4 arguments worth recording.

I think this is the cleanest test we have of whether AI assistants can hold retention past the novelty window. Four months is not enough time to see second-order churn, and the valuation assumes they already have it.

The valuation assumes nothing. It prices the option that they figure it out before the next round. You are treating a Series B like a Series D, and the evidence we have is that Forerunner and Menlo wrote the check.

My read is that staying independent was the expensive choice. The acquirer was offering an exit, and the new investors are pricing a path to $20 billion-plus, which requires Tenstorrent to take meaningful share from Nvidia in inference at scale.

The acquirer was offering a floor, not a ceiling. If the thesis is that open-weight models make proprietary inference chips a commodity, then Tenstorrent is better positioned independent than as a division inside a company that also sells cloud.

This is Amazon admitting that PyTorch is the moat and the chip is the commodity. If TorchNeuron works, AWS can undercut Nvidia on price for every app that does not need cutting-edge performance, which is most of them.

If the chip were the commodity, Nvidia would not still be supply-constrained. Amazon is spending engineering resources to save margin on a workload that is not yet large enough to matter, and developers will not rewrite for a tool that only works on one cloud.

I think the fact that it is free and downloadable today means the monetization is not figured out yet. Replika had paying users, and if Wabi is better, it should be able to charge more, not less.

Free and downloadable is the distribution model, not the business model. The monetization comes later, and the evidence that the timing is right is that she already built the audience once and knows what they will pay for.