Friday Night Wiki
TopicsTagsAbout
Buy me a coffee
Energy and Physical Infrastructure Literacy Tutorial

Telecom and Internet: The Physical Layer

The Cloud Is Somewhere

When you send a message, back up a photo, or ask an AI a question, the bits travel through physical infrastructure. "The cloud" is a metaphor; the actual place your data goes is a specific data centre with specific power and cooling requirements, connected by specific cables to your device.

Understanding this physical reality is useful because:

  • Digital services have real physical footprints
  • Internet outages are physical events (cable cuts, tower failures, power losses)
  • The AI compute build-out is a major infrastructure story
  • National sovereignty over digital services depends on physical sovereignty over cables and data centres

The Last Mile

Your device connects to the internet via some combination of:

Wired home connections

  • Fibre to the home (FTTH): light-transmitted data over glass fibre. Very high speeds (1 Gbps common, multi-Gbps increasingly). The modern standard. Rolling out in most of the developed world, fully deployed in places like Japan, South Korea, much of Scandinavia
  • Cable (coaxial): TV infrastructure repurposed for internet. Fast downloads, asymmetric (slower upload), capacity shared among neighbours
  • DSL (over copper phone lines): older, slower, still widespread
  • Fixed wireless (5G home internet, Starlink): wireless to the home, no wire to the building

Cellular / wireless

  • 2G, 3G, 4G (LTE), 5G: generations of mobile networks. 5G is the current frontier in most markets; 6G in research
  • Bands: low (long range, lower speed), mid (balance), high (mmWave; fast, short range, blocked by walls)
  • Densification: 5G high-band requires many small cells close together; low-band covers long distances

Satellite

  • Geostationary (GEO): one satellite at 36,000 km covers huge area; latency ~250 ms, limits applications
  • Low Earth Orbit (LEO) constellations: Starlink, OneWeb, Kuiper. Thousands of satellites at ~500 km; latency under 50 ms
  • Roles: rural, maritime, aviation, military, disaster response

Wi-Fi

Short-range wireless within buildings. The last few metres of most home internet connections. Standards (Wi-Fi 5, 6, 6E, 7) increase speed and handle congestion better.

The Middle: Regional Networks

Your ISP connects to the wider internet through regional networks:

  • Backbones: high-capacity networks spanning cities, countries, regions
  • Internet Exchange Points (IXPs): where networks peer with each other to exchange traffic
  • Metropolitan area networks: within and between cities

The internet is not a single network; it's a network of networks (hence "inter-net"). Traffic routes through peering agreements between operators. BGP (Border Gateway Protocol) decides paths.

Submarine Cables

The internet's international spine is fibre-optic cables on ocean floors.

  • About 500+ active submarine cables carry >99% of international internet traffic
  • Capacity per cable: modern ones can carry hundreds of terabits per second
  • Failure points: ships' anchors, trawlers, earthquakes, sabotage
  • Repair: specialised cable ships; can take weeks
  • Geopolitics: cable routes are strategic; countries increasingly pay attention

Famous cases:

  • The 2006 Pingtung earthquake damaged cables off Taiwan; affected internet across Asia for weeks
  • Red Sea cable cuts in 2024 affected Europe-Asia traffic
  • Russian submarine activity near cables in Northern Europe is a regular geopolitical concern

The "cloud" is mostly continents connected by thin glass strands under the sea. Worth visualising.

Data Centres

A data centre is a building full of computers, networking, and cooling. Modern hyperscale data centres contain tens or hundreds of thousands of servers.

What's inside

  • Racks: standardised metal frames holding servers, network gear, storage
  • Servers: thousands to millions per data centre
  • Networking: top-of-rack switches, aggregation, routers
  • Cooling: air conditioning, increasingly liquid cooling for high-density AI racks
  • Power: redundant feeds from the grid, UPS (batteries for momentary outages), backup generators (for extended outages)

Power

Data centres are electricity-intensive. A large site can draw 100-500 MW, comparable to a small city.

AI training has amplified this dramatically. A single large GPU cluster can draw 30-100 MW; training runs consume megawatt-hours in a single run. The biggest hyperscalers are now building sites of 1-2 GW capacity.

Site selection factors:

  • Power availability: enough MW available via the grid
  • Power cost: cheap electricity is a competitive advantage
  • Cooling climate: cold regions (Nordics, Iceland) reduce cooling cost
  • Latency to users: close to demand centres matters for user-facing services
  • Connectivity: proximity to major cables and IXPs
  • Land, permits, politics: the usual infrastructure siting concerns

Cooling

Servers generate heat. Cooling is a major energy cost:

  • Air cooling: traditional; blow cold air through server racks. Easy, widespread
  • Liquid cooling: circulate coolant directly to hot components. More efficient for high-density racks. Increasingly common for AI
  • Free cooling: use cold outside air directly (where climate allows)
  • Water cooling via evaporation: efficient but uses water

PUE (Power Usage Effectiveness) measures efficiency: total power used divided by power used by IT equipment. PUE of 1.0 is perfect (all power goes to computing). Modern hyperscale sites achieve 1.1-1.3. Older sites might be 1.5-2.0.

