On the third-quarter call, IBM chief executive Arvind Krishna made a claim that deserves attention from anyone buying AI. He said open-weight models now cost anywhere from 10 to 25 percent of the token cost of frontier models while being about 90 percent as effective, and that an average large enterprise could get started with two or three racks of infrastructure.
It is a vendor’s claim, from a vendor that sells infrastructure and software for running those models. It may well be broadly right for some workloads. It is not a number to put in a business case without checking it. Other coverage reports the same figures, and some commenters have called the claim self-serving, which is a fair thing to test.
What Krishna actually argued
He described the gap between the best frontier models and open alternatives as large in late 2024, with a top model scoring 70 where an open one scored 20 or 30 on broad benchmarks. By late 2025, he said, it had narrowed to something like 70 against 50, and he expects it to move towards 100 against 90. He added that the remaining challenge is operational, because running models well takes kernel engineers, network design, GPU clustering and caching expertise that many enterprises lack. The full remarks are in the earnings call transcript.
His examples of where cheaper models are good enough were summarising an email, examining a contract, answering a customer call and composing an email. He also said IBM and Red Hat are building software to make this easier, and he credited an open-source project called llm-d with helping clients run large models at five to ten times lower cost by reusing previously processed context.
Why the question is also about control
Krishna tied the cost argument to sovereignty. Using a large public model, he said, can leave traces in logs that reveal what an organisation is asking and which data it uses. Some customers therefore want models they can run on their own infrastructure, where they control the logs and the data. He said this conversation barely existed in 2024, began to pick up in 2025 and is growing in 2026.
Red Hat makes a similar case. Its enterprise inference playbook argues that a capable open model on infrastructure you control can cost meaningfully less per token than a comparable proprietary API, partly because API providers add a margin, and it suggests sending routine requests to smaller models while keeping frontier models for hard tasks.
Where the maths needs checking
A per-token comparison leaves out most of the cost of owning the answer. One independent check of the claim against list prices found that the outcome depends heavily on which frontier model you compare with and how busy the hardware is. Self-hosting only wins when the cluster is well utilised, and an idle rack costs the same as a busy one.
The full cost of self-hosting includes hardware or rental, power and cooling, the people to run it, security work, monitoring, model updates and the time spent evaluating whether a model is still good enough after each change. As one buyer’s guide puts it, open weights lower licensing and lock-in costs but do not eliminate the cost of serving predictions. That is the point to hold on to when a headline ratio is quoted.
Open weight does not mean no licence
This is the point that matters most to a licensing team. An open-weight model is not necessarily open source. Each model comes with a licence that sets out what you may do with it, and the terms differ. Some restrict commercial use above a certain scale, some restrict particular uses, and some require attribution or limit how the model can be modified or redistributed.
Krishna himself pointed out that organisations have the permission and the intellectual property rights to run these models on their own infrastructure. That is a statement about the models he had in mind. For any model you choose, read the licence and keep a record of it, the same way you would for any other software asset.
What to do with the claim
Treat it as a hypothesis. Pick two or three real workloads with steady volume, and price them three ways: a frontier API, a hosted open-weight service, and self-hosting at realistic utilisation. Include staffing. Test quality with your own data, not with a vendor’s benchmark. Check each model licence. Then decide where each kind of work should run.
It is also worth asking your existing vendors how they would respond. A supplier whose AI features are priced per action or per token may have room to move if you can show a credible alternative for part of the workload.
Governance does not go away
Moving a workload to an open-weight model you run yourself shifts responsibility to you. You are now accountable for security, for keeping the model patched and current, for monitoring quality, and for the data used with it. Logs and prompts that once sat with a supplier now sit in your environment, which is the point of the exercise, and also a new thing to protect.
Keep an inventory of every model in use, its version, its licence, where it runs and who owns it. That register is the AI equivalent of a software asset record, and it will be the first thing an auditor, a regulator or a customer asks for.
A mixed approach is likely
For most organisations the answer will not be all frontier or all open. Frontier models will keep their place for the hardest and least predictable tasks, and open models will take on routine, high-volume work where cost and control matter more than the last few points of quality. The skill is in sending each kind of work to the right place, and in keeping your contracts flexible enough to move work as the options change.
Conclusion
IBM’s chief executive says open-weight models now deliver most of the performance of frontier models at a fraction of the cost, and that two or three racks are enough to start. For routine, high-volume tasks that may prove true, and the sovereignty argument is strong for some organisations.
Test the numbers against your own workloads, count the whole cost of running the models, and read every licence. The saving may be real, but it has to be earned.
