From Vibrations to Photons: How a Signal Travels, and Where AI Will Run

How it works

Speech begins as a change in air pressure. Sound in air is a longitudinal wave: the molecules oscillate about their equilibrium positions while the disturbance moves outward [5]. A microphone turns that pressure into a changing voltage. An analogue-to-digital converter measures the voltage at regular intervals and stores the measurements as numbers, a codec compresses them, and software cuts the result into packets. At the far end a loudspeaker runs the process backwards: the signal current flows through a voice coil that sits in a magnet’s field and is fixed to a diaphragm, so the coil and diaphragm move and push the air [6]. What arrives is the pattern, rebuilt, not the air that left the speaker’s mouth.

Who competes

Microphones, converters and loudspeakers are mature parts made at very large scale. The competition is now in software: codecs, noise suppression and speech models small enough to run on the device itself.

Limits and open problems

Sound fades quickly with distance, echoes, and offers little bandwidth. It can carry data, as dial-up and underwater acoustic modems show, but slowly. Every stage adds some noise or distortion, so the far end hears a reconstruction that is close enough to understand, not a perfect copy.

How it works

Radio waves, microwaves, infrared, visible light, ultraviolet, X-rays and gamma rays are all electromagnetic radiation; they differ in wavelength [3]. Along the spectrum, frequency and energy rise as wavelength falls: NASA’s reference chart puts radio below about 3 × 10⁹ Hz and gamma rays above about 3 × 10¹⁹ Hz [4]. The same radiation acts as a wave under some conditions and as a beam of photons under others, and both descriptions are needed [2]. To carry data, a transmitter changes a property of the wave on purpose (its amplitude, phase, frequency or polarisation) and the receiver measures the result and estimates which symbols were sent. A microwave oven uses the same radiation for a different purpose: a magnetron produces microwaves, water molecules in the food vibrate, and that produces heat. It does not make food radioactive [9].

Limits and open problems

The everyday pictures mislead. A photon is not a bead with part of the message written on it, and the receiver does not need the same photons that left the transmitter. Losing some of them lowers the signal against the noise; the engineering question is whether enough signal remains to tell the symbols apart.

How it works

Glass fibres had been used since the 1930s, but lost too much light to carry information far. In the 1960s Charles Kao showed that fibres of very pure glass would carry enough of it, work that earned him half of the 2009 Nobel Prize in Physics [1]. The other ingredient arrived on 16 May 1960, when Theodore Maiman built the first working laser [12]. In a modern link a transmitter modulates a laser and the fibre guides the light, which weakens and spreads as it goes. Long routes are built as a chain of spans. Optical amplifiers raise the power of the light without decoding it, and add some noise as they do; regenerators, where used, detect the data, decide each symbol again and send a fresh signal. The ITU publishes design guidelines for submarine systems of this kind [7]. The result is the backbone of the internet: submarine cables carry over 99% of international data flows [13].

Who competes

Routes and landing points are concentrated. The ITU’s advisory body lists geographic concentration and heavy reliance on a small number of cable systems among the main weaknesses, with small island states and least developed countries the most exposed [14].

Limits and open problems

A cable is a fixed asset on the seabed. Human activity and natural hazards break cables, repairs take time, and permits slow both building and mending [14]. Fibre also cannot reach a ship, an aircraft or a remote site, which is where radio and satellites come in.

How it works

A satellite network uses two kinds of link. Starlink’s published design gives each satellite Ku-band phased-array antennas and dual-band antennas to reach customers by radio, and three lasers to reach other satellites, each rated at up to 200 Gbps. The satellites orbit at about 550 km, which Starlink says brings latency to around 25 ms against more than 600 ms for a geostationary satellite [8]. So the dish on a roof talks to the satellite by radio, and the satellites talk to each other by laser. NASA has tested laser links far beyond Earth orbit: its Deep Space Optical Communications experiment sent ultra-high-definition video from 19 million miles away at 267 Mbps in December 2023, returned data from 288 million miles in July 2024, and made its 65th and final pass in September 2025 [10]. Laser data does not travel faster than radio data. The gain is a higher data rate from a narrower beam and a smaller, lighter terminal [11].

Who competes

Starlink publishes the most detail and sells a smaller laser terminal to other satellite makers, rated at 25 Gbps [8]. NASA’s demonstrations are public and documented mission by mission [10, 11].

Limits and open problems

A laser beam is narrow, and an error of a fraction of a degree can miss the target entirely. Clouds and mist interrupt optical links to the ground, so operators need several ground stations in different places [11]. The 200 Gbps figure is a published maximum for one link, not the capacity of the network or the speed a customer receives [8].

How it works

The newest idea is to put the computer in orbit as well. Google announced Project Suncatcher on 4 November 2025 as research into a network of solar-powered satellites carrying its TPU AI chips [15]. In September 2026 it set out the reasoning and the problems. A satellite in the right low orbit is in near-constant sunlight and could collect up to eight times the solar power of a panel on the ground. In a vacuum, heat can leave only through radiators. The chips were tested in a proton beam for radiation tolerance, and laser links between satellites in a cluster are due to be tested in 2027 [16]. On 1 October 2026 a prototype built with Planet was launched on a SpaceX rideshare, contact was confirmed, and the mission will measure how the chips respond to radiation and temperature [17]. That is one satellite testing hardware. It is not a data centre.

Who competes

Google, working with Planet, runs the programme with the most public documentation [15, 17]. Other organisations have announced plans; this report cites only what a primary source documents.

Limits and open problems

Nothing here has been shown at scale. Heat rejection, radiation damage, the impossibility of repair, launch and replacement cost, and the capacity to send results down are all open [16]. Every figure comes from the company doing the work, and there are no independent cost data yet.