The unusual cost profile is actually borne by us rather than the end users, as we've already generated all 2024 embeddings and we're starting to do all of 2025. We have some going back to 2017 and will be gradually back-filling as we have compute availability.
This approach also has another interesting side effect. If models built on the embeddings would otherwise have been much larger and required significantly more compute then as the number of Tessera users scales the total compute saved can offset the compute required for centralised training and inference.
Thanks for listening and your question.
The unusual cost profile is actually borne by us rather than the end users, as we've already generated all 2024 embeddings and we're starting to do all of 2025. We have some going back to 2017 and will be gradually back-filling as we have compute availability.
This approach also has another interesting side effect. If models built on the embeddings would otherwise have been much larger and required significantly more compute then as the number of Tessera users scales the total compute saved can offset the compute required for centralised training and inference.