TBPN guests: On-premise and open models will grow, but frontier AI may grow faster

The guests said that the on-premise and open-model market would continue to grow quickly in absolute terms. But if companies can earn high returns relatively quickly from frontier AI in domains where intelligence is a key bottleneck, they predicted that the frontier-model market would grow faster relative to that segment. One guest suggested that most tokens may be served by non-frontier models, while most spending may go to frontier models.

Discussing a report that Goldman Sachs might ban or restrict Claude Code, Codex and other coding agents internally, the participants did not confirm that a decision had been made. One guest described a possible restriction for junior analysts as potentially short-sighted, but predicted that Goldman Sachs would still need to spend heavily on penetration testing.

As context, a participant cited his team’s finance benchmark: according to him, none of the models tested could successfully complete an analyst task, and the closest model received about 50% of the credit. He argued that open models would perform worse than frontier models. These were the participants’ assessment and benchmark results, not an independently verified model comparison or an established fact about Goldman Sachs policy.