Glossary

What the factory makes

Token: A chunk of text, roughly three quarters of a word. Your question is chopped into tokens on the way in and the answer is manufactured token by token on the way out. It is also the unit the AI labs charge for.
FLOP: One floating-point operation: a single multiply or add. Producing one token takes hundreds of billions of them. A chip’s speed in FLOPs per second is how many workers it has on the floor.
Training: Building the model: adjusting more than a trillion internal settings over months. Happens once per model.
Inference: Running the model: every question anyone asks, billions of times a day. Most AI computation is now inference. The money is in running the factory, not building it.
Batching: Processing hundreds of people’s questions on the same chip at the same moment, the way a kitchen cooks one big pan and plates fifty servings. It is the difference between a query costing pennies and costing pounds.
KV cache: The model’s working memory of your conversation, built while it reads your prompt so it does not have to re-read everything for every new word.

Getting power to the building

Interconnection queue: The waiting list to physically connect a new site to the electricity grid. You file a request and wait while the utility studies whether the grid can take you. Currently years, not months, and the real bottleneck on AI buildout.
Transformer (the grid kind): The large grey box that steps voltage down so power can be used. Nothing to do with the AI model architecture that shares the name. Supply-constrained, and lead times run to years.
Scheme of Control: Hong Kong’s agreement between the government and each power company. It obliges the company to meet demand at the lowest reasonable cost and allows it a regulated return. CLP’s runs to 31 December 2033.
Reserve margin (power): Installed capacity above the year’s highest demand, as a share of that demand. CLP Power’s was about 24% in 2025 (9,258 MW against 7,455 MW), so about 19% of capacity was idle at the peak.
Substation: The site where high-voltage transmission power is converted and routed onward. Building one is a multi-year civil project, not an equipment order.
Announced versus energised megawatts: Announced is capacity a developer says it will build. Energised is capacity actually connected and drawing power. The two have been diverging, and company disclosure rarely separates them cleanly.
Behind the meter: Generating power on your own site, on your side of the utility’s meter, so you never join the public connection queue. Gas turbines, fuel cells and restarted nuclear are the usual routes.
Grain-oriented electrical steel: The specialised steel used in transformer cores. Very few producers worldwide, which is part of why transformers cannot simply be made faster.
Inverter: The equipment that converts between the direct current a battery stores and the alternating current the grid uses. Every storage installation needs one, and it is a separate supply chain from the cells.
Slot reservation agreement: A commitment to a future manufacturing slot rather than a firm order for a machine. Much of the reported gas turbine order book is this rather than equipment being built, which is why a large backlog is not the same as capacity arriving soon.
Ride-through: The ability of a site to stay connected to the grid through a brief disturbance instead of disconnecting to protect itself. Long required of generators. Now being considered as a condition of connecting a large data centre.
Grid frequency: The measure of whether generation and demand are matched, 50 hertz in most of the world and 60 in North America. It falls when supply is short and rises when demand disappears, which is the direction large data centre trips push it.
NERC: The North American Electric Reliability Corporation, which sets and enforces reliability standards for the North American grid. Its alerts run from Level 1 to Level 3, where Level 3 is the most urgent.
PJM: The grid operator for Virginia and much of the mid-Atlantic United States, and the region with the densest concentration of data centre load anywhere.
Switchgear: The heavy equipment that makes and breaks high-voltage connections safely, so a fault can be isolated without taking the site down. It sits alongside the transformer in every substation and has its own lead time.
MVA: Megavolt amperes. The rating that describes what grid equipment can carry, as opposed to megawatts, which describes what is actually being consumed. Utilisation figures for a connection point are usually quoted against MVA.
Solid oxide fuel cell: A box that turns natural gas or hydrogen into electricity through a chemical reaction across a hot ceramic plate, with no flame and no moving parts. Factory-built and modular, so it can be installed on site in months rather than years.
NOx: Nitrogen oxides, the pollutants from burning fuel that form smog. How much a power source emits largely decides how hard its air permit is to obtain, and so how quickly it can start.
Gross margin: The share of each sale left after the direct cost of making and delivering the product. Used on this site only as evidence: a supplier widening its margin while volume grows is being paid for something scarce.
