Your download speed is slow because the end-to-end path between a server and your device is only as fast as its slowest link, and that bottleneck is usually local: Wi-Fi contention, a congested 2.4 GHz band, a duplex-mismatched Ethernet link, a saturated uplink, or a router whose NAT table is full. ISP provisioning and peering congestion cause a minority of real-world slowdowns. TCP congestion control, DNS latency, and server-side rate limits fill in the rest.
The mechanism: throughput is set by the narrowest pipe and the protocol that fills it
Data moves from a content server to your device through a chain of links. Each link has a capacity in megabits per second. End-to-end throughput cannot exceed the smallest capacity in that chain. A 1 Gbps fiber plan feeding a laptop on a 2.4 GHz Wi-Fi connection that negotiates 144 Mbps will deliver roughly 144 Mbps at best, minus overhead.
The protocol layer matters as much as the pipe. TCP (Transmission Control Protocol) uses congestion control to avoid overwhelming the network. It starts with a small congestion window, grows it as acknowledgments return, and backs off when packets drop. On a link with round-trip latency of 30 ms, a single TCP stream typically reaches full throughput within a second or two. On a link with 150 ms latency, that ramp takes longer and recovers more slowly from loss. This is why a speed test to a nearby server can read 900 Mbps while a download from a distant origin server crawls.
Packet loss is the multiplier most users miss. TCP treats loss as a congestion signal. At 0.1 percent loss, throughput on a 100 Mbps path barely moves. At 1 percent loss with 100 ms round-trip time, a single stream can fall below 10 Mbps regardless of link capacity. The pipe is fine; the protocol is throttling itself.
Plain language: your connection is a chain of pipes, and the data flow obeys the narrowest pipe and how aggressively the sender is allowed to fill it.
Wi-Fi adds a shared-medium problem on top. In the 2.4 GHz band, only three channels (1, 6, and 11) avoid overlap in North America. Every neighboring router, microwave, and Bluetooth device competes for airtime. A channel that looks strong on signal bars can still deliver poor throughput because airtime is shared.
Quantified context: what "slow" actually means and where the bits go
A 100 Mbps connection delivers about 12.5 megabytes per second in ideal conditions. A 25 GB game download at that rate takes roughly 33 minutes. At a measured 8 Mbps, the same file takes about 7 hours. That gap is the practical cost of a bottleneck you can often find and fix.
Overhead consumes a slice of every link. TCP and IP headers, Ethernet framing, and Wi-Fi management frames typically cost 5 to 10 percent of raw link rate. A "1 Gbps" Ethernet link commonly delivers 940 to 950 Mbps in a speed test. Seeing 940 on a gigabit plan is normal, not a fault.
Latency and throughput interact through the bandwidth-delay product. A path's capacity to hold in-flight data equals bandwidth times round-trip time. A 100 Mbps path with 100 ms round-trip time holds about 1.25 MB in flight. If the TCP window is smaller than that, throughput caps below link speed. Modern stacks use window scaling to avoid this, but older devices and middleboxes can still clamp it.
DNS resolution adds a fixed delay before any download starts. A slow or overloaded DNS resolver can add 100 to 500 ms per new domain. Page loads touching 50 domains can lose several seconds to resolution alone.
Common measured throughput ranges by connection type:
| Connection type | Typical real-world download | Main limiter |
|---|---|---|
| 2.4 GHz Wi-Fi, congested band | 5–50 Mbps | Airtime contention, interference |
| 5 GHz Wi-Fi, good signal | 200–800 Mbps | Distance, channel width, client radio |
| Wi-Fi 6E, 6 GHz, close range | 500–1,500 Mbps | Client adapter, router CPU |
| Gigabit Ethernet | 900–950 Mbps | Protocol overhead, disk speed |
| Powerline adapter | 20–200 Mbps | Home wiring quality |
| Mobile hotspot (LTE) | 5–40 Mbps | Signal, tower load, data priority |
Edge cases: when the bottleneck is not where you think
The router's NAT table is a finite resource. Consumer routers track every active connection. BitTorrent, many browser tabs, and IoT devices can push thousands of simultaneous flows. When the table fills, new connections stall or fail, and throughput collapses even though the WAN link is idle.
Bufferbloat punishes interactive use more than raw speed tests. When a router's queue is deep and unmanaged, a large download fills the buffer, adding hundreds of milliseconds of latency to every other packet. Speed tests still read high; video calls and gaming degrade badly. A router with active queue management (AQM) or SQM keeps latency bounded.
