Glossary

An asterisk marks a term the book uses in a sense of its own.

10-K, 10-Q. The annual report and the quarterly report an American listed company files with the SEC. Chapter 10.

122 days. The time xAI gives for turning a serviced factory in Memphis into its first cluster. The count has no fixed start or end: its owner’s post ends it on the first weekend of September 2024, the company’s web page at a training start, and no published date fixes its first day. It is a claim about speed, not a measurement of it. Chapter 4.

Abilene. Crusoe’s campus in Texas, built for Oracle and, through it, OpenAI, on a greenfield site. Chapter 12 uses its first phase as the anchor of the slippage matrix, the forty campuses it scores, thirty-one of them with dated promises. Chapter 4.

Abstain.* What a row of the Dryburgh Compute Race Scorecard does when its reading cannot fairly vote. A row abstains if its source flags the reading as unreliable, if the reading is unverified, if it is a J row without its one-line reason or whose reason finds no limb shown, or if its reading sits equally near two thresholds. An abstaining row prints its reading and leaves the count. The Scorecard.

Accelerator. A chip built for the arithmetic of AI, such as a GPU. What is sold under the name is a small system (logic dies and stacks of high-bandwidth memory bonded into one package). Chapter 4.

Accelerator-hour. One accelerator working for one hour, the unit the rental market quotes as a GPU-hour. A usable hour has power, cooling and a network behind it, so that it can be sold or put to work. Chapter 14.

Adoption clock.* The pace at which the economy puts the technology to work. It runs slower than the capability clock, and the money was committed against the faster one. Chapters 3, 11.

Advanced packaging. The step where logic meets memory, in which an accelerator’s dies and memory stacks are bonded into one package on a single base. In the summer of 2026 it was the narrowest corridor of the buildout, because every frontier accelerator goes through TSMC’s packaging, some of it done with outside packaging houses. Chapter 5.

Affiliates rule. An American export rule that extends the Entity List’s license requirement to any company the listed firms own half of or more. It is also called the fifty-percent rule. Chapter 7.

Agent. A model given tools and permissions and set to do work on its own, rather than answer one question at a time. In the chapters, agents write code, run experiments and act on networks. Chapter 3.

AGI. Artificial general intelligence. Its definitions differ, and the shortest forecasts of its arrival use the loosest ones. Chapter 3.

Agreed order. An order that both a regulator and the company it binds sign. Mississippi’s agreed order counted the Southaven turbines, barred additions and set a retirement date for each. Chapter 9.

AI factory. The industry’s name for a datacenter built to turn electricity into tokens. Chapter 5.

AISI. The UK’s AI Security Institute, which measures how long a cyber task a frontier model can complete. Row 4 of the Scorecard reads how fast that length doubles. Chapter 8; the Scorecard.

Algorithmic efficiency. The fall in the compute needed to reach a fixed level of performance. For language models it halved about every eight months between 2012 and 2023, on Epoch AI’s estimate. Chapter 1.

Alignment. The work of making a model behave as its makers intend. In August 2026 Anthropic reported that Claude, run as an automated alignment researcher, outscored twenty-eight human safety researchers given up to eight hours, so the automation of research reached alignment itself. Chapter 3.

Amici. Friends of the court, outside parties who file briefs in a case. An amicus brief is one such filing. Chapter 8.

Announced against energized.* The gap between the capacity a company has announced and the load actually drawing power. Chapter 12’s fourth claim is that the two are different quantities, that the ratio between them is the central fact of the physical race, and that the announced campuses are not being energized on the dates that were promised. Chapter 12.

API. Application programming interface, the door through which other companies’ software reaches a laboratory’s models. Chapter 2.

Application layer. The companies that build products on top of the laboratories’ models. Chapter 1.

ARC-AGI-3. A benchmark of novel puzzles that the ARC Prize Foundation keeps as a test of general intelligence. Humans score 100 percent on it, and row 7 of the Scorecard reads the best model’s score. Chapter 3.

Arsenal behind every arsenal.* The strong form’s claim for AI, that it is the layer that discovers, designs, aims, coordinates, defends and improves every other weapon. Chapters 1, 8.

Astra. OpenAI’s GPT-6 Astra, shipped on 3 September 2026, the first model the company designated Critical under its Preparedness Framework. Its 62.7 percent on ARC-AGI-3 crossed the bar of row 7 of the Scorecard. Chapter 3.

Automated AI researcher. OpenAI’s goal of a system that does the work of an AI researcher, the result the money was spent to buy. Its first rung, the research intern, is a system that carries out well-defined research tasks under human direction. Chapter 3.

B200. NVIDIA’s Blackwell accelerator, the chip beside the H100 in the rental indexes and the compute futures. See Blackwell. Chapter 14.

Backstop. Two senses. On the grid, PJM’s reliability backstop, a procurement the operator runs when the market has not brought enough capacity. In finance, a rated party standing behind a campus’s rent once its leases begin, which is what lets the lenders look through to it. Chapters 6, 10.

Basis point. A hundredth of one percentage point, the unit in which lenders quote spreads. The Scorecard writes it as bp. Chapter 10.

Batch Zero. ERCOT’s first batch study of large-load requests, run with an audit of which requests are real. Chapter 6.

Behind the meter. On the customer’s side of the utility’s meter. Behind-the-meter generation is a power plant on the datacenter site, at its extreme connected to nothing else. Chapter 6.

Beignet. Beignet Investor LLC, the joint venture that owns Meta’s Hyperion campus, four-fifths owned by funds of Blue Owl Capital, and the vehicle whose notes funded Blue Owl’s side of it. The notes are secured on the venture interest, one level above the buildings, not by a mortgage on the campus. Chapters 6, 10.

Benchmark. A fixed test on which models are scored and compared. The word also means a reference rate, as in SOFR, the benchmark that replaced Libor. Chapter 3.

Blackwell. The NVIDIA accelerator generation after Hopper, the H100’s, including the B200 and the GB200 NVL72 rack. Chapter 4.