Backlog: Orders a company has signed but not yet delivered, as the company itself reports it. For equipment with maintenance contracts, much of it can be service work stretching years ahead rather than machines waiting to ship.
Book-to-bill: New orders taken in a period divided by sales shipped in the same period. Above 1 means work is arriving faster than it can be delivered, so the queue of orders is growing.
Organic growth: Growth with the effects of acquisitions and currency movements stripped out, so it shows what the existing business actually sold more of.
Adjusted operating margin: The share of each sale left after the costs of running the business, with one-off items removed. Used on this site only as evidence: a margin widening while volume grows means each sale is carrying more value.
ARR: Annualised run-rate revenue. The current month of revenue multiplied out over a year. It measures how big a business is now, and it is not a forecast.
Adjusted EBITDA: Operating profit before interest, tax and the wearing out of equipment, with one-off items removed. Used on this site only as evidence that capacity earns once it is running, never as a measure of value.
Annual contract value: What a customer pays each year of a multi-year deal. For AI capacity it is often quoted per megawatt, which ties the contract directly to the power switched on.
Lead time: The wait between ordering a piece of equipment and having it delivered. For generators, large transformers and some cooling gear it now runs past a year, which is why they are ordered before the chips.
Colocation: Renting space, power and cooling inside a data centre someone else runs. Trading firms colocate beside the exchanges, where the host sets the power and cooling limits.
SMR: Small modular reactor. A nuclear reactor of up to about 300 megawatts, built from standard factory-made parts rather than as a one-off giant plant, so later units should be cheaper and quicker than the first.
PPA: Power purchase agreement. A contract in which a buyer agrees to take electricity at a set price, often for 10 to 40 years. Long contracts from reliable buyers are what lenders want before funding a power plant.
Independent power producer: A company that owns power plants and sells their output, without owning the wires that carry it or sending bills to households.
Fast reactor: A nuclear reactor that runs on neutrons left unslowed, usually cooled by liquid metal such as sodium rather than water. Sodium stays liquid at reactor temperatures without being held under high pressure.
HALEU: High-assay low-enriched uranium. Nuclear fuel enriched to between 5 and 20 per cent, against under 5 per cent for conventional reactors. Many advanced reactors need it, and commercial supply outside Russia barely exists yet.
Criticality: The moment a reactor sustains its own chain reaction for the first time. It proves the reactor works as a machine, not that it produces electricity at a cost anyone will pay.
Contracted versus connected power: Contracted power is electricity a company has secured the rights to at its sites. Connected power is what is actually wired up and feeding running chips. A neocloud is paid for the second, not the first.
BESS: Battery energy storage system. Increasingly bought not to trade electricity prices but to let a site switch on ahead of its grid connection.
Heavy-duty gas turbine: The large turbine at the centre of a power station, each producing hundreds of megawatts. Built to order, and the slowest part of building your own power to get hold of.
Aeroderivative: A smaller gas turbine adapted from a jet engine. Less power each, but quicker to ship and switch on, so it is often used as a bridge until a heavy-duty machine arrives.
Castings and forgings: The large shaped metal parts at the hot heart of a turbine, made by a small number of specialist suppliers. Their arrival sets how fast turbine output can step up.
HVDC: High voltage direct current. A way of carrying very large amounts of power over long distances or under the sea with less loss than alternating current. Each link needs a converter station at either end and takes years to build.
Bushing: The insulated sleeve that carries a high voltage conductor through the earthed steel tank of a transformer. A specialist component with its own waiting list.
GVA: Gigavolt amperes, a thousand MVA. The unit for how much power large transformers can carry, used to measure transformer demand against factory capacity.
Battery integrator: A company that designs a complete grid battery system, buys the cells and other parts, has contract factories assemble them and runs the control software. It does not make the cells.
Prohibited foreign entity: The American tax law’s term for, in short, companies tied to China, Russia, Iran or North Korea. From 2026, projects claiming clean energy tax credits must limit material from them, and the limit tightens each year.
Capacity reservation: An agreement in which a buyer books a manufacturer’s future factory or engineering capacity before placing firm orders for specific equipment. The grid equipment cousin of a slot reservation agreement.
Framework agreement: A multi-year supply deal that sets terms and an upper value in advance, with individual orders called off under it as projects firm up.
