The Intelligence Equation · Appendix C

Artificial Intelligence glossary

96 terms, defined the way the book uses them.

A

Accelerator: A chip designed for the arithmetic of training and running models: a GPU, a TPU or another custom design.
Acqui-hire: A purchase made mainly for a company's people; see licence-and-hire.
Agent: A model in a loop that plans a step, calls a tool, reads the result and continues until the goal is met or it gives up. Agents sell completed work rather than answers.
AI Act: The European Union's comprehensive AI statute, in force from 1 August 2024, which sorts AI by risk and attaches duties to providers and deployers whose systems or outputs are used in the EU.
AI Diffusion Framework: The US export rule of January 2025 that sorted countries into three tiers with caps on advanced chips; its rescission was announced on 13 May 2025.
AI winter: A period in which funding and confidence in AI collapsed after promises outran results, as in the 1970s and after the expert-systems boom.
API: Application programming interface: the metered service through which developers send requests to a model and pay, usually per token.
ARR: Annual recurring revenue: contracted subscription revenue expressed as a yearly figure. In AI it is often used loosely for run-rate.

B

Backlog: Revenue contracted but not yet delivered; as good as the customers' ability to pay and the supplier's ability to deliver.
Batching: Running many users' requests through one pass of a model's weights at once, the most important lever on the cost of inference.
Benchmark: A standard test set on which models are scored, such as MMLU, GPQA, SWE-bench or ARC-AGI.
BIS: The Bureau of Industry and Security of the US Commerce Department, which writes and enforces export controls on chips and chipmaking tools.

C

Capital cycle: The rise and fall of investors' willingness to fund AI, one of the two forces that sit around the intelligence equation like weather.
Chinchilla rule: The finding of DeepMind's 2022 paper that a compute-efficient model is trained on about 20 tokens per parameter.
Circular deal: An arrangement in which a supplier invests in, lends to or backstops a customer whose payments then return to the supplier as revenue.
Code of Practice: The voluntary EU code, published on 10 July 2025, whose signatories are presumed to comply with the AI Act's general-purpose-model duties.
Compute: The processing capacity used to train and run models, measured in operations, GPU-hours or gigawatts of data-centre power.
Contamination: The presence of test questions, or close variants, in a model's training data, which lets it partly memorise the answers.
Cost per task: The all-in cost of completing one unit of work, including inference, retries, reviews, escalations and errors; the measure that matters more than price per token.
CoWoS: TSMC's advanced packaging, which mounts a GPU die beside its memory stacks on a silicon interposer; the bottleneck of 2023.
CUDA: Nvidia's programming model and libraries for general computation on its GPUs, released in 2006–07, against which most AI code is written.
Custom accelerator: A chip designed by or for one buyer, such as Google's TPU or Amazon's Trainium, often with a design partner such as Broadcom.

D

Data platform: Software that stores, processes and governs a company's data and increasingly serves models and agents on it; Databricks's lakehouse is an example.
Decode: The stage of inference in which output tokens are generated one at a time, limited mainly by memory bandwidth.
Deployer: Under the AI Act, an organisation that uses an AI system professionally; fine-tuning or rebranding a model can turn a deployer into a provider.
Depreciation life: The number of years over which an accelerator's cost is written off; a change of one or two years moves the reported return on a data-centre build-out.
Destination user: A person who chooses to open an AI product, as distinct from an exposure user who meets an assistant embedded in another app.
Distillation: Training a smaller model to imitate a larger one, often matching it on narrow tasks at a fraction of the cost.

E

Economic margin: Revenue less inference cost and the amortised cost of training the models that earned it; harsher than inference gross margin.
Entity List: The US list of foreign companies to which exports require a licence, used against Huawei from 2019 and Zhipu AI from 2025.
EUV: Extreme-ultraviolet lithography, the machines needed for leading-edge chips, made only by ASML and never shipped to China.
Expert system: A program that applied rules written down from human specialists through an inference engine; the basis of the 1980s AI industry.
Export controls: Government limits on selling chips, tools or technology abroad; in AI, chiefly US rules on China since October 2022.

F

Fair use: The US doctrine that excuses some unlicensed copying; in AI cases the fourth factor, harm to the market for the original work, is where cases are won and lost.
Fine-tuning: Further training of an existing model on narrower data to adapt it to a task, style or customer.
FLOPs: Floating-point operations, the unit of training compute; a dense model's training run needs about 6 × parameters × tokens.
Foreign direct product rule: The US rule that extends export controls to chips made abroad with US tools or technology.
Foundry: A manufacturer that makes chips designed by others; TSMC makes almost all leading AI accelerators.
Free tier: The unpaid version of a consumer AI product; its compute, multiplied by the free base, is a cost most consumer software never had.
Frontier model: One of the most capable models available at a given time; it typically leads its nearest rival for a few months and the best open-weight model for perhaps six to twelve.

G

General-purpose AI model: Under the AI Act, a model that can serve many tasks; providers owe documentation, a copyright policy and a training-content summary, and those above 10^25 operations are presumed to carry systemic risk.
Golden set: A buyer's own collection of representative tasks with known answers, rerun whenever a model or prompt changes.
GPU: Graphics processing unit, originally built for drawing pixels and since the 2010s the main accelerator for training and running neural networks.
GPU-hour: One accelerator used for one hour; the unit in which compute is rented and training runs are costed.
Gross margin: Revenue less cost of revenue, as a share of revenue. In AI it varies with every model call and usually excludes training.