The AI data centre boom

Training large AI models requires enormous clusters. A GPT-4 scale training run uses tens of thousands of GPUs for months. Next-generation models require even more.

Hyperscalers (Microsoft, Google, Amazon, Meta) plus specialists (CoreWeave, Lambda) are racing to build capacity. Individual announced sites exceed 1 GW; some proposals exceed 5 GW.

This level of demand is unusual. In many regions, grid operators are receiving connection requests that collectively exceed existing grid capacity. Permitting new transmission, signing power purchase agreements years ahead, and building adjacent to nuclear plants are all currently happening.

For infrastructure literacy, AI compute is now a significant infrastructure story, not just a software one.

Content Delivery Networks (CDNs)

Serving every page from a single data centre is slow and expensive. CDNs (Cloudflare, Akamai, Fastly, Amazon CloudFront) cache content at "edge" locations close to users:

  • A video on Netflix probably streams from a server in your city, not from Netflix HQ
  • A static website fetches from an edge node, not from origin
  • Many APIs are served through CDN caches

This shortens latency, reduces load on origin servers, and improves reliability. From an infrastructure perspective, CDNs are distributed mini-data-centres bundled with routing logic.

Latency and Physics

Latency (time for a bit to travel from A to B) is bounded by the speed of light.

  • Light in fibre travels at about 2/3 the speed of light in vacuum
  • New York to London round trip: minimum ~28 ms
  • New York to Sydney: minimum ~150 ms
  • You to a GEO satellite and back: ~550 ms (noticeable in calls)

Latency sums: your Wi-Fi to your ISP, ISP to backbone, backbone to CDN edge, CDN to origin (if needed). Each hop adds delay.

Some applications (high-frequency trading, online gaming, cloud gaming, interactive AI) are latency-sensitive. Others (video streaming with buffering, email) aren't. Architecture depends on the requirement.

Wireless Spectrum

Wireless communication uses electromagnetic spectrum: specific frequency bands allocated to specific uses by governments.

  • Low frequencies: propagate well, limited capacity
  • High frequencies (mmWave): huge capacity, short range, blocked by walls
  • Regulations determine who can use what

Spectrum is a finite resource. Auctions (billions of dollars) allocate it to carriers. International coordination keeps uses compatible across borders.

5G uses a mix of low, mid, and high bands; each has different deployment implications.

Reliability

Modern internet is highly available but not invulnerable. Causes of outages:

  • Fibre cuts: construction equipment, natural disasters. Common; the network routes around most. Serious when multiple cables cut simultaneously
  • Power outages: data centres have backup, but long grid failures can exceed backup fuel
  • DNS failures: the system that translates domain names to IPs. A single provider's failure can take down large portions of the visible internet
  • BGP misconfigurations: routing mistakes can black-hole traffic for large regions
  • Software failures: a cascading bug at a major provider (Fastly 2021, Cloudflare, AWS) can affect vast portions of the internet

The internet's redundancy helps but doesn't eliminate these. Users notice when major services go down; infrastructure people notice the physical layer cracking.

Sovereignty and the Splinternet

A growing theme: different countries have different views of the internet.

  • China: Great Firewall; separate ecosystem; much traffic doesn't cross the border
  • Russia: moves toward domestic alternatives; sometimes disconnects from global internet
  • EU: regulation emphasis (GDPR, Digital Services Act, Digital Markets Act) that shapes what services operate
  • US: dominant platform origin; separate concerns about TikTok and Chinese infrastructure
  • Global South: varies; many countries' digital infrastructure is foreign-owned

Submarine cables, cross-border traffic, data centre locations, and chip manufacturing are increasingly geopolitical. "The internet is global" is becoming less accurate.

Common Pitfalls

"The cloud is nebulous." It's specific buildings in specific places. When you hear "we moved to the cloud", you moved to someone else's computer in their building

"Wireless is better." It's convenient; wired is still faster and more reliable in most cases. A fibre-connected desktop beats Wi-Fi for anything latency-sensitive

"5G changes everything." It changes some things (capacity in dense areas, low-latency applications). It didn't revolutionise daily use for most people; the revolution was 4G

"We don't need more data centres." AI says otherwise. Compute demand is growing faster than most predictions anticipated. This has grid consequences

"The internet routes around damage." Mostly true, historically. Modern dependencies (specific CDNs, specific cable routes for specific geographies, DNS providers) create chokepoints that routing doesn't eliminate

Next Steps

Continue to 08-energy-economics.md for the economics that decide which generation gets built and which runs.