DeepSeek-style agent infrastructure ran 1.3B sandboxes in 4 weeks, peaking at 170K concurrent

teortaxesTex · x · 2026-10-06

Working from Beam platform disclosures, the author estimates the scale of sandbox-based code agent infrastructure: 1.3 billion sandboxes over 4 weeks, 48.46M per day, with peaks of up to 170K concurrent sandboxes.
That concurrency implies a sandbox lifetime of roughly 3-5 minutes, and many may be short-lived "grading sandboxes"—unclear whether DeepSeek even counts those in its stats.
Cross-referencing DeepSeek's own figures: one DSec-scale unit does 3M, so it would need 16 units at 656 GB300s per unit (though DeepSeek likely uses ≤128 GB300-equivalents per unit). A rare glimpse into frontier-scale agent training infrastructure.

More from coding & agent