Supply capacity agreement: A multi-year deal in which a data centre operator commits to buy a supplier’s equipment in phases, so both sides can plan output. Used on this site only as evidence of how buyers secure equipment.
Contract liabilities: Cash a company has collected from customers for equipment it has not yet delivered. Used on this site only as evidence: deposits rising ahead of deliveries mean buyers are paying to hold a place.
Deferred revenue: Payments and billings a company has received ahead of delivering the product. Unlike pure customer deposits, it also includes amounts billed in advance. Used on this site only as evidence of how buyers hold their place in a queue.
Revenue mix: How a company’s sales split between kinds of offer, such as catalogue products against engineered systems. A shifting mix shows what customers are actually buying.
Profit margin before special items: Profit before one-off costs such as restructuring, as a share of revenue. Used on this site only as evidence of pricing power.
Basis point: A hundredth of a percentage point. A margin falling 200 basis points has fallen two percentage points.
Business profit: The main measure of operating profit some Japanese companies, Ajinomoto among them, report for each business. Used on this site only as evidence of how scarce a product is.
Capital markets day: A briefing where a company sets out its plans and targets to investors, usually every few years.
Restart (nuclear): Bringing a closed reactor back into service after repairs, inspections and regulatory approval. Only three closed American reactors, Crane, Palisades and Duane Arnold, have restart plans.
Uprate: Raising the output of a reactor that is already running, usually with new turbines, pumps and controls, within limits set by its design and licence.
Capacity interconnection rights: In PJM, a power plant’s right to deliver a set amount of power onto the grid at its connection point. A plant that closes gives them up; a plant being retired can, with approval, hand them to another.
FERC: The Federal Energy Regulatory Commission, the American regulator of interstate electricity transmission and wholesale power markets.
Nuclear Regulatory Commission: The federal safety regulator that licenses American reactors, including any restart.
Capacity auction: An auction in which a grid operator pays power stations, by the day, to promise they will be available when the grid is under most strain, a few years ahead. The cost is shared across every customer’s bill. PJM’s is called the Base Residual Auction.
Reliability requirement: The amount of capacity a grid operator says it needs to keep the chance of running short to about one day in ten years.
Large-load tariff: Special terms a utility sets for very large new customers such as data centres, for example paying for most of the power they reserve whether they use it or not, so other customers do not carry the cost.
Firm service: A connection the grid operator plans to keep supplied at all times, without a right to cut it short of an emergency.
Non-firm service: A connection the grid may cut first when lines are strained, offered so a site can switch on before the upgrades for firm service are built.
Network upgrades: The new lines, substations and equipment a grid operator says it must build before it can give a new site firm service.
Show cause order: A FERC order telling a grid operator to explain why its rules are still just and reasonable, or to propose changes.
Large load (Texas): A single site drawing 75 megawatts or more, the threshold Texas law sets for the special connection rules data centres face.
Large computational load: In PJM’s October 2026 proposal, a data centre or similar site of 50 megawatts or more with at least 25 megawatts behind power electronics, which must ride through faults; switching to backup power does not by itself count.
Early-connection tier: A connection that lets a site switch on before its grid upgrades are built, on condition the grid may cut it first, such as ERCOT’s Provisional Controllable Load Resource or PJM’s Interim NITS.
Net peak: The evening hour when demand minus wind and solar is highest, which the rest of the grid must cover. In California it arrives around 8pm.
Four-hour battery: A battery that can run at full power for four hours before it needs recharging, the most common size of grid battery in America.
Demand response: An agreement to use less electricity when the grid asks, usually at peak, for payment or bill credits. Google’s version slows or delays its AI computing.
Minimum demand charge: A floor on a large customer’s bill, set on the demand it contracted, paid whether or not it uses that much power.
Stranded cost: The unpaid cost of a power station or line built for a customer that leaves or never uses it, which falls on other customers unless a contract covers it.
Gigawatts under contract: GE Vernova’s count of gas turbine capacity held by firm orders and slot reservations together. Set against its yearly output, it measures the queue in years.
Netting: Counting a recent emissions cut at the same site against a new plant’s increase, so the project avoids the full review for that pollutant.
Emission reduction credit: A certified, permanent cut in pollution elsewhere that a new plant buys to offset its own emissions.
