🔍 Read the full analysis: IBM’s CEO Makes A Volume Bet On On-Prem AI on ThorstenMeyerAI.com
Get smart everyday buys 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
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, making on-premise deployment attractive for some enterprises. Independent benchmark data in the source puts the leading downloadable model at about 80% of the top proprietary model; the cost case also depends on hardware utilization, staffing and licensing.
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 frontier models while being “90% as effective,” pitching on-premise deployments as a lower-cost option for large enterprises. Independent benchmark data cited by ThorstenMeyerAI.com puts the best currently downloadable model at about 80% of the leading proprietary model, and the economics depend on how intensively companies use and manage their hardware.
Krishna said enterprises could begin with two or three racks and run open-weight models on their own systems or private clouds. He pointed to routine tasks such as summarizing email, reviewing contracts and answering customer calls as work that may not require the most capable model. He also argued that local deployment can give banks and traders greater control over sensitive data and logs.
The source reports that Artificial Analysis’ index, read on October 26, scored the leading downloadable model, Xiaomi’s MiMo-V2.6-Pro, at 46.3, compared with 57.6 for proprietary Claude Opus 5.5. That puts the downloadable model at roughly 80% of the proprietary leader on the index. Moonshot’s Kimi K3.1 Pro scored 55.1, but the source says its weights had only been promised by November 4; until they are released, it is not an available open-weight option.
Krishna’s figures were examples from the earnings-call discussion, not results from a benchmark he identified. The same source says the relative score for the strongest downloadable model rose from 51% of the proprietary leader at the end of 2024 to 80% in October 2026. It also notes that the point gap widened over that period, because proprietary models improved too. Those scores measure performance across the index’s evaluations; they do not establish that models are interchangeable for every business task.
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 Depends on Usage
The pitch matters because enterprise AI bills can scale with the number of tokens processed, while owning hardware shifts more of the expense into upfront investment and ongoing operations. For organizations handling sensitive information, keeping models and data on company-controlled systems may also address data-governance concerns. Krishna said some financial firms want AI to run “in a way that is sovereign,” with underlying data and logs kept from outside access.
But a lower token price does not automatically mean lower total cost. The source gives a list price of $172,643 for an eight-GPU H200 server and an AWS rental price of $63.30 an hour for the same GPU class. Its calculation says the purchase cost is recouped in about 114 days at full use, versus roughly 15 months at 25% utilization, before staff, cooling or networking costs. These are illustrative figures based on listed prices, not a complete comparison of operating costs or a guarantee of savings.
Krishna acknowledged that companies need people who can manage kernels and networks, and said that skills gap remains outside the most sophisticated enterprises. That qualification goes to the heart of the decision: an organization may trade cloud spending for hardware, specialist labor and the risk that equipment sits idle. IBM also stands to benefit if customers buy systems for on-prem AI, so its cost and performance claims should be read as statements from an interested vendor.
Products Worth Considering
FLEXISPOT 35-Inch Electric Standing Desk Converter with Keyboard Tray
Canon IVY 2 Mini Photo Printer Bundle, Pure White
Canon PIXMA PRO-200S Wireless Professional Inkjet Photo Printer
As an affiliate, we earn on qualifying purchases.
How the Performance Gap Is Measured
Krishna described the improvement in open models over time with illustrative comparisons: a leading model scoring 70 against an open model at 20 or 30 in late 2024, then 70 against 50 in late 2025. He used that trajectory to support the claim that a model at a fraction of the cost may be sufficient for many business uses. Those examples were not tied to a named benchmark in the remarks cited by the source.
Artificial Analysis offers a separate comparison using the same ten evaluations across models and labels whether weights are downloadable. In the source’s historical series, the best open-to-proprietary score ratio was 51% at the end of 2024, 74% at the end of 2025 and 80% in October 2026. The benchmark therefore supports a narrowing relative gap, but does not verify Krishna’s specific “90%” estimate for open models as a group.
Access to weights is not the only condition for deployment. The source notes that models carry different licences: some, including MiMo-V2.6-Pro and DeepSeek V4.1 Flash 1017, are listed under MIT licences, while others have custom or commercial terms. A high benchmark score alone does not determine whether a model can be used for a particular commercial purpose or whether an organization can run it economically.
Costs and Model Access Remain Open
The available figures do not establish the total cost of operating an on-prem AI system for a particular company. The server-versus-rental calculation excludes staffing, cooling and networking, and its payback period changes sharply with utilization. Actual costs would also depend on workload, hardware configuration, electricity, maintenance and the cloud alternatives available to each buyer.
The performance comparison is also limited. The index provides one standardized set of evaluations, not proof that a model will match a proprietary system on every task. Kimi K3.1 Pro’s reported score could narrow the gap if its weights are released, but the source says those weights were promised for November 4 and had not yet shipped as of its October 26 reading. Its licence was also not published in the material provided.
IBM’s prepared remarks also claimed its llm-d project can run large models at five to 10 times lower cost; the source says it could not independently check that figure. The company’s own financial interest in selling on-premise systems is relevant, but does not by itself establish whether the claims are accurate. Nvidia had not commented on the separate GeForce firmware issue described in the source, which is not evidence about enterprise deployment economics.
Watch Releases and Deployment Costs
The next near-term test is whether Moonshot releases Kimi K3.1 Pro’s weights as promised by November 4, and what licence accompanies them. Until the files are available and their terms are clear, its benchmark score should not be treated as evidence that enterprises can deploy it as an open-weight model.
For buyers, the practical next step is to compare a workload’s expected use against the full cost of owning and operating hardware, rather than relying on token prices or list-price payback estimates alone. The figures in the source do not settle which option will be cheaper for any specific enterprise. IBM’s claims and customer demand for on-prem systems will also be tested in subsequent company results and deployments; the source provides no confirmed sales outcome from this pitch.
Key Questions
What did IBM CEO Arvind Krishna claim?
On IBM’s October 21, 2026, earnings call, Krishna said open-weight models can cost 10% to 25% as much as frontier models while being about 90% as effective. Those figures were his claim, not a result tied to a named benchmark in the source.
Do independent benchmark figures support the 90% estimate?
Not for the leading downloadable model cited in the October 26 comparison. Artificial Analysis scored MiMo-V2.6-Pro at 46.3 and Claude Opus 5.5 at 57.6, putting the open model at about 80% of the proprietary leader on that index.
Why would a company run an AI model on-premise?
Krishna cited lower costs for some workloads and greater control over data and logs, including for financial firms. On-premise deployment also requires hardware, operating capacity and specialist staff, so savings depend on the company’s use and costs.
Does buying an AI server always cost less than renting cloud GPUs?
No. The source’s illustrative calculation says an eight-GPU H200 server listed at $172,643 reaches its purchase price in about 114 days at full use, but about 15 months at 25% utilization. The comparison excludes staff, cooling and networking and is not a complete cost estimate.
Is Kimi K3.1 Pro available as an open-weight model?
Not according to the source’s October 26, 2026, account. Moonshot had promised the weights by November 4, but they had not shipped, and the source says no licence had been published.
Source: ThorstenMeyerAI.com
