Adam Sioud

A sandbox is managed software wrapped around compute capacity. To see both layers, I track one exact four-vCPU, 8-GiB public VM offer from Akamai Linode, Vultr, Scaleway, Microsoft Azure, Amazon EC2, OVHcloud, and Oracle Cloud, then compare their median with the public processor-and-memory rates of managed sandboxes at the same requested shape. These are observed offer rates, not transaction prices, capacity guarantees, or invoices.

Both lines use fixed membership, with the median as the headline and the 25th–75th percentile as the band. The seven-vendor VM series starts with its first complete hourly check; the earlier four-vendor evidence remains in the dataset but is not relabeled as seven-vendor history. Every complete check is retained even when prices are unchanged. Regions, CPU models, tenancy, burst policy, storage, networking, lifecycle management, and delivered work still differ. VM storage treatment is shown row by row. The sandbox comparison covers processor and memory only. See the sources and dated changes.

Hourly rates still do not tell us what a completed job costs. The public StarSling HPC Sandbox Benchmark ran ten pinned Better Auth development workload on fresh sandboxes. For each complete job, I multiply its measured phase seconds by that service's matching public processor-and-memory rate. The result is an estimated cost for the same pinned software work on six services.

The latest comparable batch has 72 provider-and-job slots. Sixty-nine contain all ten measurements and appear below; three incomplete slots remain missing rather than estimated. The estimate is: measured seconds ÷ 3,600 × the matching public hourly rate. It covers processor and memory only. It is not an invoice and excludes startup, teardown, retries, storage, network, credits, and minimum billing increments.

Measured time remains underneath the estimate as auditable evidence, not the headline. The primary chart ranks cost and retains every complete job plus each service median and middle 50%. Seven source batches over five calendar days remain in the audit record, but their six harness revisions are not joined into one false performance history.

An H100 GPU-hour, a four-vCPU VM/VPS hour, and a managed sandbox hour are different products. Their raw dollar prices do not belong on one linear axis, so the next chart rebases each series to 100 at its own first retained observation. H100 uses the first print with at least ten providers; VM/VPS starts when the first complete seven-vendor check exists; sandbox uses the rate in force at the H100 starting print.

H100 and VM/VPS are observed hourly. The sandbox line steps only when a reviewed public rate changes. Select a series to bring its price-dispersion band forward while the other median lines remain visible. This compares direction and dispersion, not absolute cost, traded volume, or executed transactions. Absolute VM/VPS and sandbox prices remain in the first chart above.