Combined cycle gas plant: A power station that burns gas in a turbine, then uses the waste heat to raise steam for a second turbine.
Electric service agreement: The contract between a utility and a large customer to supply and pay for power, often with minimum monthly charges and a fee for leaving early.
Investment casting: Making a metal part by pouring molten metal into a ceramic shell formed around a wax copy. How turbine blades are made.
Superalloy: A nickel-based metal that keeps its strength close to its melting point; the material of the hottest turbine blades.
Associated gas: Natural gas that flows out of an oil well with the crude, so its supply follows oil drilling. Most Permian gas.
Rig count: The number of drilling rigs at work, counted weekly by Baker Hughes.
Bcf a day: Billion cubic feet of gas a day.
LNG (liquefied natural gas): Gas chilled until liquid so a ship can carry it.
Liquefaction train: One production line in an LNG export plant.
Active power: Power drawn by installed, running computers and available to earn revenue. Nebius’s term; the stage customers pay for.
Delivered capacity: Data centre capacity handed over to the tenant who will run computers in it. It comes after a building is energised and before the computers run at full load.
Capacity in service: A builder’s term for capacity ready for customers and fully fitted out. DayOne counts it separately from capacity merely energised.
Prepayment: Money a customer pays before capacity is delivered, often covering part of the cost of building it.
Enrichment: Raising the share of uranium-235, the atom that splits, by spinning uranium gas through fast centrifuges. Ordinary reactors use up to about 5 per cent; HALEU runs from 5 to 20.
Cascade (enrichment): A line of centrifuges linked so each passes slightly richer gas to the next. Centrus’s demonstration cascade in Ohio has sixteen.
SWU: Separative work unit, the standard measure of enrichment work. American reactor owners bought about 12.7 million in 2025.
First core: The full initial load of fuel a new reactor needs before it can start, far larger than a yearly reload.
MISO: The grid operator and wholesale market across much of the American Midwest and South.
Suspension (of a power plant): In MISO, a plant that stops running but keeps its connection, for at most 36 cumulative months in any five years.
Refuelling outage: The planned shutdown, every 18 to 24 months at Constellation’s units, when a reactor takes new fuel and heavy equipment can be swapped.
Extended power uprate: The largest kind of uprate, approved for up to 20 per cent more power, needing new heavy equipment and an NRC licence amendment.
Construction permit: The NRC’s permission to build a reactor of the design described; a separate operating licence is needed to run it.
Test reactor (non-power): A reactor licensed for research under section 104c of the Atomic Energy Act; the law caps the share of its running costs that can go to selling energy.
Thermal and electrical megawatts: A reactor’s heat rating (MWt) is larger than its electrical output (MWe), because a steam plant turns well under half its heat into power.
Boiling water reactor: A water-cooled reactor in which the water around the fuel boils and its steam drives the turbine directly.
Years of work: A company’s order backlog divided by its revenue over the previous four quarters. It shows how long its factories would take to clear the orders already signed if no new order came in. A proxy for the wait, not a quoted delivery time: price rises and early bookings also lengthen it.
Maximum demand: The most power a customer may draw at once, booked in its supply agreement with the utility. A data centre books it for its finished size and fills it over years.
Energy supply agreement: In Malaysia, the contract between Tenaga Nasional and a large new customer that fixes the maximum demand the grid will provide.
Green Lane Pathway: Tenaga Nasional’s fast-track connection route for data centres, launched in August 2023, which cut supply delivery from 36 to 48 months to about 12.
Connected load charge: In Malaysia, a penalty on a large new customer drawing less than 85% of the demand it declared, phase by phase, in its first four years, charged per kilowatt of shortfall.
Ultra high voltage tariff: Malaysia’s tariff class, from 1 July 2025, for data centres connected above 50 kilovolts. Its energy charge is higher than ordinary high voltage users pay; its capacity and network charges are the same.
Subsea power cable: An armoured high voltage cable laid on the seabed to carry power between countries. Made to order in a few specialist factories.
Converter station: The plant at each end of a direct current link that turns alternating current into direct current and back. Built to order.
Electricity Importer Licence: In Singapore, the full licence from the Energy Market Authority that lets an import project build and operate. It comes after a conditional approval or licence.
Data Centre Call for Application: Singapore’s competitive process, run by EDB and IMDA, for handing out new data centre capacity under efficiency and green power conditions.
