$450 Billion in AI Infrastructure, Most of It Wasted: ApexData Launches RidgeScope

This is less a product launch than a pricing signal for the AI stack: if operators can systematically surface wasted compute, the market will start underwriting AI capex as a controllable operating lever rather than a sunk cost. Near term, that is margin-positive for hyperscalers and the largest model builders because even modest efficiency gains can translate into meaningful capacity without incremental GPUs; the budget owners who win are MSFT, AMZN, GOOGL and other platforms that can absorb the software into existing spend.

The second-order loser set is the low-quality end of the GPU-rental ecosystem: neoclouds, overloaded cluster operators, and anyone selling “compute” without provable throughput. Once buyers can benchmark waste, they will push harder on SLAs, credits, and utilization audits, which should compress pricing power and raise churn for weaker operators. For the semiconductor complex, the read-through is mixed: better visibility can initially support orders by exposing bottlenecks, but over 6-18 months it raises the risk that the market starts discounting a larger installed base as underproductive rather than scarce.

Consensus may be too quick to call this bearish semis. In the first 1-3 months, optimization tools usually unlock more spend, not less, because teams reinvest savings into more experiments and bigger runs. The real contrarian risk is a later-cycle narrative shift: if multiple hyperscalers publicly tie AI ROI to utilization discipline, the market could rotate from "build more" to "prove productivity," which would pressure high-multiple GPU suppliers and reward software/infrastructure names. Falsifiers are simple: capex guides that keep rising despite better tooling, or no evidence of utilization improvement in next-quarter commentary.