IBM’s CEO Says Businesses Should Take AI Into Their Own Hands

🔍 Read the full analysis: IBM’s CEO Says Businesses Should Take AI Into Their Own Hands on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

IBM CEO Arvind Krishna urged businesses to consider running open-weight AI models on their own infrastructure, arguing on the company’s October 21 earnings call that they can cost a fraction of frontier systems while delivering roughly 90% of their effectiveness. A benchmark comparison available October 26 supports the direction of his argument but puts the strongest currently downloadable model at 80% of the leading proprietary model, leaving both the performance gap and the practical cost of deployment in view.

Krishna said open-weight models cost “25% to 10% of the token cost” of frontier models and are “90% as effective.” He said a large enterprise could begin with “two or three racks.” Open-weight models allow users to run model weights on their own systems or private clouds, rather than relying solely on a provider’s hosted service. The precise savings vary by model and deployment; the earnings-call statement is IBM’s assessment, not a universal price comparison.

Artificial Analysis’s Intelligence Index, read on October 26, scored Claude Opus 5.5 at 57.6 and the best model then listed with downloadable weights, Xiaomi’s MiMo-V2.6-Pro at 46.3. That puts the open model at about 80% of the proprietary leader on the index. At the end of 2024, the comparable share was 51%, according to the source’s historical comparison. The index measures results across ten evaluations; it does not establish that every model performs equally well on a particular company’s tasks.

A potential exception is Kimi K3.1 Pro, which scored 55.1 on the same index, or about 96% of Opus 5.5. But its maker, Moonshot, had only promised to release its weights by November 4. Until those weights are available, the model cannot be counted as downloadable. The figures also do not settle questions about model licences, hardware requirements, staffing or total operating costs.

The Cost of Running AI In-House

The choice affects more than a model’s price per token. Running AI within a company’s own environment may appeal to organizations handling sensitive information: Krishna said banks and traders want control over underlying data and logs. Open weights can also give customers the permission and intellectual-property rights to host a model themselves, subject to the specific licence attached to it.

That control comes with operational demands. Krishna identified shortages of kernel engineers and networking specialists as a constraint for many businesses, saying that gap remains outside the most sophisticated enterprises. Companies may need to buy or rent accelerators, operate the systems, and account for power, cooling, networking and staff. The potential savings therefore depend on how heavily and consistently equipment is used—not just on a model’s token price.

Krishna’s advice also comes from a company with a commercial interest in on-premises computing. IBM sells infrastructure and expects demand for its Spyre cards to benefit its business. That interest does not negate his claims, but it is relevant when weighing them against independent benchmarks and a company-specific cost analysis.

Popular right now

Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Black

AmazonView Latest Price

Vitamix Propel Series 510 Professional-Grade Blender, 3 Programs, Black

AmazonView Latest Price

Dyson Ball Animal 2 Plus Corded Vacuum

AmazonView Latest Price

As an affiliate, we earn on qualifying purchases.

From Benchmark Gaps to Deployment

Krishna said the performance difference between open and frontier models had narrowed in stages. In his account, the leading model scored about 70 against 20 or 30 for an open model in late 2024, then about 70 against 50 in late 2025. He presented tasks such as summarizing emails, reviewing contracts and answering customer calls as examples where a business may not need the most capable available system.

The independent index comparison broadly shows open models gaining ground, but it does not reproduce Krishna’s illustrative scores. It reports scores of 15.2 versus 7.7 at the end of 2024 and 57.6 versus 46.3 on October 26, 2026, for the top proprietary and downloadable-weight models, respectively. The percentage share improved, but the point difference grew from 7.5 to 11.3 because the leading models’ scores also changed.

Hardware economics vary with usage. The source material lists an eight-GPU H200 server at $172,643 and AWS rental of the same GPU class at $63.30 per hour. Its estimate says the purchase price could be recovered in about 114 days at full use, or about 15 months at 25% utilization, before staffing, cooling and networking costs. These are list-price comparisons and illustrative payback periods, not a complete estimate of ownership costs or a forecast for any particular company.

Open Weights and Costs Still Vary

It remains unclear whether Moonshot will release Kimi K3.1 Pro’s weights by its promised November 4 date, what licence would apply, or what computing resources would be required to run the model. The source material says the model has 2.8 trillion parameters and describes it as a data-centre-scale project; the benchmark score alone does not show whether it is practical for a given business.

Nor is there one cost comparison that applies across companies. Hosted prices, usage patterns, hardware utilization and engineering requirements differ. DeepSeek’s hosted V4.1 Flash 1017 is listed as costing 3% to 6% of Claude Opus 5.5’s list price, but the same material gives a different low-cost example: closed model GPT-6 Luna is listed at $0.10 per million input tokens and $0.50 per million output tokens. These examples show why “open” does not automatically mean cheaper than every hosted alternative.

Model permissions also differ. The source lists MIT licences for MiMo-V2.6-Pro and DeepSeek V4.1 Flash 1017, while other models use custom or commercial terms. Companies need to check what each licence allows; benchmark rank alone does not determine whether a model fits their business or legal requirements.

Watch for Weights and Real-World Costs

The next near-term checkpoint in the supplied material is Moonshot’s promised November 4 release of Kimi K3.1 Pro’s weights. If they appear, the model’s licence and practical hardware requirements will help determine whether its high index score translates into a usable open-weight option.

For businesses considering a deployment, the relevant next step is to compare models on their own tasks and estimate total costs under expected utilization. That means factoring in hosting, hardware, staff and infrastructure alongside token prices, and checking each model’s licence. IBM’s earnings-call comments make the case for considering that approach; the benchmark and cost figures do not establish that in-house systems will suit every enterprise.

Key Questions

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

On IBM’s October 21 earnings call, Krishna said open-weight models could cost 10% to 25% as much as frontier models while being about 90% as effective. He also said large enterprises could start with two or three racks of equipment.

Do current benchmarks show open models at 90% of the leader?

Not for the top model with downloadable weights listed in the October 26 comparison. Artificial Analysis scored Xiaomi’s MiMo-V2.6-Pro at 46.3 against Claude Opus 5.5 at 57.6, or about 80% of the proprietary leader on that index.

Could Kimi K3.1 Pro change that comparison?

Its score of 55.1 is about 96% of Opus 5.5’s score. But Moonshot had only promised its weights by November 4, so it was not yet a downloadable model in the comparison described.

Does running an open-weight model always cost less?

No. Costs depend on the model, hosted alternatives, usage and hardware utilization, as well as staffing and infrastructure. The supplied examples include an open model hosted at a small fraction of a proprietary model’s list price, but also a low-priced closed model.

Why would a company run AI on its own infrastructure?

Krishna pointed to data control and the ability to run models on-premises or in private clouds. Companies may also seek control over deployment, but they need suitable hardware, specialist staff and a licence that permits their intended use.

Source: ThorstenMeyerAI.com