Download Time Calculator

In 2026, a game-studio launch manager pushing a 65 GB title to 500,000 Steam pre-loaders on day-of-release needs to know not the marketed 1 Gbps fiber line, but the realistic ETA each customer will experience. The headline number sells plans; this calculator estimates what users actually feel. The gap — usually 12-30% — comes from a stack of small physics-and-protocol penalties: TCP slow-start, TLS handshake, HTTP/2 framing, IP/TCP/Ethernet headers, retransmissions, congestion control back-off, and the slowest hop in the path.

The Internet Engineering Task Force (IETF) standardised TCP in RFC 793 (1981) and refined congestion control after the 1986 Internet Collapse — Van Jacobson and Michael Karels documented slow-start and AIMD (Additive Increase Multiplicative Decrease) in 1988. Modern Linux kernels default to BBR (Bottleneck Bandwidth and Round-trip propagation time), authored by Neal Cardwell's team at Google in 2016, which converges on the path's real bandwidth-delay product rather than guessing via packet loss. BBR-enabled servers (YouTube, Spotify, Cloudflare) reach line rate faster — but the first 1-2 seconds of any new TCP connection still under-utilises bandwidth.

HTTPS adds its own cost. TLS 1.3, standardised in RFC 8446 (Eric Rescorla et al., 2018), reduced the handshake to 1 round-trip from TLS 1.2's 2 round-trips, but still imposes AES-GCM encryption on every byte. HTTP/2 (RFC 7540, 2015) adds 9-byte frame headers per stream; HTTP/3 over QUIC (RFC 9114, 2022) replaces TCP entirely with a UDP-based protocol that eliminates head-of-line blocking. Cloudflare, Google, and Facebook all serve significant fractions of traffic over HTTP/3 in 2025. Pages that toggle the "Raw" protocol option above approximate QUIC-class overhead (~4%) rather than HTTPS (~15%).

File-size accounting is the other major source of confusion. The IEC 80000-13 standard (2008) formalised KiB / MiB / GiB (binary, 1024-based) versus KB / MB / GB (decimal, 1000-based), but operating systems disagree. Windows shows GiB but labels them "GB"; macOS and iOS use decimal GB; Linux file systems vary. Hard-drive manufacturers always use decimal GB (so a "1 TB" drive holds only ~931 GiB when Windows formats it). This calculator uses the common convention: size inputs interpret 1 GB = 1024 MB (binary), and bandwidth uses 1 Mbps = 1,000,000 bits/sec (decimal) — matching how ISPs and OSes report each.

The bandwidth dial here uses a log scale because the realistic range of consumer connections spans four orders of magnitude — 3 Mbps (3G mobile) through 10,000 Mbps (XGS-PON business fiber). Linear scales squash the lower tiers; log scales give equal visual real estate to each decade. Akamai's State of the Internet report (last published 2017 before going subscription) showed global average broadband speeds growing ~25% YoY through the 2010s. Ookla's Speedtest Global Index for Q1 2026 lists 256 Mbps median fixed broadband globally, 95 Mbps median mobile.

For ML engineers downloading model weights, the 70-billion-parameter Llama 3 at Q4_K_M quantisation is ~40 GB — at 100 Mbps with 85% efficiency that's ~56 minutes; on a 1 Gbps line ~5.6 minutes. The Hugging Face Hub, where most open-weight models are distributed, uses LFS (Large File Storage) on top of git; their CDN (Cloudflare R2 in 2025) typically negotiates 80-92% of consumer-link speed. Tools like huggingface-cli download resume on failure and parallelise via multipart-range requests — bumping the effective efficiency to ~95% of raw bandwidth.

Finally, the slowest hop wins. A 1 Gbps fiber line into a router connected via 100 Mbps Ethernet pegs at 100 Mbps. A WiFi 5 link to a phone in a far bedroom averages 80 Mbps regardless of fiber speed. CDN edges (Cloudflare, Akamai, Fastly) try to keep the last-mile the bottleneck — but a small percentage of paths cross congested peering links that throttle to 50-200 Mbps regardless of plan. The ETA in the lantern dial above is an honest estimate, but reality is path-dependent — your mileage may vary by ±20%.