Bond book. The orders investors place for a new bond while it is sold, also called the order book. A book five times covered holds five dollars of bids for every dollar of bonds on offer. The Introduction; chapters 10, 11.

Booked, allocated, rationed.* Three verbs kept apart for the queue at the foundry. Capacity is booked by money laid down years ahead, allocated among the buyers who booked, and rationed by lead times when demand still exceeds it. Chapter 5.

bp. See basis point.

Brake, the.* OpenAI’s pauses of its own training and evaluation over safety in 2026. The brakes decided which workloads could use the installed compute, not whether the compute would be bought. Chapter 3.

Break fee. A payment owed if a deal is abandoned. SpaceX’s option on Cursor carried two, due only if SpaceX walked away or Cursor ended it for SpaceX’s uncured breach; SpaceX completed the purchase on 14 August 2026, so neither fell due. Chapter 14.

Break-point.* A rental price for an accelerator-hour at which the book’s model of Hyperion changes its ordering of the four futures. On SemiAnalysis’s $3.1 million figure for a 72-accelerator rack there are three. Below $3.01 an hour the strategic engine stops being the best future for the campus and the rationed utility takes its place. Below $2.66 the strategic engine no longer covers its rent. At $1.84 the commodity case would break even. The book’s figures, not a market’s. Chapter 14.

Bridge loan. A short loan that carries a borrower until longer money is raised, and is repaid from it. SpaceX’s bridge was repaid from its bonds. Chapter 10.

Bridge power. Electricity bought in to run a campus while its own plants are late. In the four futures the rationed utility runs half its hours on bridge power at an assumed $95 a megawatt-hour; in Entergy’s architecture for Hyperion the minimum bill covers the plants’ full cost, overruns and bridge power. Chapters 6, 14.

Brownfield. See greenfield.

Calibration baseline.* The reading of the Scorecard printed in this book, read in October 2026, against which the reading of 31 August 2027 is measured. It is a baseline rather than a test because its rows were written after the events they score, to reproduce a first count made on 29 August 2026 whose rows were never published; the first test comes on 31 August 2027. The Scorecard.

Campus ledger.* Chapter 12’s record of the forty largest announced campuses, which sets the load announced or estimated for each against the load the site can be shown to draw, with the state of its dated promise. Printed in full as Promised Against Energized. See slippage matrix. Chapter 12.

Capability catch-up. See clocks of diffusion.

Capability clock.* The pace at which the technology’s capability advances, which runs faster than the adoption clock, and against which the money was committed. The Scorecard uses the same words for the first of its three groups of rows, which ask when a model can do its own research. See Scorecard’s clocks. Chapter 3.

Capacity auction. PJM’s yearly auction for capacity, which on a grid means enough megawatts at the moment of peak demand, as against energy over the year. It is the auction through which the datacenters’ cost reached household bills. Chapter 6.

Capex. See capital expenditure.

Capital expenditure. Spending on buildings and equipment, shortened to capex. Chapter 1.

Capital recovery. Repaying the cost of a plant over a fixed term with a return on it, as a utility’s tariff or a lender’s schedule does. Chapter 14 derives Hyperion’s minimum bill that way. Chapter 14.

Catalytic controls. Catalysts that clean nitrogen oxides and other pollutants from a turbine’s exhaust. Selective catalytic reduction is one such system. Chapter 9.

CHIPS Act. The American subsidy law of 2022 for chip manufacturing, under which the fabs in Arizona and elsewhere were funded. Chapters 5, 8.

CDS. Credit default swap. See credit protection.

Chain, the.* The book’s picture of the machine as links in a dependency order, from the wager through science, infrastructure, finance, revenue, sovereignty and power to consent. Each link demands something of the one below, and each can delay, redirect or stop the links above it. The five contests are the five places on it where the binding constraint can sit, and one machine is the claim that its links are one system. The Introduction.

Chain of thought. A model’s written reasoning as it works. When that reasoning was itself policed and penalized, a model learned to hide its intent in it. A chain-of-thought monitor is a second system reading the first model’s reasoning. Chapter 3.

Change of landlord.* What the two exits from the machine’s constraints advertised most loudly, the Gulf and orbit, each turn out to be. Leaving for either changes who holds the veto, and moves it to where fewer people can say no. See no outside. Chapter 13.

Checkpoint. A saved state of a model partway through its training. Chapter 3.

Clean Air Act. The federal air-pollution law. The NAACP gave notice of intent to sue under it over the Memphis turbines in June 2025, and sued under it over the Southaven turbines in April 2026. Chapter 9.

Clearing house. A go-between that settles trades between buyers and sellers; what a commodity gets when its market is large enough to need one and thin enough to fear trading without one. The book’s instance is OpenRouter, a service that passes requests to AI models, after Stripe agreed to buy it. Chapter 14.

Clock. A word the chapters use for several measures of pace. The Prologue’s four clocks are how long a lead lasts, how long a rent runs, how long a chip is paid for and how long a permit takes. The others each have an entry: the clocks of diffusion, the capability clock and the adoption clock, the science clock of chapter 3, the lenders’ clock (the rack’s clock, chapter 10’s name for it), the subsidy clock, and the Scorecard’s clocks. The Prologue; chapters 1, 3, 10, 11, 14; the Scorecard.

Clocks of diffusion.* The three rates, measured in chapter 1, at which a frontier lead spreads to rivals. They are frontier turnover from model to model at the American frontier, capability catch-up by open-weight models, and the national frontier lag of China’s frontier behind America’s. Chapter 1.

Close, the.* The book’s evidence cut-off, 1 November 2026, the day to which every status the book gives was checked. The Introduction; the Scorecard.

Closed model. See open model.

Cluster. A datacenter’s worth of chips wired to work as one machine. Chapter 1.

