What happens to your engineering team if you lose access to the model it runs on? Stefan [Surname], Lead for AI and APIs at sipgate, joins Beyond Vibe Coding to explain why the company is bringing its AI in house.

After Anthropic temporarily suspended access to Claude Fable 5 earlier this year, sipgate ran an experiment. It rented a cluster of B300 GPUs, hosted the open-weight model GLM 5.2, and asked its developers to drop Claude Code for a week. The verdict surprised Stefan: developers found it faster, rated the quality close to what they were used to, and some asked when they could have it back. sipgate is now moving the hardware into its own data centre.

The conversation widens to where AI actually sits in a telephony business, why fine-tuning small models on anonymised customer data can beat a large general model, and whether central infrastructure beats a powerful machine on every desk. Stefan also makes a case for specialist distilled models over shrunken versions of the frontier ones, and predicts that European hosted models will be the default for sipgate's customers by the end of the year.

His most memorable line, borrowed from a colleague: "I don't know how to click anymore. I need an agent to do it." Stefan's point is that products without an agent layer will soon be products people can't use.

Listen to the full episode: https://open.spotify.com/episode/6KydPigXl7rgmDedFMlO1U