A chip, a cluster, or an inference server gets filed when we can name the vendor, the spec, and the date. A rumor about capacity does not. When a model post includes the machine it runs on, as EmbeddingGemma 2 does with on-device memory, that fact is recorded with the model and can be cross-filed here once it changes how teams buy or host compute.
This shelf is a frame until the next primary document. We would rather leave it quiet than fill it with secondhand capacity charts.