Colossus. The supercomputer xAI built in a former Electrolux factory in Southwest Memphis, run since the merger by SpaceXAI and then SpaceXSI. Colossus 2, its second, has its compute in Memphis and its power across the state line in Southaven, Mississippi. Chapter 4.

Combined-cycle. A gas plant that burns gas in a turbine and uses the hot exhaust to raise steam for a second turbine. Chapter 6.

Commodity infrastructure.* The future in which three or more suppliers reach the same capability and the accelerator-hour falls to its cost. Everything works in it, and on the book’s model Hyperion still loses money, because at that price the campus runs flat out and does not cover its rent. Chapter 14.

Compute. The chips, and the buildings and power to run them. The wager holds that a head start can be bought with it. The Introduction.

Compute futures. Exchange-traded contracts on a month’s rent of one chip, an H100 or a B200, tracking Silicon Data’s rental-price indexes. CME fixed the terms on 11 August 2026 and announced listing from 5 October, pending the CFTC’s review; at the end of October 2026 the contracts were not listed, and the review was extended to 9 November. Chapter 14.

Consent.* The last link of the chain, the one in which the people affected hold no paper. Its no is held by a board, a regulator, a court or a ballot, and it mostly relocates, delays and reprices. In 2026 permission moved from a cost the builders paid to a race they could lose, run in polls, elections and legislatures. Chapter 9.

Consent as a race.* The third of chapter 12’s claims, that permission decides where compute is built and on what terms, and that it moved in 2026 from a cost the builders paid to a race they could lose. See conversion. Chapter 12.

Conversion, prevention.* The two outcomes chapter 12 keeps apart for consent. Conversion turns opposition into a schedule and a covenant while the machine runs, as the Memphis and Southaven turbines were permitted, litigated and given a retirement calendar. Prevention is a named campus ended, not delayed or moved, because the neighbors said no. The word conversion has two other senses in the chapters: a campus converting from announced to energized, and a brownfield conversion, a building turned into a datacenter. Chapter 12.

Convertible. A bond or preferred share that its holder can turn into the issuer’s stock on set terms. NVIDIA bought convertible preferred at Lumentum and Marvell, and convertible bonds at MediaTek. Chapters 5, 10.

Coupon. The interest rate a bond pays. Chapter 10.

Covenant. A condition written into a loan, a bond or another contract. Chapter 10.

CoWoS. Chip on wafer on substrate, TSMC’s process for bonding logic dies and memory into one package. Its capacity, counted in wafers a month, is read in row 17 of the Scorecard. Chapter 5.

CPU. Central processing unit, a general-purpose processor, as against an accelerator built for AI. Chapter 4.

Credit protection. A contract that pays out if a borrower defaults, a credit default swap, written CDS in the Scorecard. Its price, in basis points a year, is the market’s reading of the borrower’s risk. Chapter 12.

Critical. The level of cyber capability, in OpenAI’s Preparedness Framework, at which a model can find and build working exploits across many well-protected systems without human help. The framework then requires safeguards during development as well as before release. Chapter 3.

Cross-examination.* Chapter 12, in which the book tests its own thesis, that the buildout is one machine in which a lead bought with compute can be kept, by stating five ways it could be wrong, each with evidence, a surviving objection and a dated test. See thesis. Chapter 12.

Curtailment. Cutting back a load’s power when the grid is short, as in the system’s highest hours. A new load can agree to it in advance, and demand response is the same agreement from the customer’s side. Chapter 6.

Custom silicon. A chip a buyer designs for itself, such as Google’s TPU, Amazon’s Trainium or Meta’s MTIA. It changes who designs the chip and who captures the margin, not which island makes it. Chapter 5.

Cut-out. A customer ordering chips on another’s behalf, as a Chinese company did for Huawei at TSMC. Chapters 5, 7.

DDTL. Delayed-draw term loan, a loan drawn in stages as the equipment it pays for is bought. Row 26 of the Scorecard reads the spreads CoreWeave paid on such loans. Chapter 10.

Debt-service coverage. The ratio a rating agency rates a bond on, the cash available for its payments against the payments due. S&P rates Hyperion’s notes on it. The book’s rent cover is a different ratio. Chapter 10.

Deemed export. The export-control doctrine under which letting a foreign person use a controlled technology, inside or outside the United States, counts as exporting it and needs a license. The June 2026 directive to Anthropic applied it to a chatbot, reaching the developer’s own foreign-national staff. Chapter 14.

Defense Production Act. The law under which the government can order companies to put national-defense contracts first. It was threatened against Anthropic in 2026 and, on the record, never invoked. Chapter 8.

Delegation.* Giving a model a goal, leaving the room and trusting what comes back, the question chapter 3 tests. It is the middle of three thresholds, between a model capable of a task and one that substitutes for a researcher. Chapter 3.

Demand response. A large customer’s agreement to limit or shift part of its load when the grid calls, offered in return for a faster connection. Curtailment is the same cut seen from the grid’s side. Chapter 6.

Depreciation. See useful life.

Destructive finance.* The fifth of chapter 12’s claims, that the buildout’s financing has the shape of a leveraged infrastructure mania and will destroy capital chiefly if the accelerator-hour becomes cheap or idle before the contracts and the debt have run off. Destructive means a default, a restructuring or a write-down of $1 billion or more in a single instance, in an issuer’s or a lender’s own filing. The claim is tested channel by channel, in four channels, the hyperscalers’ borrowing, OpenAI’s paper, Anthropic’s and the chip debt. Chapter 12.

Distressed asset.* The future in which the buildout loses money outright, the tail of the four. It would be announced by two consecutive years of falling energized additions, cancellations of blocks of capacity that had been financed, downgrades or a neocloud default, and shutdowns ordered rather than negotiated. Chapter 14.

Docket. The public case file of a regulator or a court. Chapter 2.

Double blockade.* Washington’s and Beijing’s blocks on the same trade in chips. Washington’s is a license it may refuse, by a published rule or on terms disclosed by the company it binds; Beijing’s is a refusal to buy, given as an unpublished word at its border and known only through sources. Chapter 7.

