A team can make a workload ten times faster by tripling its infrastructure cost. What determines whether that is worth it?

What determines it

Whether the speed crosses a threshold that changes behaviour or unlocks value. Performance improvements are not linearly valuable; they matter at discontinuities:

• A pipeline finishing before the business day starts rather than after is worth a great deal; finishing at 4am instead of 6am, when nobody looks until 9, is worth nothing.
• An interactive query returning in two seconds rather than thirty changes how analysts work; two seconds versus one does not.
• A latency budget being met rather than missed is worth its cost; being met with more margin is not.

The question is which side of a threshold the current state is on, and if there is no threshold nearby, the improvement is probably not worth paying for.

The alternatives that are usually cheaper

Before accepting the cost multiplier:

• Data layout. Partitioning, clustering and file sizes dominate how much a query must read, and fixing layout is frequently a larger improvement than adding compute — at no recurring cost.
• Query and job optimisation, where a small number of shapes usually account for most of the work.
• Doing less: incremental processing instead of full recomputation, filtering earlier, and not producing outputs nobody consumes. A meaningful proportion of scheduled data work produces artefacts nothing reads, and finding them is cheaper than making them faster.
• Caching results, since repetition is the signal that pre-computation pays.

The trade-off to state explicitly

Elasticity makes it trivially easy to spend money. A platform that autoscales to satisfy any workload will do so silently, and the constraint has moved from capacity to cost — a softer constraint that is easier to ignore than a full disk.

The controls that keep it honest: per-workload budgets, timeouts, auto-suspend, and cost visibility at the moment the query is written rather than in a monthly report.

The metric that settles it

Cost per unit of business value delivered — per report produced, per model trained, per customer served — trended over time. Total spend rises with growth and tells you nothing, while unit cost falling as usage grows is the only evidence the work is compounding.