Critical IT load: The most power a data centre is designed to deliver to its servers. Singapore licenses halls of 3 MW and up and states it on each licence.

Inside the building

Rack: The cabinet, roughly two metres tall, that holds the computers. Rack density means how much power one draws. A normal server rack is 5 to 10 kilowatts; a current AI rack is near 200; the next generation is designed for up to a megawatt.
PUE: Power usage effectiveness. Total electricity entering the building divided by the electricity that actually reaches the computers. A perfect score is 1.0; old air-cooled sites run near 2.0 and modern liquid-cooled ones near 1.1. Everything above 1.0 is overhead, mostly cooling.
Direct-to-chip cooling: A metal plate with liquid channels sitting directly on each chip, hosed to a pump unit. The mainstream answer now that air has run out of road.
UPS: Uninterruptible power supply. A large battery system between the grid and the servers that carries the whole load through a dip or cut in utility power until the generators start.
Busway: Enclosed metal power rails running above the racks, into which each cabinet taps its supply. The overhead wiring of a data hall.
Cold plate: A metal block with liquid channels clamped directly onto a chip. Coolant flowing through it carries the heat away at the source, which is how most of an AI rack is now cooled.
Oversubscription: Connecting more equipment to a power supply than it could carry if everything peaked at once, on the bet that it never will. It squeezes more out of a fixed connection, at the risk of tripping a breaker if the bet is wrong.
Buffer tank: A large tank of chilled water that keeps servers cool for the minutes the chillers need to restart after a power interruption. Effectively a battery for cold.
Neocloud: A cloud company built only to run AI. It rents out large GPU clusters, from by-the-hour use to dedicated multi-year capacity, rather than the websites, databases and email the big clouds also sell.
GPU cluster: Thousands of GPUs wired together so tightly that they behave as one computer. If one chip is slow the rest wait, so the network, storage, power and cooling around them matter as much as the chips.
Hyperscaler: The industry term for the giant cloud companies, such as Microsoft, Amazon, Google and Meta, that build data centres at the largest scale.
CDU: Coolant distribution unit. The pump and heat-exchange box that drives liquid around the rack. Effectively the rack’s heart.
Immersion cooling: Submerging whole servers in a non-conductive fluid. The most extreme option, and the one that handles the most heat.
800 VDC: Eight hundred volts of direct current, delivered to the rack instead of converting inside it. Higher voltage means less current, which means far less copper for the same power. Borrowed from the electric vehicle charging supply chain.
Busbar: The thick copper bar carrying current inside a rack. At one megawatt the old low-voltage design needs about 200 kilograms of it per rack, which is what forced the change.
Power stage: A small chip that switches on and off very fast to step a voltage down. Several sit around each AI processor, each carrying a share of the very high current it draws at below one volt.
eFuse: An electronic fuse. It cuts power in microseconds if something shorts, and, paired with a hot-swap controller, lets an engineer pull a live server tray without crashing the rack.
Wide-bandgap semiconductor: Silicon carbide and gallium nitride switching chips. They handle high voltage efficiently, which is what makes the 800 volt architecture practical.
Prefabricated module: A power or cooling room built, wired and tested in a factory, then shipped to the data centre site in one piece and connected. It trades scarce site labour for factory time.
Commissioning: Testing equipment under power on site before it is allowed to carry the real load. Every panel, UPS and cooling loop in a data centre has to pass it.
Inside the fence: Equipment on the data centre site itself, between the grid connection and the racks, as opposed to the substation and grid on the utility’s side.
Digital twin: A working computer simulation of a building or system, used to design and test it before it exists and to run it afterwards.
Cooling tower: An open unit, often on the roof, that cools water by letting some of it evaporate. Cheap in electricity, but the evaporated water is lost for good.
Dry cooler: A large radiator with fans that sheds heat to the outside air without evaporating any water. Saves water, costs more electricity on hot days.
Chiller: A large refrigerator for a building’s cooling water. It runs harder, and draws more power, when no water is evaporated.
Water withdrawal and consumption: Withdrawal is water taken in and mostly returned, as at many power stations. Consumption is water lost for good, mostly as vapour. For data centres, consumption is the figure that matters.
