🔍 Read the full analysis: IBM’s CEO Explains The Volume Economics Of On-Prem AI 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
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, and said large enterprises could begin with two or three racks. An independent benchmark cited in the source puts the leading downloadable model at 80% of the leading proprietary model; actual on-prem savings depend on utilization, infrastructure and staffing.
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 at a fraction of frontier-model token costs, while retaining more control over data. He said such models are “90% as effective” at “10%” of the cost, but an independent benchmark cited in the source puts the leading downloadable model at 80% of the top proprietary model as of October 26.
In the call’s question-and-answer session, Krishna said open-weight models cost “25% to 10% of the token cost of the frontier models.” He argued that many routine enterprise jobs—including summarizing email, reviewing contracts and answering customer calls—may not require the most capable available model. He said a large enterprise could start deploying on-prem AI with two or three racks. These are Krishna’s estimates and examples, not independently verified cost or deployment measurements.
The comparison in the source draws on the Artificial Analysis Intelligence Index, which evaluates models across ten assessments. Its October 26 reading gave proprietary Claude Opus 5.5 a score of 57.6 and downloadable Xiaomi MiMo-V2.6-Pro a score of 46.3, or about 80% of the proprietary model’s score. That is a narrower relative gap than the 51% recorded between the listed leaders at the end of 2024, although the raw score gap grew from 7.5 points to 11.3 points. The source stresses that these are index comparisons, not Krishna’s named benchmark.
A possible exception depends on a future release: Moonshot’s Kimi K3.1 Pro scored 55.1, about 96% of Claude Opus 5.5, but its weights had not shipped as of October 26. Moonshot had promised them by November 4, according to the source. Until downloadable weights are available, it should not be counted as an open-weight model. Model licences also vary, so a high score alone does not establish that a model can be used freely for a particular commercial deployment.
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
The economic case rests on a trade-off: buying and operating compute may reduce recurring inference costs and give an organization more control over its data, but only if the hardware is used enough to justify its fixed costs. The source gives list-price examples: an eight-GPU H200 server starts at $172,643, while renting the same GPU class from AWS costs $63.30 an hour. Its estimate says purchasing breaks even in about 114 days at full utilization, but takes about 15 months at 25% utilization—and excludes staffing, cooling and networking.
Those figures are illustrative rather than a complete total-cost comparison. Actual costs will vary with negotiated prices, workload, power, hardware configuration and the level of service needed. Krishna also acknowledged a practical constraint: enterprises need people who can optimize kernels and manage the underlying network. That skills gap can make cloud services the easier route, even if they cost more per token.
For banks and other organizations handling sensitive information, the decision is not only about price. Krishna said some businesses want AI to run in a “sovereign” way, without outside access to underlying data or logs. On-prem systems or private clouds may offer greater control, but the source does not establish that every deployment will meet a particular organization’s security or regulatory requirements.
Products Worth Considering
Apple MacBook Pro 16-inch Laptop with M5 Pro, 18-Core CPU, 20-Core GPU, 48GB Memory, 2TB SSD, Space Black
Apple MacBook Pro 14.2-inch with M5 Pro, 15-core CPU, 16-core GPU, 48GB Memory, 1TB SSD, Silver
Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, and Liquid Retina Display
As an affiliate, we earn on qualifying purchases.
IBM’s Case and Its Interests
Krishna described model performance as having narrowed over time: he said that in late 2024 a leading proprietary model might score 70 against 20 or 30 for an open model, and that by late 2025 the comparison was closer to 70 against 50. He presented the cost-performance trade-off as a reason to use less expensive models for routine work and reserve frontier systems for tasks that need them.
IBM has a direct commercial interest in wider on-prem AI adoption: it sells infrastructure, including its Spyre accelerator cards. That does not invalidate Krishna’s argument, but it is relevant when weighing his cost and performance claims. The source also reports that IBM said its llm-d project can run large models at “5 to 10 times lower cost”; it provides no independent measurement supporting that figure. Krishna himself identified specialist staffing and network expertise as obstacles for all but the most sophisticated enterprises.
The source’s pricing examples also show why model hosting matters. It says hosted DeepSeek V4.1 Flash 1017 costs 3% to 6% of Claude Opus 5.5’s list price, while closed GPT-6 Luna is listed at $0.10 per million input tokens and $0.50 per million output tokens. These examples compare particular products and prices, not a universal open-versus-closed cost ratio; hosting, usage and pricing terms can change the calculation.
Costs, Scores and Weights Still Vary
The figures do not settle whether on-prem AI is cheaper for a particular company. The server-versus-cloud calculation omits staff, cooling and networking and assumes a utilization rate; actual workloads may not keep hardware busy. The source does not supply a full comparison of total costs, or independently verify IBM’s llm-d savings claim.
Benchmark scores also cannot establish that models perform equally well on every company’s tasks. The index comparison measures a selected set of evaluations, while Krishna’s 90% figure was not tied to a named benchmark. The source reports that Kimi K3.1 Pro’s weights were promised for November 4, but says they had not been published by October 26 and that no licence had been released. Its availability and terms therefore remained uncertain at the time of reporting.
Finally, open weights do not automatically mean unrestricted commercial use. The source identifies different licence terms across models, including custom commercial and community licences. Organizations must check the applicable terms and their own technical and compliance needs before treating a model as deployable.
Watch Model Releases and Utilization
The next specific milestone cited is Moonshot’s promised November 4 release of Kimi K3.1 Pro weights. Whether those weights arrive, what licence accompanies them and what hardware they require will affect how much weight to give the index comparison that places the model near the current proprietary leader.
For enterprises weighing deployment, the practical next step is to compare workloads against the full cost of owned infrastructure, including staffing and operating expenses, rather than relying only on token prices or a best-case break-even estimate. IBM’s earnings-call comments set out the company’s case; the source provides no confirmed company-by-company adoption plan or independently audited estimate of savings.
Key Questions
What did IBM CEO Arvind Krishna say about on-prem AI?
Krishna said open-weight models can cost 10% to 25% as much per token as frontier models while being about 90% as effective, in his view. He also said a large enterprise could start with two or three racks. These are statements made on IBM’s earnings call, not independently verified estimates.
Does the benchmark confirm that open models are 90% as capable?
Not in the figures cited. Artificial Analysis’s October 26 index listed the leading downloadable model at 46.3, compared with 57.6 for the leading proprietary model—about 80% by score. The index uses ten evaluations and does not measure every organization’s tasks.
When can buying AI servers be cheaper than renting?
It depends on hardware use and full operating costs. The source estimates an eight-GPU H200 server could match rental costs in about 114 days at full load, or about 15 months at 25% load, before staff, cooling and networking. These list-price calculations are examples, not a guarantee of savings.
Were Kimi K3.1 Pro’s model weights available?
Not as of October 26, 2026, according to the source. Moonshot had promised the weights by November 4, but their release and licence were still pending at the time of reporting.
Source: ThorstenMeyerAI.com
