🔍 Read the full analysis: IBM’s CEO Says The Road To On-Prem AI Runs Through Volume on ThorstenMeyerAI.com
Get audio and creator gear 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 the effectiveness of top proprietary models at 10% to 25% of their token cost, and suggested large enterprises could start with two or three racks. An Artificial Analysis comparison puts the strongest currently downloadable model at 80% of the proprietary leader, while the economics of owning hardware depend on keeping it busy and covering staffing and infrastructure costs.
IBM CEO Arvind Krishna argued on the company’s October 21 earnings call that enterprises can run open-weight AI models on their own infrastructure for 10% to 25% of the token cost of leading proprietary models, while getting about 90% of their effectiveness. The pitch makes on-premises AI a potential option for companies with enough usage and technical capacity, but an independent benchmark comparison and hardware-cost figures show that neither the performance gap nor the savings are settled for every workload.
Krishna said large enterprises could begin with two or three racks and use open-weight models for routine work such as summarizing email, reviewing contracts, answering customer calls and drafting messages. He also pointed to data control: banks and trading firms, he said, may want AI systems that keep underlying data and logs within infrastructure they control. Open weights, he said, give organizations permission and intellectual-property rights to run models on premises or in private clouds.
The performance comparison is less favorable than Krishna’s 90% figure for the strongest model whose weights were downloadable in the cited scoreboard. Artificial Analysis’s Intelligence Index, read October 26, scores MiMo-V2.6-Pro at 46.3 and proprietary Claude Opus 5.5 at 57.6. That places the open model at about 80% of the proprietary model’s score on the index. The comparison uses the index’s evaluations; it does not establish that every model performs at the same relative level on a company’s specific tasks.
Hardware economics depend on utilization. The source cites an 8-GPU H200 server listed from $172,643 and AWS rental of the same GPU class at $63.30 an hour. A simple comparison puts the purchase cost at roughly 114 days of continuous rental, or about 15 months if the system is used 25% of the time. Those estimates exclude staff, cooling and networking, and do not account for other purchase or cloud costs.
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 On-Prem AI Can Pay Off
Krishna’s argument puts a practical question before companies weighing AI infrastructure: whether a cheaper model, run frequently on owned equipment, can meet business needs without paying frontier-model prices for every request. For organizations processing large volumes of routine work, the combination of lower token costs and control over data location may make self-hosting worth evaluating.
But lower model prices alone do not establish a lower total cost. An enterprise must account for hardware utilization, engineering staff, networking, cooling, deployment and ongoing operations. Krishna himself identified a skills gap, asking where companies would find kernel engineers and network specialists; he said that gap remained outside the most sophisticated enterprises. Businesses with uneven demand or limited infrastructure teams may find cloud services easier, even if the per-token price is higher.
The proposal also comes from a company that sells enterprise hardware. IBM expects on-premises AI demand to support sales of its Spyre cards, according to the source material. That commercial interest does not invalidate Krishna’s claims, but it is relevant when evaluating his recommendation. The figures should be treated as IBM’s case for a business model, not as an independent cost study.
Products Worth Considering

