Based on Microsoft's apparent losses on CoPilot, I suspect the latter. My initial research on ML usage suggested the thing wasn't profitable when commoditised which it well and truly is now hence why I've avoided investing in it (wisely so far). The energy cost versus capital cost versus actual utility makes it a net negative ROI for the technology supplier. The industry is currently living on some vain hope that hardware improvements will decrease cost enough for it to be profitable. With die and process shrinks starting to get problematic (3nm isn't going as well as people want it to) and increasing energy costs, cost reductions are unrealistic.