Duplex mismatch on wired links silently halves throughput. If one end of an Ethernet link negotiates full duplex and the other half duplex, collisions and retransmissions appear. The link "works" but delivers a fraction of its rated speed.
The counterintuitive detail: a faster Wi-Fi connection can make downloads slower in a crowded environment. Wider channels (80 MHz or 160 MHz) occupy more airtime and are more likely to overlap with neighbors, increasing retransmissions. A narrower, cleaner channel sometimes wins.
Red flags that point to a specific fault:
Throughput collapses only in the evening: upstream congestion or peering saturation, not your equipment. Speed tests are fast but real downloads are slow: server-side rate limits, CDN routing, or DNS. Wired is fast, Wi-Fi is slow at close range: channel congestion or a legacy client radio. Speeds degrade after hours of uptime: router memory leak, NAT table exhaustion, or thermal throttling. One device is slow while others are fine: that device's adapter, drivers, or background sync.
Misconceptions: bandwidth, speed tests, and the "ISP throttling" story
Misconception one: more bandwidth always means faster downloads. Bandwidth sets the ceiling. Latency, loss, and server limits set what you actually get. A 1 Gbps line with 2 percent packet loss can underperform a clean 100 Mbps line for single-stream transfers.
Misconception two: speed tests are ground truth. Speed tests connect to deliberately nearby, well-provisioned servers and often use multiple parallel streams. They measure peak capacity, not the performance of the specific server you care about. A test reading 500 Mbps and a real download at 30 Mbps are both "correct" measurements of different things.
Misconception three: the ISP is throttling you. Some ISPs do manage traffic, but most measured slowdowns trace to Wi-Fi, in-home wiring, or the remote server. Before assuming throttling, test wired at off-peak hours against multiple destinations.
Misconception four: Wi-Fi signal bars equal speed. Bars show signal strength, not airtime availability, interference, or channel width. A strong signal on a congested channel delivers poor throughput.
Misconception five: restarting the router is a fix. A reboot clears the NAT table and any memory leak, which is why it often helps temporarily. If the slowdown returns within hours, the underlying cause remains.
Practical actions: a diagnostic sequence with thresholds
Work from the inside out. Each step isolates one link in the chain.
Step one: test wired directly to the router or modem. Use a Cat5e or Cat6 cable. If wired throughput is near your plan's rated speed (within 10 percent), the WAN side is healthy and the problem is local. If wired is also slow, the fault is upstream.
Step two: test Wi-Fi on 5 GHz or 6 GHz within 10 feet of the router, line of sight. If throughput jumps above 200 Mbps, distance or interference caused the original slowdown. If it stays low, check the client adapter's negotiated rate in the OS network settings.
Step three: run a latency test under load. Start a large download, then ping a nearby server. If latency rises above 100 ms or varies by more than 50 ms, bufferbloat is present. Enable SQM or AQM on the router if it supports it.
Step four: check for packet loss. A sustained loss above 0.5 percent on a wired link indicates a cable, port, or upstream fault. Swap cables and ports before contacting the ISP.
Step five: change DNS to a fast public resolver and retest page loads. This will not change raw download throughput but often fixes perceived slowness on many-site pages.
Step six: for sustained large downloads, use a wired connection and a download manager with multiple connections. Parallel streams route around per-connection TCP limits and server-side per-stream caps.
Thresholds to reconcile: below 25 Mbps, most single-stream work is bandwidth-limited and any fix helps. Between 25 and 200 Mbps, latency and loss matter more than raw bandwidth for interactive use. Above 200 Mbps, the bottleneck is almost always Wi-Fi, the client device, or the remote server. These thresholds overlap because the same connection can be bandwidth-limited for one task and latency-limited for another.
Limitations: what this cannot tell you and where measurement stops
No remote diagnosis can identify your specific fault. The chain has too many links and too many vendor-specific behaviors. The mechanisms here narrow the search; they do not replace measurement on your own equipment.
Speed test results depend on the test server, the number of streams, and the time of day. A single test is a sample, not a verdict. Run three tests at different times against different servers before drawing conclusions about your ISP.
Encrypted DNS, VPNs, and carrier-grade NAT add layers that obscure the path. A VPN routes traffic through an additional hop with its own capacity and latency, and it can reduce throughput even when your local link is healthy.
Some slowdowns are outside your control: peering congestion between your ISP and a content network, server-side rate limits, and regional routing. These are real and common, but they are also the last explanation to reach for, not the first.
This article is general educational information about network performance, not a diagnosis of your specific connection or equipment. For persistent faults, contact your ISP or a qualified network professional with your measurements in hand.