Rising cloud waste turns cost control into a delegation problem

The 29% arrives without a composition. Neither the estimate as relayed nor the argument built on it says how much of that waste is model spend, how much is idle capacity, and what the figure was the year before ^[14]. That gap decides which diagnosis is right. If the increment is AI workloads, whose cost scales with the complexity of a decision rather than with predictable traffic and whose single request fans out into model calls, retrieval, tool use and retries ^[9], then reservations and thresholds were never the right instrument. If it is the same underused compute teams have been watching for years ^[2], the binding constraint is permission to act, not visibility.

One year is not yet a trend, and a survey estimate of waste is a self-report by the people being surveyed. Fair on both counts, and the supporting figure does not shore it up: Datadog's finding that 83% of container costs go to idle resources is a share of container spend, not of the total bill ^[2], so the two numbers cannot be stacked into a larger one. What survives the objection is the mechanism claim, which needs no new figure. Rajput's contention is that the waste persists in environments that already have dashboards, because the gap between knowing and doing is measured in human latency ^[7].

The fastest-moving number in the record is about scope rather than savings. The FinOps Foundation reports 98% of teams now manage AI spend, against 31% two years ago ^[3], a rise of 67 percentage points and roughly a threefold increase ^[12]. Nearly every cost team now owns a line item that static rules were not built for, on Rajput's reading, which is a change in the job description before it is a change in the toolchain ^[9].

In Rajput's design, an agent acts within defined guardrails or escalates to the right owner ^[8], and humans sign off on anything high risk ^[11]. The recoverable saving is therefore bounded by the set of actions someone pre-authorised. The trade-off is blast radius against latency: widen the guardrails and you recover more of the 29% while accepting that a machine will terminate or resize something material without asking; keep every consequential action behind an approval and the queue that created the waste stays intact. Whoever writes that list is making the decision, and it is not the person who owns the dashboard.

The board-deck version is short: waste is up for the first time in five years ^[1], rules-based automation has plateaued ^[5][6], buy autonomy. What it leaves out is a price. Rajput concedes the point himself, that piping every metric and log line through a large model can rebuild the waste it was meant to remove ^[10], and answers it by keeping collection and execution deterministic while invoking reasoning only on anomalies or scheduled reviews ^[11]. That architecture is defensible, but it is not costed, since the record carries no figure for what such a program consumes against what it returns.

Which sets the sequencing. The guardrail list drafted this quarter is the ceiling on what any of this recovers next year, because everything outside it still waits for a human. An organisation that delegates nothing buys better diagnosis at a higher run rate.