What did the filing actually say? DeepMind's grant US10936949B2 (issued 2021-03-02) says: pick the training tasks that make the model learn fastest, and you waste fewer expensive compute cycles. Assigned to DeepMind Technologies Limited and classified CPC G06N 3/08, it is training-efficiency IP from the lab whose parent, Alphabet, foots the compute bill.
The economics are stark. Training a frontier model is the largest discrete expense in the AI business — a single run can consume an enormous slice of a quarter's infrastructure. Anything that gets equivalent capability from fewer training steps is, in effect, a cost-reduction technology with the same dollar logic as a cheaper supply contract.
Alphabet's disclosures fold this into aggregate research-and-development and infrastructure investment; there is no “training efficiency” line in the 10-K, and we don't invent one. What the patent provides is a dated, owned record that the efficiency of the most expensive step in the pipeline was an explicit research target in 2021.
For a markets reader, the value is in connecting the cost narrative to a primary document. When Alphabet talks about disciplined AI investment, the discipline is partly technical — it lives in patents like this that try to make each training dollar buy more learning. The filing is the evidence the discipline is real, not rhetorical.
House rule: a grant is invention and ownership, not revenue. We attribute no figure to it and note that efficiency gains are notoriously hard to isolate in aggregate financials. The defensible claim is narrow — the training-cost lever was being patented at DeepMind in 2021.