IBM’s CEO Says On-Prem AI Could Make Sense At Volume
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: IBM’s CEO Says On-Prem AI Could Make Sense At Volume on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

TL;DR

IBM CEO Arvind Krishna said open-weight AI models can deliver about 90% of frontier-model performance at 10% to 25% of the token cost, making on-premises deployments worth considering for large enterprises. Independent benchmark data in the source puts the best downloadable model at 80% of the top proprietary model as of Oct. 26, 2026; the economics also depend on usage, infrastructure and staff.

IBM CEO Arvind Krishna said on Oct. 21 that open-weight AI models can cost 10% to 25% as much per token as leading proprietary models while being “90% as effective,” a case for enterprises to consider running AI on their own infrastructure. The claim comes from an IBM earnings call, and a benchmark comparison cited by ThorstenMeyerAI.com puts the strongest downloadable model at 80% of the leading proprietary model as of Oct. 26, underscoring that the cost case depends on the task and the deployment.

Krishna made the comparison during the question-and-answer portion of IBM’s third-quarter earnings call. He said the performance gap between open and proprietary models had narrowed, and argued that many routine business tasks do not require the best available model. He named email summaries, contract review and customer calls as examples. The source reports that he put the potential token cost of open-weight models at “25% to 10%” of frontier-model costs, while describing them as “90% as effective.”

Krishna also pointed to data control as a reason some organizations might choose on-premises systems or private clouds. He said banks and trading firms want to use AI “in a way that is sovereign,” without allowing others to access underlying data or logs. Open-weight models, he said, give customers the permission and intellectual-property rights to run them on their own infrastructure or in private clouds. These are Krishna’s statements; IBM sells hardware and services that could benefit from more on-premises deployments.

The independent comparison cited in the source uses the Artificial Analysis Intelligence Index, which applies the same ten evaluations across models and identifies whether weights can be downloaded. Read on Oct. 26, it listed Claude Opus 5.5 at 57.6 and downloadable Xiaomi MiMo-V2.6-Pro at 46.3. That places the latter at about 80% of the proprietary leader. The source says the comparable share was 51% at the end of 2024. Kimi K3.1 Pro scored 55.1, but its weights had not yet been released; Moonshot had promised them by Nov. 4.

At a glance
reportWhen: Comment made on Oct. 21, 2026; benchmar…
The developmentOn IBM’s Oct. 21 earnings call, CEO Arvind Krishna argued that open-weight models and high usage volumes could make running AI on enterprise-owned infrastructure economically attractive.
Thorsten Meyer AI
Open vs. closed · Oct 2026

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.

80%best open vs. best closed model, 26 Oct
11.3points gap (46.3 vs. 57.6)
96%if Kimi K3.1 Pro's weights ship
51%the same ratio at the end of 2024

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.

  • End of 2024Qwen2.5 72B 7.7 vs. o1 15.251%
  • End of 2025MiMo-V2-Flash 22.4 vs. GPT-5.2 30.474%
  • End of June 2026GLM-5.2 33.7 vs. Claude Fable 5 49.668%
  • October 26, 2026MiMo-V2.6-Pro 46.3 vs. Claude Opus 5.5 57.680%
  • If Kimi K3.1 Pro ships weights55.1 vs. 57.6 · promised “by Nov 4”96%

What Krishna said

Illustrations from the earnings call, not a named benchmark.

  1. Late 2024"70" vs. "20 or 30"Krishna, Q3 call, 21 Oct · no named benchmark
  2. Late 2025"70 to 50"Krishna, Q3 call, 21 Oct · no named benchmark
  3. 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.

ModelIndexLicence
MiMo-V2.6-Pro
Xiaomi
46.3MIT
DeepSeek V4.1 Flash 1017
DeepSeek
45.1MIT
GLM-5.3
Z AI
44.8own commercial licence
Step 5 Preview
StepFun
43.7own commercial licence
Muse Spark Open
Meta
38.9licence 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

The choice between cloud AI and enterprise-owned hardware is not settled by a model’s token price alone. A company buying servers takes on the cost of GPUs, power, cooling, networking and the people needed to operate the systems. The source gives a list price of $172,643 for an eight-GPU H200 server and compares it with AWS rental at $63.30 per hour for the same GPU class. Its calculation estimates that purchase costs could be recouped in about 114 days at full utilization, compared with roughly 15 months at 25% utilization, before operating expenses and staffing.

Those figures illustrate why deployment volume matters: a system used steadily can spread its fixed cost across more work, while idle capacity makes ownership harder to justify. Krishna himself acknowledged that staffing and systems expertise are constraints, asking where enterprises would find kernel engineers and network specialists. He described the frontier cloud as a more expensive but easier option. On-premises AI is not automatically cheaper; the business case varies with workloads, utilization, hardware prices and the organization’s ability to run the systems.

Model scores and licences also affect the decision. The source says some high-scoring open models use permissive MIT licences, while others carry custom commercial terms. A downloadable model is not necessarily unrestricted for every business use. Enterprises need to check licence conditions as well as benchmark results and compute requirements before choosing a model.

