rising from 51% of the top proprietary score at the end of 2024 to 80% on October 26, 2026. The absolute gap in index points was 7.5 at the end of 2024 and 11.3 in the October comparison, because proprietary models also improved. The index’s scores are one benchmark, not a universal measure of performance across every task.
Costs also depend on where and how a model is accessed. The source says hosted DeepSeek V4.1 Flash 1017, whose weights are MIT-licensed, costs 3% to 6% of Claude Opus 5.5’s listed price. It also cites OpenAI’s closed GPT-6 Luna at $0.10 per million input tokens and $0.50 per million output tokens. Those examples show why the phrase “a tenth of the price” cannot be applied to every model or service: model, provider, token type, and deployment method all affect the bill.
IBM also said in its prepared remarks that its llm-d project can run large models at five to ten times lower cost. That is an IBM claim; the supplied material does not provide an independent measurement. IBM sells infrastructure that could benefit if companies choose to run models in their own data centers.
Benchmark and Cost Caveats
The 90% performance figure is Krishna’s characterization, not a finding from the Artificial Analysis index. That index puts MiMo-V2.6-Pro at 80% of Claude Opus 5.5’s score, while Kimi K3.1 Pro reaches 96% in the listed comparison only if its weights become available. The source says they were promised by November 4, 2026; it does not establish that the release occurred.
It is also unclear how representative the price examples are of a particular organization’s actual spending. Listed rental and server prices do not include all operating costs, and a break-even calculation depends on utilization and workload. The supplied material does not give a complete, independently audited comparison across cloud providers and on-premises systems. Model licenses differ too: a downloadable model is not automatically unrestricted for commercial use.
Krishna’s statements come from an IBM earnings call, and IBM has a business interest in hardware and on-premises deployments. The source also reports that IBM expects on-premises demand to support sales of its Spyre cards, while citing a 20% decline in mainframe revenue for the quarter. Those details provide business context, but do not independently validate the model cost or performance claims.
Model Releases and Enterprise Trials
The next concrete point to watch is whether Kimi K3.1 Pro’s weights arrive by November 4, as promised in the source material, and what license accompanies them. If released, the model’s index score would make it a closer open-weight comparison with Claude Opus 5.5, but its availability and license would still need to be checked.
For enterprises, the practical test is likely to be workload-specific trials: compare model quality on actual tasks, calculate token and infrastructure expenses at realistic utilization, and check whether the license permits the intended use. The source material does not identify a new IBM deployment date or a confirmed follow-up announcement. Further evidence on real operating costs, IBM’s infrastructure sales, and models’ performance on business tasks will clarify how far Krishna’s case applies beyond the earnings-call pitch.
Key Questions
What did IBM CEO Arvind Krishna argue?
Krishna said open-weight models can cost 10% to 25% as much as frontier models while being “90% as effective,” and argued that some enterprises could run them on their own infrastructure. Those figures are his claims, not independent measurements.
Does an independent benchmark support the 90% figure?
Not for the leading downloadable model cited in the source. Artificial Analysis listed MiMo-V2.6-Pro at 80% of Claude Opus 5.5’s score on October 26, 2026. Kimi K3.1 Pro scored 96% of the leader, but its weights had not yet shipped.
Is running AI on-premises always cheaper?
No. Buying hardware can reduce repeated rental or token charges when equipment is used heavily, but the calculation also depends on utilization, staffing, power, cooling, and networking. The cited price comparison excludes several of those costs.
Why might a company run a model itself?
Krishna said companies, including banks and trading firms, may want greater control over data and system logs. Running models on premises or in a private cloud can support that goal, but requires suitable infrastructure and technical staff.
Source: ThorstenMeyerAI.com