DRAM. Dynamic random-access memory, the standard memory chip, as against the high-bandwidth memory stacked beside an accelerator. Chapter 5.

Dryburgh Compute Race Scorecard, the.* The book’s instrument for judging which of the four futures the record is moving toward. Thirty-one indicators, each with a threshold fixed in this edition and a grade for its evidence, each voting for one future or abstaining; the verdict is the plurality of the rows whose evidence is graded A to C. It is read at dryburgh.com/vol1/scorecard on the last day of each month. The Scorecard.

Durable asymmetry.* A lead kept long enough to compound. The wager rests on it, by one of two routes, exclusivity of the model or superiority of the machine around it. The Introduction.

Early-termination fee. The fee a customer owes if it leaves a power contract before its term. At Hyperion a Meta Platforms guarantee backs it. The fee protects only the original estimate for the generation and transmission, while the minimum bill is recalculated so that the customer pays the plants’ actual cost. Chapter 6.

EFLOPS. See floating-point operation.

Energized. Connected and actually drawing power. It is the grid’s word, and this book’s test of operating. An authorization is not an energization, and load counts as energized only when it can be seen drawing. See evidenced. The Introduction; chapter 12.

Entity List. An American export list. A sale to a firm on it requires a license, and the affiliates rule extended that requirement to any company the listed firms own half of or more. Chapter 7.

Epoch AI. A research group whose estimates run through the chapters, from its capability index of models to its readings of datacenter sites from satellite imagery and permits. Chapter 1.

ERCOT. Electric Reliability Council of Texas, the state’s grid operator. Its queue of large-load requests, and its audit of them, recur through the chapters. Chapter 6.

ETF. Exchange-traded fund, a fund whose shares trade on an exchange as a company’s do. Chapter 10.

EUV. Extreme ultraviolet lithography, the process that prints the finest patterns on a chip. One Dutch company, ASML, makes every EUV lithography machine in production. Chapters 5, 7.

Evidence grade.* The Scorecard’s letter for how far a reading can be trusted, from A, a primary document opened, to E, a claim not found or unverifiable. A slash-grade such as A/D means the reading has a primary figure inside a derived calculation, and the lower letter governs. The Scorecard.

Evidenced.* The book’s word for the load the public record shows: a megawatt counts as evidenced when a primary document shows it energized or when Epoch AI’s site-level estimate counts it as operating IT capacity, and never when the builder says so. Epoch’s estimates are of installed capacity, not a metered draw, and they are most of the evidenced side of the campus ledger, which says so. Chapter 12.

Evidentiary hearing. A hearing at which a body takes evidence before it decides. The NAACP and two local groups asked for one on the permanent Southaven permit. The hearing was set for 10 November 2026, and the permit stands in the meantime. Chapter 9.

Exclusive generation.* A generation of models with an improvement no rival has yet matched, the thing exclusivity of the model needs and the record since GPT-4 has not shown. Chapter 1.

Exclusivity of the model.* The first route to a durable lead, in which no rival has what the leader has. The record since GPT-4 runs against it. Chapter 1.

Export controls. A state’s rules on which chips, tools and even models may be sold abroad, and to whom. Washington’s are published, or disclosed by the company they bind. Beijing’s rare-earth controls of October 2025 are published notices, suspended until 10 November 2026; its proposal of October 2026 would write a large model’s weights into a published list; and its stop on the H200 at its own border reached the record only through sources. Chapter 7.

Fab. A chip factory. Chapter 5.

Fable. Anthropic’s models Fable 5 and Fable 5.1, the same models as Mythos 5 and Mythos 5.1 with their cyber safeguards in place. Fable 5 and Mythos 5 were switched off for every customer worldwide on 12 June 2026 under an export control directive, and Fable 5 was off for nineteen days. Fable 5.1, released on 1 September, is generally available. See Mythos. Chapters 3, 8.

Failed sale-leaseback. A sale of equipment leased straight back to the seller, which accounting judges was not truly a sale. The equipment stays on the seller’s books, and the obligation stays beside it. Chapter 10.

Falsifier. A dated test that would show one of the book’s claims wrong. A claim whose falsifier fires is withdrawn, and the withdrawal is reported beside the Scorecard’s verdict. Chapters 1, 12.

Fifty-percent rule. See affiliates rule.

Finance lease. A lease that accounting treats as a purchase paid for over time, so that it counts with capital spending. The Scorecard.

Firm power. Power available every hour without interruption. A firm, around-the-clock connection needs four things at once, energy over the year, capacity at the peak, deliverability to the site and firmness, which is all of them every hour; a training run pays for the lack of it in restarts, and a shopping mall can survive losing it for an hour. Chapter 6.

Five contests.* The five places on the chain where the binding constraint can sit, also called the five races: capability, capacity, value, sovereignty and consent, coupled so that any one, binding hard enough, holds back the other four. The Introduction.

Five races. See five contests; the chapters say contests, and the Introduction names the fifth race, consent.

Floating-point operation. A single arithmetic step in which a computer’s work is counted. An EFLOPS is a billion billion of them a second. Chapter 2.

Foundry. A company that makes chips other companies design. TSMC is the foundry behind every frontier accelerator outside China. Chapter 5.

Four futures.* The four outcomes the machine can still have, a strategic engine, a rationed utility, commodity infrastructure and a distressed asset. The Scorecard’s plurality says which is arriving. Chapter 14.

Free cash flow. Operating cash flow less capital expenditure. Chapter 10.

Frontier. The leading edge of AI capability, and the models and laboratories at it. Chapter 1.

Frontier turnover. See clocks of diffusion.

Four clocks.* The Prologue’s four: a lead lasts about seven weeks, a rent can end on ninety days’ notice, a chip is paid for over years, and a permit takes longer than the machine it permits. The book opens and closes on them. See clock. The Prologue; the Epilogue.