WUE: Water usage effectiveness. Litres of water a data centre consumes on site for each kilowatt-hour its computers use. Lower is better; zero means no water evaporated for cooling.
Airside economiser: Cooling a data hall with outside air whenever the weather allows, so the chillers can rest. Often paired with a little evaporated water on hot days (adiabatic cooling).
Waterside economiser: Cooling a building’s water loop with a cooling tower or cold outside air instead of a chiller whenever conditions allow. Efficient, but a tower evaporates water.
Adiabatic cooling: Cooling air by evaporating a small amount of water into it or over pads in front of a coil. Uses far less water than a cooling tower, and only on hot days.
Heat rejection: The last step of cooling, where a building’s heat is released to the outside air, by evaporating water or through a dry radiator.
Water risk (watershed): How likely the local water supply is to run short or be depleted. Google rates the sources it draws from as low, medium or high risk and reports the split.
Reserve margin (water): Spare water treatment capacity above average intake. Malaysia’s regulator SPAN targets 10 to 15%.
Megalitre: A million litres. Malaysian water capacity is quoted in megalitres a day.
Solid-state transformer: A transformer built from power semiconductors rather than only copper and steel, converting medium-voltage grid power straight to direct current. Silicon carbide does the switching.
Rectifier: Equipment that turns alternating current into direct current. In an 800 VDC building one large rectifier replaces many small power supplies.
Power rack: A cabinet of power converters standing beside the computer racks, feeding them 800 VDC from an existing AC power room. Nvidia’s first 800 VDC product.
Power shelf: The unit inside today’s rack that turns 415 VAC into 54 VDC for the trays.
Hot-swap controller: A chip that lets a live server tray be pulled or inserted without shorting the rack’s power.
Content per megawatt: What one supplier’s equipment is worth in each megawatt of data centre computing power, as the supplier estimates it.

Chips, memory and the wiring between them

GPU: The accelerator doing the bulk of AI maths: thousands of small identical cores performing the same operation at once, which is exactly the shape of neural network arithmetic.
TPU: Tensor processing unit, the AI chip Google designs for its own data centres. It needs a different cooling and power design from Nvidia’s racks, so a building planned for one may not suit the other.
HBM: High-bandwidth memory. Memory chips stacked eight to twelve high and glued directly beside the processor, so data travels millimetres instead of centimetres.
Memory-bandwidth-bound: When the chip is waiting for data rather than waiting to finish a calculation. Most of the time an AI model is writing its answer, this is what is limiting it.
Scale-up versus scale-out: Scale-up is the very fast link between chips inside one rack. Scale-out is the network between racks across the hall. They use different technologies and have different economics.
Optical fibre: A strand of glass thinner than a human hair that carries data as pulses of light. The core in the middle carries the light and the cladding around it keeps it in, so a signal can travel kilometres where copper fades within a couple of metres.
Total internal reflection: The effect that keeps light inside a fibre. The cladding bends light slightly less than the core, so light striking the boundary at a shallow angle bounces back inward instead of escaping.
Passive connectivity: The fibre, cable and connectors of a network: equipment with no electronics that only carries the signal. Unlike transceivers and switches, it usually stays in place through several chip generations.
Optical transceiver: The thumb-sized device at each end of a fibre link that converts between electricity and light. Copper fails after a few metres, so between racks everything becomes light. They are consumables, replaced at each upgrade.
Optical DSP: The digital signal processor inside an optical transceiver. It cleans up and reshapes the signal so electricity can become light and back again without errors, and it is the hardest part of the plug to make.
PAM4: The signalling standard used by fast links: four levels of brightness or voltage instead of on and off, so each pulse carries two bits. Harder to read, which is why the DSP is needed.
800G and 1.6T: How much data one link moves per second: 800 gigabits, or 1.6 terabits for the generation now ramping. Each step doubles the traffic through the same small plug.
XPU: An industry label for a custom AI chip, designed with a chip partner for one company’s own workload rather than sold to everyone. Cheaper per calculation than a general GPU, for that one job.
EML: Electro-absorption modulated laser. A laser the size of a grain of sand that both makes a beam of light and switches it on and off billions of times a second to encode data. The light source inside most fast data centre transceivers.
