At theoretical 10 Gbps line rate, one decimal terabyte takes thirteen minutes and twenty seconds. The current calculator estimates 14.9 min one way using its 12% planning-overhead assumption. Reaching that range requires the source and destination to sustain roughly 1.1–1.25 GB/s, plus a tuned, clean path.
On a 10 GbE LAN, storage is often the first ceiling. Across a WAN, latency, packet loss, host buffers, site uplinks, and remote service limits can reduce throughput dramatically. “10 Gbps” describes an interface or service boundary, not the complete transfer.
Separate theoretical, modeled, and observed results
One decimal TB is 8 × 10^12 bits:
8 × 10^12 bits ÷ 10 × 10^9 bits/s = 800 seconds
That theoretical result excludes protocol and operational work. The calculator's 12% allowance multiplies the payload by 1.12, producing the component-backed result displayed above. The allowance is a planning assumption, not a law or measured 10 GbE efficiency.
Observed time is the actual result for a named source, destination, path, and workload. Capture all four; a best-case LAN test should not become a promise for an intercontinental delivery.
The storage system must feed 10 GbE
At line rate, 10 Gbps is 1,250 decimal MB/s. Under the current planning model, the implied payload rate is about 1,116 MB/s. The source must read and the destination must write continuously near that range.
The Seagate Exos X18 datasheet lists a maximum sustained transfer rate of 270 MB/s for that enterprise hard drive. Four drives at that rating total 1,080 MB/s and five total 1,350 MB/s before RAID, controller, filesystem, workload, and inner-track effects. Therefore “four or five drives” is only first-order arithmetic, not an array design or performance guarantee.
A SATA SSD may also fall below the target, especially for sustained writes after cache exhaustion. A capable NVMe device or striped array can exceed it for favorable sequential work, but thermal behavior, queue depth, filesystem, free space, encryption, and simultaneous activity matter. The destination has to match the source.
Test the real media with representative files. One large file exercises sequential throughput; hundreds of thousands of files add metadata and request work. Include any packaging and extraction time in the workflow, not only network time.
The host and bus can become ceilings
Confirm the entire local path:
- a 10 Gbps-capable NIC negotiated at the expected rate;
- a suitable PCIe lane configuration;
- current drivers, firmware, and offload settings;
- enough CPU and memory for protocol, encryption, hashing, and application work;
- storage connected through a bus that can sustain the target;
- no 1 GbE switch port, firewall, uplink, or dock in the route.
ESnet's Data Transfer Node guidance emphasizes balanced CPU, NIC, storage, and external connectivity. A 10 Gbps workstation behind a 1 Gbps site uplink is still a 1 Gbps internet endpoint.
SMB Multichannel and SMB Direct can improve appropriate managed Windows environments by using multiple connections or RDMA-capable hardware. They require compatible systems and deliberate administration. Do not disable SMB signing, encryption, or endpoint security simply to improve a benchmark; evaluate required controls and performance together.
WAN 10 Gbps is not LAN 10 GbE
Within one building, low latency and a clean managed network make near-interface performance plausible when endpoints keep up. Across distance, a high-bandwidth path needs substantial in-flight data. ESnet's host-tuning guidance estimates roughly 120 MB of buffer to expect a single stream to fill 10 Gbps over a 100 ms round trip.
Loss is especially costly. In one ESnet research-network observation, 0.0046% packet loss on a 90 ms path coincided with 490 Mbps in the lossy direction versus 8.2 Gbps in the clean direction. That is an observed case illustrating sensitivity, not a universal ratio or expected internet result.
Use multiple well-managed streams where the application and service support them, but do not assume more concurrency always helps. Eventually CPU, storage, gateway, request, or account limits become the bottleneck.
Cloud ceilings and object limits are separate
Cloud services can have more headroom than one workstation while still imposing scoped limits. Azure documents standard storage-account defaults of 60 Gbps ingress and 200 Gbps egress in listed regions, and 25/50 Gbps elsewhere. Those are regional account targets for specified configurations, not guaranteed throughput.
Object size is another scope. For current service versions, Azure block blobs support about 190.7 TiB. Google Cloud Storage caps a completed object at 5 TiB. Since December 2025, Amazon S3 supports up to 50 TB by multipart upload, expressed as 48.8 TiB in its quota table; AWS separately limits a single GET to 5 TB and requires ranged or part-based requests above that.
None of these object ceilings says that a 1 TB desktop upload will run at 10 Gbps. Region, quota, object layout, request concurrency, account history, credentials, and the public path remain relevant.
A practical 10 Gbps test plan
- Benchmark source read and destination write separately.
- Test a memory-to-memory or fast-storage LAN path to validate NICs and switches.
- Record negotiated speed, CPU, disk, and errors during a large-file test.
- Repeat with a representative small-file package.
- Test the intended remote region and measure round-trip time and loss.
- Increase concurrency cautiously while watching host and service constraints.
- Verify the received data and preserve logs.
- Plan from repeatable sustained throughput, not the best short interval.
If even an ideal model misses the deadline for a much larger package, compare online delivery with physical media using the shipping-versus-transfer calculator.
Where TeraAirlift fits
TeraAirlift is a Windows-first desktop operations console with queue progress and ETA, cancellation, retry-friendly operations, transfer history, diagnostics, recipient-controlled delivery, and SHA-256 integrity verification. It gives operators evidence during large jobs, but does not promise 10 Gbps throughput or an SLA. The same planning applies to post-production and LiDAR teams.
Sources
- Seagate — Exos X18 datasheet
- ESnet — Data Transfer Nodes
- ESnet — Host tuning background
- ESnet — Packet loss
- Microsoft — SMB file-server performance tuning
- Microsoft — Standard storage-account targets
- Microsoft — Blob Storage scalability targets
- Amazon Web Services — Uploading objects to S3
- Amazon Web Services — Downloading objects from S3
- Google Cloud — Cloud Storage quotas and limits