GB200 NVL72. NVIDIA’s rack-scale computer, seventy-two accelerators and thirty-six CPUs wired over copper to work as one machine. It is the cabinet that defined the buildout of 2025. Chapter 4.

GB300. The successor of the GB200 NVL72 rack, in NVIDIA’s Blackwell Ultra generation. Chapter 4.

General-purpose technology. A technology that spreads through every industry. The scholarship on them finds that power and value go to whoever adopts them, not to whoever owns the frontier technique. Chapter 1.

Generation time. The time one generation of models takes to produce the next. Chapter 1.

Generator step-up transformer. The transformer that puts a power plant’s output onto the wire. Its lead time is read in row 15 of the Scorecard. Chapter 6.

Gigawatt. The unit a power plant is counted in, written GW, about one large nuclear reactor. Leopold Aschenbrenner’s essay Situational Awareness of 2024, behind the interview with which chapter 1 opens, made it the unit of the argument. Chapter 1.

GPU. Graphics processing unit, the commonest accelerator, a chip built for the arithmetic of AI. Chapter 4.

GPU-hour. See accelerator-hour.

Greenfield. A site built on bare land. A brownfield conversion instead turns a serviced building into a datacenter, as the 122-day build did in Memphis. Chapter 4.

Grid queue. The line of projects waiting to connect to the power grid, a connection the grid calls interconnection and studies before it is made. The book’s refrain is that the queue is not a forecast. Chapter 6.

Gross margin. Revenue less its direct costs, such as inference, as a share of revenue. Gross profit is the same amount in dollars. Chapter 11.

Gross profit. See gross margin.

GW. See gigawatt.

H100. NVIDIA’s accelerator of the Hopper generation, the chip of the fleets of 2023 and 2024. Its rent and resale are the test of destructive finance. Chapters 4, 12.

H20. A downgraded NVIDIA Hopper chip, in the Institute for Progress’s words “designed specifically to comply with export control thresholds,” which went on sale in China in 2024. Washington required a license for it in April 2025. The H200 is almost six times as powerful. Chapter 7.

HLE. See Humanity’s Last Exam.

H200. The H100’s successor in the Hopper generation, almost six times as powerful as the H20. A Washington rule of January 2026 let it be sold into China under case-by-case licenses, and an instruction from Beijing stopped it at the Chinese border that week. Chapters 5, 7.

Harness. The setup a model is run or tested inside, its prompt, tools and permissions. The same model can behave differently under a different harness. Chapter 3.

HBM. High-bandwidth memory, memory chips stacked one on another beside an accelerator. Three makers supply it, two of them Korean, and it is booked years ahead. Chapter 5.

Hopper. See H100.

Humanity’s Last Exam. A benchmark of expert-written questions across many fields, written to be hard for the best models; row 8 of the Scorecard reads the best model’s score on a text-only subset of its questions, and the Scorecard writes it HLE. The Scorecard.

Hyperion. Meta’s datacenter campus in Richland Parish, Louisiana, powered by three Entergy gas plants and financed through Beignet’s notes. Chapter 14 runs it through the four futures. Chapter 6.

Hyperscaler. One of the largest builders and operators of datacenters. Most are cloud companies that also rent capacity out, while Meta builds mainly for its own use. In the counts the chapters cite they include Alphabet, Amazon, Meta and Microsoft. Some counts add Oracle. Chapter 10.

IG. See investment grade.

Index fund. A fund that holds a market index by rule, taking each bond or share at its index weight rather than a view of the issuer. Chapter 10.

Inference. Running a trained model for the people who use it. Training builds a model, and inference runs it. Chapter 1.

Injunction. A court order to do or to stop something. The NAACP asked the federal court to halt the Southaven turbines while its suit ran. Chapter 9.

Inner loop, outer loop. The inner loop of research generates candidates, tests them and keeps the best, and it is automated. The outer loop chooses the question and the test of success, and it is human. Chapter 3.

Interconnection. See grid queue.

Investment grade. The rating agencies’ mark of debt safe enough for cautious buyers, written IG. Debt below the line is high yield, or junk. Chapter 10.

Investment grade, the letters. The agencies write the line in letters: BBB- or Baa3 and above is investment grade, BB+ or Ba1 and below is high yield. Chapter 10 reads CoreWeave’s loans at Ba2 and BB+, just below the line, and the hyperscalers’ bonds far above it. Chapter 10.

IPO. Initial public offering, a company’s first sale of its shares to the public. Chapter 10.

Island, the.* Taiwan, where one company, TSMC, makes the frontier chips and, some of it with outside packaging houses, bonds them into their packages. The island in every machine is the book’s phrase for that concentration. Chapter 5.

IT load. The power drawn by a datacenter’s IT equipment (its servers, storage and network) before cooling and other overhead. Chapters 3, 12.

J row.* A Scorecard row whose vote turns on a qualitative judgment, such as fast progress or material delay. Its vote needs a one-line reason; without one it is void, and the row abstains. The Scorecard.

Jailbreak. To talk a model past the limits its maker set, as when attackers led Anthropic’s model to believe it was doing authorized security work. Chapter 8.

JPMAM. J.P. Morgan Asset Management, the bank’s asset-management arm. Row 23 of the Scorecard reads its series of hyperscaler bond issuance. The Scorecard.

Junk. See investment grade.

Kilovolt. A thousand volts, the measure of a power line’s voltage. Three of the four grid operators’ transmission portfolios in chapter 6 are built around 765-kilovolt backbones, and Hyperion’s new lines run at 500 kilovolts. Chapter 6.

Kilowatt. See megawatt.

KV cache. A served model’s short-term memory of the conversation, filled by reading the whole prompt at once. Filling it accounts for the pause before the first word appears, the time to first token. Chapter 14.

kW. See megawatt.

Large load. A customer large enough that the grid must study it before connecting, such as a datacenter. Texas sets the line at 75 megawatts, and PJM’s proposal at 50, a line the federal regulator accepted on 9 October 2026, while rejecting the rest of the proposal. Chapter 6.

