How much internet speed streaming actually needs

One 1080p stream needs roughly 5 Mbps. One 4K stream needs anywhere from 15 to 50 Mbps, depending entirely on which service you ask: Netflix publishes 15 Mbps or higher for Ultra HD, Apple recommends a minimum of 25 Mbps, and Max asks for 50+ Mbps for the best 4K experience. If a speed test already clears those numbers and video still stalls, a faster plan will not fix it. The stall is coming from latency, queueing or Wi-Fi position inside your own house, and none of that appears in the figure your ISP sells you.

What the services actually publish

Every number below is what the service prints on its own help pages, worded as a recommendation rather than a hard floor. Note where the ladders stop: YouTube's runs from 360p to 4K and gives no 1440p or 8K figure at all.

Two rows deserve suspicion. Disney's recommended-speeds article is a help-center shell that will not render for automated retrieval, so its 5 and 25 Mbps come from the page's indexed text rather than from a page anyone quoting it has read. Comparison sites report the same pair, which is reassuring but not independent.

Prime Video is messier. Amazon's help article on live-stream problems recommends a minimum download speed of 1 Mb/s for SD content and 5 Mb/s for HD content. For 4K, the only figure traceable to Amazon sits on a Fire TV hardware page and reads at least 15 Mbps. The 25 Mbps widely attributed to Prime Video comes from comparison sites, not from Amazon, so do not repeat it as Amazon's number.

Apple's guidance has aged in two directions at once. Its current article on 4K, HDR and Dolby Vision recommends a minimum of 25 Mbps. Its third-generation Apple TV specifications page still recommends a sustained 8 Mbit/s for 1080p, 6 Mbit/s for 720p and 2.5 Mbit/s for standard definition. That is hardware guidance for a box Apple discontinued in 2016, and a surprising amount of streaming advice is still quietly anchored to it.

Plan on 25 Mbps per simultaneous 4K stream. Netflix's 15 is the low outlier, Max's 50 the high one, and budgeting around either extreme buys you stutter or an overpriced plan.

The FCC table everyone copies says it is not measured

Most per-activity charts trace back to one source, the FCC's Broadband Speed Guide: 1 Mbps for general browsing and email, under 0.5 Mbps for online radio and VoIP, 3 to 4 Mbps for SD video, 5 to 8 Mbps for HD video, 25 Mbps for Ultra HD 4K, 1.5 Mbps for an HD video call and 4 Mbps for online multiplayer.

Then read the disclaimer printed beside that table. The agency calls these "rough guidelines" that "are not based on surveys or experiments conducted by the FCC," and sends households running several devices at once to a separate Household Broadband Guide instead. Any article presenting a tidy Mbps grid as an empirical requirement is dressing up numbers the issuing agency declines to call measurements.

Buffering is usually latency, not throughput

Apple's WWDC 2022 session on network quality put numbers on the gap. Going from 1 to 2 Mbps reduced page load time by almost 40 percent, but after 4 Mbps each incremental increase produced almost no improvement. Latency behaved nothing like that. For every 20 millisecond decrease, load time improved linearly.

The same session walks through a rebuffering case with no capacity shortfall in it. Skipping ahead in a video caused several seconds of rebuffering because oversized buffers, 4 MB at TCP, 256 KB at TLS and 4 MB at HTTP, left fresh packets stuck behind stale ones. Shrinking those to 128 KB, 16 KB and 256 KB fixed it immediately, with nothing about the connection changed, and moved Apple's responsiveness score, measured in round trips per minute, from poor to high.

That failure mode is bufferbloat. Jim Gettys coined the word on 3 December 2010, and the reference write-up is the piece he published with Kathleen Nichols in ACM Queue in November 2011.

Cloudflare's Aggregated Internet Measurement score, published on 18 April 2023, works from the same premise. Its streaming scenario weights download bandwidth alongside unloaded latency, packet loss and the delta between loaded and unloaded latency, which is bufferbloat measured directly. Its video-calling scenario drops bandwidth entirely. Thresholds are specific: packet loss above 5% scores zero and below 1% the full ten, while jitter above 20 ms scores zero and below 10 ms ten.

The case study in that post is worth the detail. A worker sat in the office farthest from the wireless access point and kept dropping video calls. Moving the access point toward the middle of the office lifted the streaming score from 25 out of 70 to 45, and the video-call score from 15 out of 50 to 35. Nobody called the ISP.