Samsung Galaxy Tab A11+ Kids Edition, 128GB, Royal Blue

DJI Mini 5 Pro Fly More Combo with DJI RC 2, 1-Inch CMOS

Apple MacBook Air 15-inch Laptop with M5 Chip, 24GB Memory, 1TB SSD, Midnight
As an affiliate, we earn on qualifying purchases.
The Benchmark Gap Since 2024
Krishna described the open-versus-proprietary performance gap as narrowing over time. In his account, a top model scored around 70 while an open model scored 20 or 30 in late 2024; by late 2025, he said, the comparison was about 70 to 50. Those were illustrative comparisons, not scores tied to a named benchmark in the remarks.
Artificial Analysis provides a separate reference point. Its chart lists the best open model at 51% of the proprietary leader at the end of 2024, 74% at the end of 2025, and 80% on October 26, 2026. The current scores show a smaller relative gap than in 2024, though the point difference in the listed scores is 11.3 now, compared with 7.5 at the end of 2024. Different models lead each period, so these figures do not show a like-for-like comparison of the same systems.
One possible exception is Kimi K3.1 Pro, which scores 55.1 on the index, or about 96% of Claude Opus 5.5’s 57.6. The source says Moonshot had promised to release its weights by November 4, but they had not shipped as of the October 26 comparison. Until weights are available, it is not an open-weight option that companies can download and run themselves.
Costs, Licences and Model Access
The available figures do not establish whether on-premises AI will cost less for a particular company. The server-versus-rental calculation is a simplified comparison, not a full ownership analysis: it excludes staffing, cooling and networking, and does not model changing hardware prices or utilization. Actual savings will depend on how much a system is used and which tasks it runs.
Model access and licensing also vary. The source says MiMo-V2.6-Pro and DeepSeek V4.1 Flash 1017 use MIT licences, while some other models carry custom or commercial terms. A high benchmark score does not by itself establish that a model can be used for every commercial purpose. Companies would need to check the relevant licence and deployment requirements.
Prices cited for hosted models are not a universal comparison of equivalent services. DeepSeek V4.1 Flash 1017 is reported at 3% to 6% of Claude Opus 5.5’s list price, while the source also gives a low listed price for closed model GPT-6 Luna. Differences in model capability, service, usage and pricing terms make a single token-price comparison incomplete. IBM’s separate statement that its llm-d project can run large models at five to ten times lower cost is also an IBM claim; the source says it could not independently verify that figure.
Finally, Kimi K3.1 Pro’s reported score cannot yet be counted as a downloadable-model result on the facts provided: the weights were promised for November 4, and the source says no licence had been published. Its availability, licence and ability to run at practical cost remain open questions.
What Enterprises Should Verify
The next near-term milestone in the source’s account is whether Moonshot releases Kimi K3.1 Pro’s weights by November 4, and under what licence. That would determine whether the model can be evaluated as an open-weight option rather than only compared by its reported index score.
For enterprise buyers, the next step is not simply to compare headline token prices. They would need to test candidate models on their own workloads, review licences, estimate utilization and include the cost of people and supporting infrastructure. The source material does not provide a company-specific deployment plan or a full cost comparison, so it remains unclear how many organizations could achieve the savings Krishna described.
IBM’s remarks also leave unanswered how quickly enterprises can fill the engineering and networking roles Krishna identified, and whether its hardware and software offerings will meet the needs of those deployments. The case for on-prem AI will depend on demonstrated performance and total operating costs—not only on a model’s benchmark score or nominal price per token.
Key Questions
What did IBM CEO Arvind Krishna say about open-weight AI?
Krishna said open-weight models could cost 10% to 25% as much per token as frontier models while being about 90% as effective. Those are his claims from IBM’s October 21 earnings call, not a finding from a named benchmark.
How did the strongest downloadable model compare on the cited index?
On the Artificial Analysis Intelligence Index read October 26, MiMo-V2.6-Pro scored 46.3, compared with 57.6 for proprietary Claude Opus 5.5. That is about 80% of the proprietary score on that index.
Does owning AI hardware automatically save money?
No. A purchase-versus-rental comparison depends heavily on utilization and omits costs such as staff, cooling and networking. Companies also need to assess model performance and the work required to operate the system.
Why might companies want to run AI on premises?
Krishna said some enterprises, including banks and trading firms, want control over where data and logs are handled. Running models on owned infrastructure or a private cloud may support that goal, but the exact protections depend on the system and its deployment.
Is Kimi K3.1 Pro available as an open-weight model?
Not according to the source material as of October 26, 2026. Moonshot had promised its weights by November 4, but they had not yet shipped and no licence had been published.
Source: ThorstenMeyerAI.com
