It seems to me the ROI on the research may be diminishing. If they aren't turning a profit with the infrastructure already deployed and the models already trained then buying more GPUs and spending billions more on training doesn't seem like it would improve the correct side of the balance sheet.