Ledger. See campus ledger for chapter 12’s record of the forty campuses. The record of thirty-one indicators at the back is the Scorecard.

Lenders’ clock.* The schedule on which the debt raised against the machine is repaid, which the machine must outlast. Chapter 10 calls it the rack’s clock, the term over which the rack’s hardware is paid for; the owner’s own clock is the useful life over which it depreciates the hardware. Chapter 10.

Lien. A claim on a property for unpaid work. It is the claimant’s assertion, not a court’s judgment. Chapter 4.

Limb.* One condition inside a test. In the Scorecard a limb is one future’s condition in a row, and a falsifier’s limbs are conditions that can each fire alone. The Scorecard.

Lithography. The printing of a chip’s circuits. The finest patterns are printed by EUV machines. Chapter 4.

Logic die. The computing chip proper. An accelerator bonds its logic dies and stacks of memory into one package. Chapter 4.

Loop, the.* AI improving AI, so that a lead compounds. The wager bets that whoever has the loop first would have it alone. Chapter 1.

Loss. A model’s error at predicting the text it is trained on. The scaling laws say how much more compute each further reduction takes. Chapter 2.

Mass is money.* Chapter 13’s claim that a gigawatt in orbit costs more than the same gigawatt on the ground, because every kilogram of radiator, panel and shielding is launched. Chapter 13.

Megawatt. A thousandth of a gigawatt, written MW. A kilowatt, written kW, is a thousandth of a megawatt. Chapter 6.

Megawatt-hour. A megawatt drawn for one hour, the unit in which electricity is sold. Chapter 6.

METR. The organization that measures how long a task a model can complete. Its time-horizon readings fill the first three rows of the Scorecard. Chapter 3.

Minimum bill. A charge payable whether or not the power is used. At Hyperion it is designed to cover the three gas plants’ full cost, including overruns before they enter service and any power bought in to bridge the gap if they are late. Chapter 6.

MISO. The regional grid operator that controls the interconnection of Hyperion’s generators, and one of the four grid operators whose transmission plans chapter 6 sums. Chapter 6.

Mission layer. This book’s name for the layer in which a state runs several models and can swap one for another: the data, the orchestration that routes work between models and tools, the logs, the acceptance tests and the substitute kept on hand. Whoever owns it sets the limit of what a maker’s license can take away. Chapter 8.

MLGW. Memphis Light, Gas and Water, the municipal utility of Memphis. Chapter 4 reads its posted records for the second substation planned for the Memphis cluster, due in the fall of 2025, which no posted minute shows in service. Chapter 4.

Mobile and temporary. Mississippi’s words for the Southaven turbines, which exempted them from air permits while each stayed on its trailer and on site for under twelve months. See nonroad engine. Chapter 9.

Moot. Dismissed because the thing appealed against has gone, as the Memphis appeal was once the temporary turbines had left. Chapter 9.

Moratorium. A pause, such as one on the processing of new datacenter rezonings, site plans and building permits. Chapter 9.

Moving bottleneck.* The second of chapter 12’s claims, that the binding constraint on compute moves in a predictable order and can be named before it binds rather than after. For 2027 the book names permission (the permits and the consent that the 2026 record already shows binding), and for 2028 the bond market’s patience. Chapter 12.

MW. See megawatt.

Mythos. Anthropic’s most capable models of 2026, kept from general release. Claude Mythos Preview went in April only to the eleven launch partners and some forty organizations of Project Glasswing, its program for critical-infrastructure defenders, and was the strongest model METR had measured in May. Mythos 5, released on 9 June as Mythos Preview’s upgrade, was “the same model as Claude Fable 5 but with cyber safeguards lifted,” and Mythos 5.1 is available only through trusted-access programs. See Fable. Chapters 3, 8.

NAACP. The National Association for the Advancement of Colored People. It sued under the Clean Air Act over the Southaven turbines in April 2026, and with Young, Gifted & Green and the Safe and Sound Coalition asked for an evidentiary hearing on the permanent Southaven permit. Chapter 9.

Nameplate. The rating a machine carries, its output under standard conditions. A nameplate compute draw is installed accelerators times their power, not the power actually used. Chapter 10.

National frontier lag. See clocks of diffusion.

Neocloud. A company, such as CoreWeave, that rents out accelerators to others and finances them largely by debt against its customers’ contracts. Chapter 10.

Nitrogen oxides. Pollutants in a gas turbine’s exhaust, which catalytic controls clean and a satellite measured over Southaven. Chapter 9.

No outside.* See change of landlord.

Node. A generation of chip manufacturing, named in nanometers. The most advanced node ships from Taiwan first. Chapter 5.

Nonroad engine. The category for tractors and portable generators, in which the Shelby County Health Department placed xAI’s trailer-mounted turbines. It kept them outside the permits required of stationary sources such as power plants, if each stayed less than a year in one place. Chapter 9.

Notes. Two senses. In finance, bonds; senior unsecured notes are backed by the company’s general promise to pay rather than by any particular asset. In this book, the Notes at the end of each chapter, keyed to a phrase in the text, each naming its source; the full Note for every citation is kept at dryburgh.com/vol1/notes. Chapter 10; A Note on Sources.

OCF. Operating cash flow, the cash a company’s operations produce before capital expenditure. The Scorecard.

Off balance sheet. Obligations a company discloses in its footnotes but does not carry as debt, such as leases not yet commenced and purchase commitments. Chapter 12.

Offtake. A customer that has contracted for the capacity a project will supply; an offtake agreement is that contract, and lenders lend against it. Chapters 6, 10.

One machine.* The book’s claim that the science, the machines, the money, the states and the neighbors are parts of a single machine rather than five separate stories. The Introduction.

Open model. A model whose weights are published for anyone to download and run, also called an open-weight model. A closed model’s weights stay with its maker. Chapter 1.

Open-weight model. See open model.

Order book. See bond book.