H

HBM: High-bandwidth memory, stacked DRAM packaged beside an accelerator, supplied by SK Hynix, Samsung and Micron.
High-risk system: Under the AI Act, uses such as hiring, credit scoring, education, critical infrastructure and medical devices, which require risk management, documentation, logging, human oversight and conformity assessment.
HSR review: US merger review under the Hart-Scott-Rodino Act, which can take a year and which licence-and-hire deals were structured to avoid.
Hyperscaler: One of the largest cloud companies, such as Microsoft Azure, AWS and Google Cloud.

I

Inference: Running a trained model to produce an answer; the recurring cost of every use, and by 2025–26 the larger part of the industry's compute bill.
Intelligence equation: This book's claim that an AI company creates durable value only when four factors are all present, and that they multiply rather than add.
Interconnect: The links that join accelerators into one system, such as NVLink within a rack and InfiniBand or Ethernet between racks.

L

Lakehouse: An architecture that merges the data warehouse and the data lake, promoted by Databricks.
Licence-and-hire: A deal in which a large company takes a non-exclusive licence to a start-up's technology and hires its leaders and much of its team, leaving the company and its investors behind; used for Inflection, Adept, Character.AI, Windsurf and Groq.

M

MFU: Model FLOPs utilisation: the share of an accelerator's peak arithmetic actually spent on the model, the rest lost to communication, memory movement, restarts and idle time.
Mixture of experts: A model design in which only part of the network is active for each token, so compute is set by active rather than total parameters.
Model lab: A company that trains frontier or near-frontier models and sells access to them, such as OpenAI, Anthropic, Google DeepMind or Mistral.
Model routing: Sending each request to the cheapest model able to handle it, reserving frontier models for the hardest work.

N

Neocloud: A specialist cloud that rents large clusters of the newest accelerators on multi-year contracts, usually financed with debt secured on the chips and contracts; CoreWeave is the model.

O

Open weights: A model whose trained parameters are published for others to run and modify, under licences that range from permissive (MIT, Apache) to restricted.
Outcome pricing: Charging per completed result, such as a resolved support ticket, so that the vendor pays for failures and answers for quality.

P

Parameters: The numbers inside a model, its weights, adjusted during training; a rough measure of its size.
Post-training: The stages after pre-training, such as fine-tuning and reinforcement learning, that turn a raw model into a usable assistant.
Powered shell: A data-centre building with its power connection in place, ready for the customer's equipment; a late one can move a quarter's revenue.
Preemption: The displacement of state law by federal law; in AI, the contested federal effort to override state AI statutes.
Pre-training: The first and largest training stage, in which a model learns to predict the next token on a very large body of text and other data.
Prompt caching: Storing repeated input so that it is read back at a fraction of the normal price.
Provider: Under the AI Act, an organisation that develops an AI system or places it on the market, carrying heavier duties than a deployer.
Public benefit corporation: A for-profit company whose charter commits it to a public purpose as well as to shareholders; OpenAI Group PBC since 28 October 2025.

Q

Quantisation: Storing a model's weights in fewer bits to cut the memory and cost of serving it.

R

RAISE Act: New York's frontier-model law, signed on 19 December 2025, with a 72-hour incident deadline, effective 1 January 2027.
Reasoning model: A model that generates hidden "thinking" tokens before it answers, improving hard answers while multiplying the tokens consumed.
Retrieval: Fetching documents at query time and feeding them to a model; where the strongest copyright claims and recurring licence fees now sit.
RLHF: Reinforcement learning from human feedback: training a model on people's ratings of its answers, the technique behind ChatGPT's dialogue tuning.
RPO: Remaining performance obligation: revenue customers have contracted for but not yet received, disclosed in filings.
Run-rate: A recent period's revenue, often a single month, multiplied to a year; not audited, and an overstatement in fast-changing usage businesses.

S

Saturation: The point at which leading models score near the maximum on a benchmark, so that it no longer separates them.
SB 53: California's frontier-AI law, signed on 29 September 2025 and in force from 1 January 2026, which applies to developers of models trained with more than 10^26 operations, requires critical safety incidents to be reported within 15 days and requires the largest developers to publish safety frameworks.
Scaling laws: The empirical finding that a model's loss falls predictably as a power law in compute, data and parameters.
Seat: A price per user per month, the standard software unit, which shrinks when agents replace the users.
Services as software: Selling the work a person would do, priced against the labour it replaces, rather than a tool priced per seat.
Sovereign AI: Compute, models and data that a government wants under national control.
Speculative decoding: Using a small model to draft tokens that a larger model checks in one pass, cutting the cost of generation.
Systemic risk: Under the AI Act, the presumed risk of general-purpose models trained with more than 10^25 operations, which adds evaluation, adversarial testing and incident duties.

T

Take-or-pay: A contract under which the customer pays for committed capacity whether or not it uses it.
Tender offer: A sale of employees' or early investors' shares to new buyers, the main source of liquidity at a private lab.
Time horizon: METR's measure of the length of task, in skilled human time, that a model completes half the time.
Token: The unit in which a model reads and writes text, roughly three-quarters of a word; the unit in which inference is priced.
TPU: Tensor processing unit, Google's custom accelerator, used for its own models and sold to others including Anthropic.
Training run: The computation that sets a model's weights; the final run is a fraction of the research programme behind it.
Transformer: The neural-network design introduced in "Attention Is All You Need" (2017), which trained fast on GPUs and scaled better than earlier designs.

U

Usage pricing: Charging per token, call, action, credit or minute, which shares the variance of compute cost with the customer.
Utilisation: The share of paid accelerator time spent doing useful work; as large a lever on cost as the price of the chip.

V

Vector database: A database that stores numerical representations of documents so that a model can retrieve relevant passages.

W

Weights: The parameters of a trained model; the asset that open-weight releases give away.
Wrapper: A thin interface on someone else's model, whose margin and features the supplier can take away.