Indium phosphide: InP. The crystal that data centre lasers are grown on, because silicon is very poor at giving off light. Its wafers are small and brittle, yields are hard won, and only a few factories in the world work it at volume.
CW laser: Continuous-wave laser. A simpler laser that stays on rather than blinking, feeding light to a silicon photonics chip that does the encoding. Competes with the EML on shorter links.
Silicon photonics: Building most of an optical transceiver in ordinary silicon, fed by an external CW laser. Cheaper to scale on short links, though it still needs an indium phosphide laser somewhere.
Co-packaged optics: Putting the light conversion on the switch chip’s own package instead of in a plug at the front of the switch. Saves power and moves the translation job to whoever makes the switch chip.
Linear pluggable optics: A transceiver with the DSP removed, leaving the switch chip to clean up the signal. Uses less power, at the cost of shorter reach and tighter engineering.
Switch chip: The processor inside a network switch, sending each piece of data arriving from many chips to the right destination. Sold to any switch builder, it is called merchant silicon. Its capacity decides how many chips one network layer can connect.
Ethernet: The open standard for how machines talk over a network, used in offices for decades and now adapted for AI. No single company owns it, so any maker’s chip can join an Ethernet network.
Network tier: One layer of switches in a data centre network. Chips plug into the first tier, which plugs into the next. Every extra tier adds switches, optical plugs, power and delay, so bigger switch chips that cut a tier save all four.
Tape-out: The moment a finished chip design is sent to the factory for manufacture. Everything before it is design work; everything after it is making and testing the real silicon.
Analog chip: A chip that handles real-world quantities such as voltage, current and temperature rather than doing calculations. Many manage power, and some sell for cents.
Wafer: The thin silicon disc chips are made on before it is cut up. A 300 millimetre wafer has about 2.25 times the area of a 200 millimetre one, so each pass through the factory yields more than twice the chips.
Foundry: A company that manufactures chips other companies design. TSMC is the largest, and makes most leading AI accelerators whoever designed them.
Fab: Short for fabrication plant, the factory where chips are made on wafers. TSMC says a new one takes two to three years to build and one to two more to reach full output.
CoWoS: TSMC’s advanced packaging that joins an AI processor to its stacked memory on one base. A separate factory step from making the chip, and in 2026 the tightest one.
Advanced packaging: Assembling several pieces of silicon into one component, including stacking memory beside a processor. Frequently the real constraint rather than the chip fabrication itself.
Die: A single chip cut from a wafer. An Nvidia B200 joins two large logic dies so they behave as one processor.
Package: The finished chip component you could hold, with dies, memory and wiring mounted together on a base.
Qualification: The months or years of testing a customer runs before trusting a new supplier’s part. It is why a cheaper alternative cannot simply be swapped in.
Mask: A plate carrying the circuit pattern for one layer of a chip. The most advanced masks are mirrors rather than stencils.
Lithography: Printing a chip’s circuit pattern onto a wafer with light, layer by layer.
EUV: Extreme ultraviolet lithography, using light of 13.5 nanometres for a chip’s finest layers. Only ASML builds the machines.
Photoresist: The light-sensitive coating on a wafer that lithography prints the pattern into. Made by a few Japanese chemical companies and slow to change.
Interposer: A slab of extremely fine wiring that the processor and its memory sit on side by side, so they can talk at short range.
Package substrate: The laminated circuit board under the interposer that carries power and signals out to the rest of the computer.
ABF: Ajinomoto Build-up Film, the insulating film between the wiring layers of a package substrate. Ajinomoto says it is the product of choice for nearly all high-performance CPUs.
Thermosetting: Describes a material that hardens permanently when heated and cannot be melted again, like baked enamel. The insulating film in a package substrate is one.
EUV scanner: The machine that prints a chip’s finest layers with extreme ultraviolet light, scanning the pattern across the wafer one section at a time.
Numerical aperture (NA): How wide a cone of light a lithography machine’s optics gather. A wider cone prints finer detail. Standard EUV machines are 0.33 NA, called low NA.
High NA: The newest EUV machines, with a numerical aperture of 0.55 instead of 0.33, so they print finer detail.
Installed base: All the machines a supplier has already sold that are working in customers’ factories. Servicing and upgrading them is a large business of its own.