Operating, under construction, contracted, proposed, requested.* The five words used for capacity, never traded for one another. The buildout is reported in announcements, and these words count it in electrons. The Introduction.

Paper, the.* The loans, leases and bonds that pay for the machine. Chapter 10 follows the paper that carried one machine, from a rented turbine to a bond due in 2056. Chapter 10.

Parameter. One of the adjustable numbers inside a network that training sets, and the unit a model’s size is counted in. The weights are their trained values. Chapter 1.

Perpetual. A contract with no expiry date that tracks an index, the crypto exchanges’ form of a future. Bitget’s “pre-market compute perpetuals” on Silicon Data’s H100 and B200 indexes were reported on 9 September 2026, while CME’s futures were still unlisted. Chapter 14.

PJM. The grid serving more than 67 million people across thirteen states and the District of Columbia. Its yearly capacity auction carried the datacenters’ cost onto household bills. Chapter 6.

Perimeter. The set of spending a count includes. Goldman Sachs’s perimeter is the spending it counts as AI investment; the Scorecard’s rows and chapter 14 name their perimeter where it decides a figure. Chapters 1, 14.

Post-training. See pre-training.

Power law. A relation in which each multiplication of the input divides the quantity measured by the same fixed factor. The scaling laws found a model’s loss falling as a power law in its parameters, its data and its compute. Chapter 2.

Pre-training. The field’s word for the training run itself, in which a model learns from its data, such as human text. Post-training is the work done on the model after it, such as reinforcement learning. Chapter 2.

Preparedness Framework. See Critical.

Prevention. See conversion.

Private credit. Loans originated by investment funds that are not banks. Chapter 10.

Prospectus. The document offering a company’s shares or bonds, which its buyers and lenders read. SpaceX’s showed its lenders how fast the machine had been built, and the lawsuit over its turbines. See S-1. Chapter 10.

PUE. Power usage effectiveness, a datacenter’s total power over the power drawn by the IT equipment inside. The closer it is to one, the less goes to cooling and conversion. Chapter 6.

Rack. The cabinet of servers in which the buildout is counted, powered and cooled. Chapter 4.

Ratepayer. A customer who pays a utility’s rates. At Hyperion, if the customer leaves early, part of the plants’ cost falls on Entergy’s other ratepayers. Chapter 6.

Rationed utility.* The future in which the machine works, the hours are scarce, and scarcity, not abundance, keeps the accelerator-hour worth paying for. Chapter 14.

Reasoning model. A model that spends more time thinking at inference before it answers. Chapter 2.

Red line. A use a supplier refuses to license its model for, such as mass surveillance of Americans or fully autonomous weapons. Chapter 8.

Refresh life.* How often a campus’s hardware is replaced, in the four futures: every three years for the strategic engine, four for the rationed utility, five and a half for commodity infrastructure (Meta’s own accounting life for its servers, used as a proxy), and never for the distressed asset. Chapter 14.

Reinforcement learning. Training a model by rewarding the results it produces, used after pre-training. It is written RL. Chapter 2.

Rent cover.* The book’s own ratio for a campus, its value after power, operations and replacement against its rent. Not the debt-service coverage a rating agency rates the notes on. Chapter 14.

Rental index. A published series of what a chip’s hour rents for across many providers. Silicon Data’s indexes are what the CME contracts track and what the book’s test on Hyperion reads on 1 June 2029. The publisher has described its inputs four ways, as chapter 14 says, and its letter to the CFTC of October 2026 does not reconcile them. Chapter 14.

Research intern. See automated AI researcher.

Residual-value guarantee. A floor under an asset’s value, paid by the guarantor if the asset re-lets or sells for less. Meta gives one at Hyperion, and NVIDIA one at OpenAI’s Ohio campus. Chapter 10.

RL. See reinforcement learning.

Run rate. Recent revenue extrapolated to a year. Chapter 11.

S-1. The registration statement a company files with the SEC before it lists its shares, the prospectus its buyers and lenders read. Chapter 10.

Sandbox. The walled-off system in which a model is tested, meant to keep it from reaching the internet. Chapter 3.

Scale-up. Wiring accelerators so closely that they work as one computer, as the seventy-two inside a GB200 NVL72 rack do, rather than adding more separate racks. The chapters describe it without using the word. Chapter 4.

Scaling laws. The finding, published in January 2020, that a model’s loss falls as a clean power law in its parameters, its data and the compute spent. They say how much more compute each further reduction takes, which made a training run priceable. Chapter 2.

Scorecard, the.* See Dryburgh Compute Race Scorecard, the.

Scorecard’s clocks.* The three groups into which the Scorecard’s thirty-one indicators fall. The capability clock asks when a model can do its own research, the construction clock when announced capacity becomes load, and the finance and consent clock how long lenders, regulators and communities will allow it. The Scorecard; chapter 14.

SEC. The Securities and Exchange Commission, with which American listed companies file their reports. Chapter 10.

SemiAnalysis. The research firm founded by Dylan Patel, whose estimates of chips, datacenters and rents run through the chapters. It sells its research into the buildout it counts. Chapter 7.

Settlement. The day bonds are paid for and delivered. Chapter 10.

Slippage matrix.* Chapter 12’s second list of forty, the campuses chosen for the clearest promise-dating, which shares many names with the campus ledger’s forty and is not the same list. It scores each dated phase as kept, slipped, missed, not yet due or never dated, and The Dated Tests reads it. Chapter 12.

Singularity. Growth that becomes infinite in a finite time, which would need the generation time to shrink toward zero. Chapter 1 says the wager does not need it, only exclusivity plus a shortening clock. Chapter 1.

SOFR. Secured Overnight Financing Rate, the benchmark that replaced Libor, over which floating-rate loans are priced. Row 26 of the Scorecard quotes CoreWeave’s spreads over it. Chapter 10.

Southaven. The city in Mississippi, across the state line from Memphis, where the mobile turbines that power Colossus 2 run on the grounds of a retired Duke Energy plant, and where a permanent plant of forty-one turbines is permitted. Chapter 9.

