Show me the line item — and absent one, show me the cost mechanism. Microsoft's grant US11526679B2 (“Efficient transformer language models with disentangled attention and multi-step decoding,” issued 2022-12-13) is a cost mechanism. Assigned to Microsoft Technology Licensing, LLC, with inventors including Jianfeng Gao, it targets the efficiency of the transformer architecture that powers modern language models.
The architecture detail matters because transformers are expensive by default. “Disentangled attention” and “multi-step decoding” are techniques to get strong language performance for less compute — the DeBERTa line of work that Microsoft Research is associated with. Cheaper-per-token is, for a company selling language AI, directly a margin question.
Microsoft routes its AI revenue through cloud and productivity segments and does not isolate model-architecture economics, which is normal. The grant is the technique-level record under the segment story: dated 2022, owned, and aimed at the efficiency of the dominant language-model architecture.
For a disclosure-first reader, the value is grounding. When Microsoft talks about scaling AI services profitably, the profitability rests partly on architectural efficiency like this. The patent is the primary document that the efficiency work is real and dated, even though its dollar effect is folded into aggregate results.
House rule, restated: a grant proves invention and ownership, not revenue. We attribute no figure. The claim is that the IP making transformer language models cheaper to run was held and dated by Microsoft in 2022.