Products Worth Considering

Jackery Explorer 1000 v2 Portable Power Station, 1024Wh LiFePO4 Battery

Jackery Explorer 1000 v2 Portable Power Station, 1024Wh LiFePO4 Battery

Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation

Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation

As an affiliate, we earn on qualifying purchases.

IBM’s Pitch and the Model Gap

Krishna described model performance as having moved closer over time. In the call, he recalled a broad benchmark comparison in late 2024 where a leading model scored around 70 and an open model 20 or 30, followed by a 70-to-50 comparison in late 2025. He presented the later improvement as support for choosing a lower-cost model for tasks that do not need frontier performance. The source cautions that these were illustrative examples, not results tied to a named benchmark.

The Artificial Analysis figures offer a separate, dated reference rather than verification of Krishna’s exact percentages. Its cited comparison shows the open-model share of the top proprietary score rising from 51% at the end of 2024 to 80% in October 2026. In raw score points, however, the gap in that comparison widened from 7.5 points to 11.3 because the proprietary model’s score also rose. The source says the share would reach about 96% if Kimi K3.1 Pro’s weights are released, but that model remained closed at the time of publication.

IBM has a commercial interest in the shift: it sells on-premises infrastructure and expects demand to support sales of its Spyre AI cards. The source also reports that IBM claimed its llm-d project can run large models at “5 to 10 times lower cost,” but provides no independent measurement for that statement. It notes that IBM mainframe revenue fell 20% in the quarter, citing The Register. These business details add context, but they do not establish whether a particular customer will save money by moving AI workloads on-premises.

Costs and Model Access Still Vary

The available information does not establish that open-weight models are generally 90% as effective as proprietary models. Krishna’s figure was not tied to a named evaluation, while the cited index puts the leading downloadable model at about 80% of the top proprietary score on its own tests. Benchmarks also cannot show how a model will perform on every company’s specific tasks.

The cost comparison is similarly conditional. The source provides hardware and rental list prices, but its payback estimates do not include staff, cooling or networking, and they assume particular utilization levels. It does not supply a complete cost model for a specific enterprise workload. Token prices also differ by provider, model and service: a cheap hosted model may be a better fit than buying and operating servers.

Some model weights cited as forthcoming were not available as of Oct. 26. Kimi K3.1 Pro’s promised release date was Nov. 4, and the source says no licence had been published. Its score should not be treated as evidence of an available open-weight option until the weights and terms are released. IBM’s own cost claims have not been independently verified in the material provided.

Watch Model Releases and Deployment Data

The next near-term test is whether Moonshot releases Kimi K3.1 Pro’s weights and publishes licence terms by the date it promised. Any release would allow businesses to assess whether the model is genuinely downloadable for their intended use and what hardware it requires. The Artificial Analysis index may also change as new models and evaluation results are added.

For enterprises, the decision will require comparing hosted-model prices with a workload-specific estimate for owned infrastructure. That assessment should account for expected usage, peak demand, idle time, staff and facilities costs, as well as data-control requirements. IBM’s earnings-call remarks make the case for considering on-premises deployment; they do not establish a universal cost advantage. Future customer deployments and independently checked cost figures would help show where the approach pays off.

Key Questions

What did IBM CEO Arvind Krishna say about open-weight AI?

On IBM’s Oct. 21, 2026, 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 cited routine business tasks as possible uses for lower-cost models.

Does the cited benchmark confirm the 90% figure?

No. The Artificial Analysis comparison cited by ThorstenMeyerAI.com puts the leading downloadable model at 46.3 out of 57.6 for the top proprietary model, or about 80%. The benchmark does not directly verify Krishna’s illustrative claim.

When can on-premises AI be less expensive?

It may make financial sense when an organization can keep expensive GPU systems busy enough to spread purchase and operating costs across substantial workloads. The source’s payback illustration changes from about 114 days at full utilization to roughly 15 months at 25%, before staff, cooling and networking costs.

Why might a company choose on-premises AI apart from cost?

Krishna cited data control and sovereignty, particularly for banks and trading firms that want to limit access to underlying data and logs. A company’s security, regulatory and operational requirements will determine whether a private deployment is suitable.

Is Kimi K3.1 Pro an open model now?

Not according to the source’s status on Oct. 26, 2026. Moonshot had promised to release the weights by Nov. 4, but they had not shipped and no licence had been published. Until release, it should not be counted as an available open-weight model.

Source: ThorstenMeyerAI.com

You May Also Like

Analyzing The $400 Million Public AI Fund: Infrastructure For Sovereignty Or Political Rhetoric?

A detailed examination of the $400 million public AI initiative, its progress, funding transparency, and implications for AI sovereignty and policy.

How Check Point Accelerates Vulnerability Research With Daybreak

A headline links Check Point’s vulnerability research to Daybreak, but the available material gives no method, examples or measured speed gains.

Where The Agent Economy Is Being Built: Four AI Deals To Know

Deepgram, Mintive, Dream and Meta announced moves spanning voice, agent identity, building design and personalized shopping on Oct. 29.

Google’s AI Attack Findings: Which Claims Are Evidence, And Which Are Projections?

Google reports one server that automated vulnerability discovery and exploitation, while more autonomous attacks remain a forecast.