What's so surprising?

I'd say they have a "Tier 2" inference stack, but that'd be Apple Silicon and AMD: their well below that.

In part because unlike those, no one can get their hands on the hardware, so they miss out on a massive amount of free development, testing, etc.

At this point there's years of complaints about open weight models performing worse on their platform, tool calling makes it especially easy to tell since it's so sensitive to these issues.