AWS Changed GPU Cost Reports. Count the Renders That Survive Review.

A GPU instance can run all afternoon and produce clips that nobody can use. AWS’s November 6 update makes that afternoon easier to account for. Cost Explorer and Cost and Usage Report (CUR) 2.0 now amortize EC2 Capacity Block fees across the reservation, showing costs associated with launched instances and unused capacity. For an experiment team, the next step is to connect that spending to the outputs that passed review. Keep the rejected attempts in the calculation. AWS announcement

AI generated and fact-checked this article against the linked public sources. It was prepared through an automated workflow, and Keith has not reviewed it.

Stereo Depth Lab gives this distinction a concrete shape. The public case study describes a working local prototype that estimates relative depth and exports stereo images or video. It offers depth warping, point splats, and a lightweight persistent-Gaussian rendering path. Video continuity remains a research problem, and a heavier reconstruction backend is future work. The cloud experiment below is hypothetical. The documented tool runs locally. Stereo Depth Lab case study

Put the rejected render beside the bill

Suppose a team reserves a block and spends $240 on its reservation fee. This is an invented accounting example. Of that fee, the report attributes $180 to instance usage and $60 to unused capacity. Twelve clip-rendering attempts finish, and four meet the team’s visual acceptance criteria.

Dividing the whole $240 by twelve gives $20 per completed attempt. Dividing it by four gives $60 per accepted output. Both describe the same batch. If someone reports only the $180 associated with running instances, the unused share disappears from the experiment’s cost. Keep the full reservation fee in the batch total, with its used and unused portions shown underneath. Add storage, transfer, operating system, and other charges separately when they apply. AWS’s billing guide explains that the reservation fee is paid upfront and operating system usage can bring additional charges. Capacity Blocks billing guide

Define acceptance before starting the batch. For a stereo video experiment, the review might check whether edges around moving subjects remain tolerable across frames, whether newly exposed areas are usable, and whether the final output meets the requested resolution. Keep the source clip and review conditions fixed when comparing two methods. Record which criterion rejected a result so the team can distinguish a rendering defect from an export failure.

A rejected experiment can still answer a valuable research question. Record that finding alongside its cost. It belongs in the research budget even when its output never enters a delivery folder. If the batch produces zero accepted outputs, report the total spend and zero accepted outputs directly. There is no meaningful cost-per-accepted-output number to calculate for that batch.

Give each attempt an identity

For a cloud pilot, write a run record when the worker starts. Include the experiment ID, source version, rendering configuration, instance ID, reservation ID, and start time. On completion, add the end time, output location, and execution status. Review can then append an acceptance result and a reason. A process exit code establishes one part of that record. The visual decision arrives later.

Use the reservation and instance identities to connect billing periods to the experiment. An instance can spend time loading a model, waiting for input, or retrying a failed export. Account for those periods in the batch total. For a shared instance, choose and document how overlapping jobs share its cost, and reconcile the allocations to the billed total. A small first pilot can avoid that allocation problem by assigning the block to one experiment.

There is also a deadline to include in the plan. AWS documents that running instances begin terminating before the Capacity Block ends, with a 30-minute cleanup interval for instance blocks and 60 minutes for UltraServers. Finish and export results with enough margin for that boundary. Test that the saved outputs remain retrievable after the worker is gone. How Capacity Blocks work

At the end of the pilot, the team should be able to open an accepted clip, find its run record, and reconcile the experiment’s used and unused reservation costs. That gives the next research decision a price grounded in what the batch actually produced.