What happened
E2B published results from an evaluation of NVIDIA Vera hardware for its isolated agent environments. The company describes sandboxes as microVMs that give agents a shell, filesystem, dependencies and processes; developers manage them through E2B’s SDK. E2B says its runtime resumes snapshots, loading memory as needed, and supports pausing and returning to a session. The evaluation ran October 9–13, 2026, on one pre-production dual-socket Vera server with 176 physical cores and 1,479 GiB of RAM. E2B compared component tests with an AMD EPYC 9B45 reference measured in August.
In a bare-host single-thread compute microbenchmark, Vera reached 7,759 events per second versus 5,131 on the AMD system, or 1.51 times the throughput. The test used Vera’s 64K-page kernel; E2B reports medians of at least three runs. In a separate memory-population test using 4K kernels, 2 MiB huge pages and a 2 GiB region, Vera reached 62.8 GB/s against 17.4 GB/s, or 3.61 times the throughput. E2B says that test measures memory population, not complete sandbox startup.
For a full-stack test, E2B used Embed, its single-node distribution of the open-source runtime, running directly on bare metal. It measured 50 ms warm sequential sandbox starts and 3.8 ms command round trips at low load; those were separate tests, not fleet-load latency figures. In a mixed-load capacity test, each sandbox repeated a CPU burst, a 10 MiB disk write and a pause, with one in ten running continuously. Across four tuned runs, the server sustained 1,400–1,550 loaded sandboxes within the company’s latency and error objectives. At 1,550, p95 command round-trip latency was 255 ms. A separate test in which every sandbox ran continuously reached 550.
Host configuration materially affected the capacity result. E2B says early runs reached 320 sandboxes; with network-neighbor tuning and then a higher device limit, its best fully tuned run reached 1,550. This is a configuration comparison, not a Vera-versus-x86 speedup. The company also reports 22,145 create-and-destroy cycles with zero errors. It plans a multi-node Vera cluster, cold resumes backed by object storage and an x86 comparison using the same E2B packaging and host tuning.
What the archive adds
New Runtime’s earlier coverage on October 22, 2026, described Modal’s benchmark finding that CPU behavior under contention can matter for heterogeneous, bursty agent sandboxes (NVIDIA Vera CPUs double throughput for Modal Sandboxes). E2B’s results are a new benchmark from another sandbox provider on Vera, not an independent replication of Modal’s measurements.
Trend signal
This supports, modestly, agent economics moves to completed work: the practical unit for infrastructure is increasingly how many useful sandbox workloads a host can sustain within latency and error objectives. E2B’s figures are capacity measurements for a specified workload, not cost per completed task; without cost, power and comparable x86 fleet-capacity data, they do not yet establish an economic advantage.
Why now
E2B frames the tests as an evaluation of hardware for its agent cloud and says it will extend the work to a multi-node Vera cluster and cold resumes from object storage. It also identifies a same-packaging, same-tuning x86 comparison as a next step, which would broaden the current single-server evaluation.
Why it matters
Teams building coding agents often run many isolated environments at once, and the host must handle both active computation and the memory work of restoring snapshots. E2B’s results suggest that memory population and single-thread performance are worth testing separately rather than treating processor choice as a single headline number. Just as importantly, the jump from 320 to as many as 1,550 sandboxes shows how runtime caps and host settings can dominate capacity measurements. These results can help teams design their own workload-shaped tests, but they do not show that Vera makes agents 1.5 or 3.6 times faster, or that it beats x86 in equivalent fleet capacity. The evaluation also leaves cost and power open, so it cannot yet answer whether the observed capacity is cheaper or more efficient to operate.
Limits and uncertainty
E2B describes this as a pre-production evaluation on one server. Its component benchmarks compare specific operations, not end-to-end agent speed or complete sandbox startup. Capacity results depend on the tested mix of CPU bursts, disk writes and pauses; continuously busy sandboxes reached a different limit. The company says some restore paths remain faster on x86 references and nested virtualization needs more work. It did not measure cost or power, and its planned same-configuration x86 comparison and multi-node evaluation are not reported here. The results have not been independently checked.
Reader takeaway
Treat the figures as a test plan, not a purchasing verdict. Measure snapshot memory population, command latency and fleet capacity separately, then tune host limits against your own workload mix. Until comparable x86 capacity, operating cost and power data are available, E2B’s benchmark shows Vera’s promise for its runtime, not a general hardware win.