IBM chief executive Arvind Krishna told investors on the company's third-quarter earnings call that open weight models now make a lot of sense for running on a company's own servers.
In his view the gap with closed frontier models has nearly closed: open models deliver about 90% of their quality at 10–25% of the token cost.
How the gap closed
Krishna put it in benchmark terms. In 2024 the best closed model might score 70 on a broad test while an open one scored 20 or 30. By late 2025 it was roughly 70 against 50, and today open models have almost caught up.
That leaves businesses with a simple choice: pay ten times more for the very best, or take something 90% as good. Krishna believes most tasks, such as summarising email, reviewing contracts, answering customers and drafting messages, work just as well on the cheaper option.
Two or three racks, not enough engineers
IBM is seeing fast-growing interest in running AI on premises. Krishna thinks a useful setup fits in two or three racks, but building it is hard:
- you need engineers who understand kernels and networking;
- the model must sit on GPUs that are close to each other on the network;
- results have to be cached properly so responses stay fast.
He offered no ready answer beyond promising IBM software to ease deployment and predicting more demand for its Spyre mainframe accelerators.
Our take
For anyone building bots or AI features for websites, the lesson is practical: routine jobs like customer replies, summaries and text analysis rarely need the most expensive model. An open one, self-hosted or from a cheap provider, can cut token bills several times over. Keep in mind that IBM sells exactly the hardware and software for such setups. More in The Register.