DRAM: Dynamic random access memory, the working memory in every phone, laptop and server. HBM is DRAM too, stacked, so the two compete for the same factories.
DDR5: The standard memory used in servers and PCs today, mounted on boards a few centimetres from the processor.
Trade ratio: How many wafers of ordinary DRAM a memory maker gives up to make the same number of HBM bits. Micron puts it at about three, and says it rises with each new generation.
Through-silicon via: A vertical connection drilled through a memory chip, so a stack of chips can be wired top to bottom. It is what lets HBM be piled up beside a processor.
Cleanroom: The dust-free hall where chips are made. A chipmaker’s output is limited by how much cleanroom it has, and building a new one takes years.
Take-or-pay agreement: A contract in which the customer pays for an agreed volume whether or not it takes delivery. Some memory makers now sign multi-year ones.
Nearline drive: A high-capacity hard drive for data that is kept rather than constantly used. The cheapest place per terabyte to store the data AI produces.
Areal density: How much data fits on each patch of a hard drive’s disk. Raising it is how drive makers add storage without adding drives.
HAMR: Heat-assisted magnetic recording. A tiny laser warms a spot on the disk for an instant so it can hold a smaller magnetic bit. Seagate sells it as Mozaic.
Scale-across: Linking AI data centres in different places, sometimes hundreds of kilometres apart, so they can work on one training job. Only the least talkative work can cross, because light takes about five microseconds per kilometre of fibre.
Optical circuit switch: A box of tiny tilting mirrors that bounces light from one fibre into another without reading it. It uses almost no power and works at any data rate, but takes milliseconds to change, so it sets the shape of a network rather than routing each packet. Google’s is called Palomar.
Optical engine: The small assembly that turns a chip’s electrical signal into light and back. In co-packaged optics it sits on the switch package.
External laser source: A removable front-panel module holding the lasers that feed co-packaged optics, so the part most likely to wear out can still be replaced by hand. The OIF standard form is ELSFP.
Tensor parallelism: Cutting each calculation across several chips, which must swap results many times a step. It needs the shortest links, so it stays inside one building.
Data parallelism: Giving each group of chips its own copy of the model and different data, then averaging what they learned once a step. It is the work that can cross between sites.
AI WAN: Microsoft’s wide area network built only to carry AI training traffic between distant data centres, such as its Fairwater sites.
Reference architecture: A vendor’s published blueprint for building a cluster from its products, naming which switch, cards and cabling to use.
Export licence (advanced chips): US permission needed to ship advanced AI chips to any company headquartered in China, or whose parent is, wherever its subsidiary sits. In Hong Kong the rule also follows the hall’s address: since 15 January 2026, H200-class chips shipped from the US to end users there are reviewed case by case.
Presumption of denial: A US licence review policy under which an export application is expected to be refused unless shown otherwise.
Case-by-case review: A US licence review policy under which each export application is judged on its merits. Since 15 January 2026 it applies to H200-class chips exported from the US to end users located in China, Hong Kong included.
Petaflops (PFLOPS): A thousand million million calculations a second. Hong Kong’s government counts computing power this way without stating the precision, so its figures compare only with each other.
Base die: The logic chip at the foot of an HBM stack that talks to the processor. Micron designs and makes its HBM4 base die in-house; from HBM4E, a buyer such as Nvidia can supply the design itself.
Memory wall: The gap between how fast a chip can calculate and how fast its memory can feed it. When an AI model writes an answer, the chip spends most of its time waiting on memory.
KV cache: The working notes a model keeps for each conversation while it answers, held in memory beside the chip. Longer conversations and more users need more of it.
Memory controller: The circuit that tells memory what to read and write. It used to sit on the processor; in custom HBM it can move into the base die of the memory stack.
NVHBM: Nvidia's own design for the base die of an HBM stack, which memory makers build to its specification. Nvidia offers it from more than one maker, to its own chips and its partners'.
Low NA EUV: ASML’s standard EUV machines, the ones its yearly capacity figures (around 65 in 2026) count. High NA is the newer, finer model.
Exabyte: A billion gigabytes. Hard drive makers report how much storage they ship in exabytes.
Supply allocation: A maker’s commitment of future output to named customers. Seagate reports how far ahead its drive output is allocated.

Missing something? Tell me and I will add it: tommyllk2003@gmail.com.