The delivery industry separates these causes in its own metrics. Conviva's Connection Induced Rebuffering Ratio counts only non-seek rebuffering, treating a stall as connection-induced only when it begins 10 seconds or more after a seek event. An industry that needs a 10 second exclusion window has already conceded that rebuffering has more than one cause.

Codec efficiency moved the numbers in 2025

Advice written before 2025 overstates what Netflix needs. Its engineering blog reported on 1 December 2025 that AV1 accounts for approximately 30% of all Netflix streaming and uses one-third less bandwidth than both AVC and HEVC, while scoring 4.3 VMAF points above AVC and 0.9 above HEVC. AV1-HDR10+ launched in March 2025 and Film Grain Synthesis was productized that July. One Film Grain Synthesis comparison sets a regular AV1 stream at 8274 kbps against an FGS stream at 2804 kbps that Netflix calls clearly better quality, a 66% reduction. Hardware is not the obstacle: across 2021 to 2025, 88% of large-screen devices submitted for Netflix certification supported AV1.

The same post reports 45% fewer buffering interruptions on AV1 sessions. A compression change cut buffering, and nobody's internet plan got any faster.

YouTube publishes a bitrate ladder too, and it gets misquoted constantly. Its recommended values, 8 Mbps for 1080p SDR rising to 35 to 45 Mbps for 4K and 80 to 160 Mbps for 8K, are upload targets written for creators. YouTube re-transcodes every upload into its own adaptive delivery ladder and does not publish those playback bitrates, so the upload table is not what YouTube sends your TV. It does establish one thing worth carrying across: HDR costs more at the same resolution, 8 to 12 Mbps for 1080p SDR against 10 to 15 Mbps for 1080p HDR, with 480p and 360p HDR not supported at all. A flat number per resolution, silent on HDR and frame rate, is already stale.

Working out how many streams your plan supports

No service, ISP or regulator publishes a combined method, so here is the arithmetic. Start from usable throughput at the device, not the figure on your bill, since a shared 2.4 or 5 GHz network delivers well below the advertised rate once walls and interference are involved. Divide by the highest per-stream bitrate in play, then cap that at the service's own device limit.

Take a 100 Mbps plan delivering about 70 Mbps at the TV. That covers four 15 Mbps Netflix 4K streams, or two 25 Mbps Apple 4K streams plus an HD stream, or exactly one Max 4K stream at 50+. Bandwidth is rarely the binding constraint in a normal household. The plan cap usually is. Netflix ties resolution and screens to tier: Standard with ads and Standard each give 1080p on 2 supported devices at a time, Premium gives 4K plus HDR on 4, and Netflix states it does not limit how many devices you use, only how many watch at the same time, 1 to 4 screens by membership. A family on Standard with a 500 Mbps line still gets two streams.

So buy the cheaper plan unless three or more people genuinely watch 4K at once, and put the savings into a better-placed access point or a wired run to the main TV.

Audio barely registers, with one exception

Spotify runs about 24 kbps at Low, 96 at Normal, 160 at High and 320 at Very High, all AAC, plus a Lossless tier equivalent to 24-bit/44.1 kHz FLAC. Web playback is capped separately, at AAC 128 kbps on Free and 256 kbps on Premium. Apple Music encodes most of its catalog in ALAC from 16-bit/44.1 kHz up to 24-bit/192 kHz. Apple's lossless page states no kbps figure anywhere, so the 256 kbps widely attached to standard-quality Apple Music is historically cited rather than currently published.

Tidal is the exception, and its own numbers do not quite agree. It lists High as 16-bit/44.1 kHz FLAC at up to 10 MB per minute and Max as HiRes FLAC up to 24-bit/192 kHz at up to 40 MB per minute. Convert those: 10 MB per minute is about 1.3 Mbps, but 40 MB per minute is about 5.3 Mbps, more than Netflix asks for 1080p video. Tidal still states that High and Max need only 1.5 to 2 Mbps, which holds only if 40 MB is a peak at the top sample rates rather than a sustained rate. Assume the higher number on a weak link.

What to check when a stream stalls

Move the access point, or move to the same room as it, or plug in over Ethernet, and retest before changing anything else. That alone moved two Cloudflare scores 20 points each. Next, check whether the stall lines up with someone uploading, because a saturated uplink inflates latency for every device in the house, downstream video included. Check the service's device cap before blaming the connection, since two concurrent streams on Netflix Standard is the documented limit rather than a fault. Then measure latency under load instead of idle ping. Idle latency is what a plain speed test reports; the loaded figure predicts rebuffering. Aim for packet loss under 1% and jitter under 10 ms.