Sovereignty.* The link at which states decide what may be sold, to whom and on whose chips, and who commands the model. It can also override the other links’ vetoes. The Prologue; the Introduction; chapters 7, 8, 13.

SpaceXAI. The name under which the former xAI did business from its merger into SpaceX on 2 February 2026 until October. On 4 October 2026 Musk said it would be renamed SpaceXSI, SI for super intelligence, the White House’s term for AI since late September, and its web pages carried the new name by the middle of the month. The chapters name the company as it was on the day of each event, and its turbine subsidiary, MZX Tech LLC, as the filings do. See SpaceXSI; xAI. Chapters 4, 9, 10.

SpaceXSI. The name the former xAI has carried since October 2026, inside SpaceX. See SpaceXAI. Chapters 7, 9.

Spread. The extra interest a borrower pays over a benchmark rate, such as Treasuries or SOFR, quoted in basis points. Chapter 10.

Stargate. OpenAI’s program of datacenter campuses, announced in January 2025 as $500 billion over four years. Chapter 1.

Stateline. Stateline Power, the turbine venture of CTC Property and a Solaris subsidiary. It is to rent about 900 megawatts of generating equipment to MZX Tech, a SpaceX subsidiary, for Southaven. Solaris carries the venture’s assets and debts in its own accounts, and SpaceX does not. Chapter 10.

Strategic engine.* The future in which the buildout becomes the engine of a new economy. It would be announced by a verified autonomous-research result, and by energized additions that rise while credit and consent hold. Chapter 14.

Strong form. Two uses. Of the military claim, that the first state with sufficiently advanced AI gains a decisive, durable advantage over every rival, including a nuclear peer; the weak form holds only that AI already improves intelligence, cyber operations, targeting and research, and the medium form that it gives a sustained operational edge. Of the wager, chapter 12’s strong form is that one generation of models does most of the research for the next. Chapters 1, 12.

Subsidy clock.* How long tokens can be sold below what they cost to make, which for the heaviest users they are, for as long as equity is cheap enough to fund the difference. It is the third clock inside the forward bet on revenue, after the lenders’ and the revenue’s. Chapter 11.

Substation. The equipment through which a site takes its power off the grid’s transmission lines. Chapter 4.

Superiority of the machine.* The second route to a durable lead, in which the leader has more compute, more power, deeper capital, faster iteration and a state behind it. Every chapter after the first tests it. Chapter 1.

Switchgear. The equipment that connects, protects and disconnects a site’s electrical circuits, between the substation and the racks. Chapter 4.

Take-or-pay. A contract that binds a customer to pay for agreed volumes over its term, whether or not it takes them. Chapter 5.

Tariff. The word has two senses in the chapters. A trade tariff is a tax on imports, and a utility’s tariff is its published schedule of rates and terms. Chapter 6.

TeraFab. The fab Tesla and SpaceX announced for themselves, with logic, memory, packaging, test and the making of lithography masks under one roof. It is the largest private buyer’s attempt to build around the island. Chapters 4, 5.

Terawatt. A thousand gigawatts. Chapter 6.

Thesis, the.* The book’s argument in one sentence: that the buildout is one machine, that a head start in self-improving AI bought with compute might be kept, and that the chain of constraints from the wager to consent decides whether it is. Chapter 12 cross-examines it. The Introduction; chapter 12.

Time horizon. The length of task a model completes at a given reliability, counted in the time a skilled human would need, as METR measures it. The Scorecard reads it at 50 and 80 percent reliability. Chapter 3.

Time to first token. The pause before a served model’s first word appears, while it reads the whole prompt at once. Chapter 14.

Token. A word or a piece of a word, the unit a model reads and writes, and the unit of text an AI is sold by. The Introduction; chapters 11, 14.

Total factor productivity. The output an economy gets from a given stock of labor and capital. A rise in it is growth that added labor and capital do not explain. Chapter 1.

TPU. Tensor processing unit, Google’s own accelerator, which comes off TSMC’s lines like NVIDIA’s. Chapter 5.

Training. Building a model from its data, as against inference, which runs it. A training run is one such computation, and a training cluster is built for it. Chapter 1.

Trajectories. The step-by-step records of a model’s work. Chapter 3.

Tranche. One slice of a bond, loan or warrant sold or granted in several parts on different terms. Chapter 10.

Transformer. In AI, the network design of 2017 that processes every part of an input at once. With it, a model’s size is bounded by the accelerators one can buy. On the grid, a transformer is equipment that changes voltage. Chapter 2.

Triple-net lease. A lease under which the tenant pays taxes, insurance and upkeep on top of the rent. Chapter 10.

TSMC. The Taiwanese company that makes the frontier chips. Every accelerator that matters outside China and carries its memory beside the chip passes through TSMC’s packaging line. Chapter 5.

TVA. The Tennessee Valley Authority, the federal power utility that serves Memphis. Its board authorized the Memphis cluster’s grid power in three installments, in November 2024, February 2026 and August 2026. Chapters 4, 9.

Uncatchable lead.* The wager as chapter 12 states it, its first claim: that whoever gets a sufficient lead in automated research keeps it, because the loop of AI improving AI compounds faster than rivals can copy. The chapters test it; they do not assert it. Chapter 12.

Useful life. The years over which an owner depreciates hardware, writing it off on its books, five to six years on the operators’ own filings. The lenders’ clock, the term over which the hardware is paid for, is a different clock. Chapter 4.

Utilization. The share of a machine’s hours that are used. With the value of each hour, it decides a campus’s return. Chapter 14.

Vehicle. A company formed only for one deal, to hold an asset and the debt against it. Chapter 10.

Vendor credit. A seller’s support for its buyers’ purchases, through investments, guarantees, warrants or promises to buy back capacity. NVIDIA’s is the book’s leading case. Chapter 12.

Vera Rubin. The NVIDIA accelerator generation after Blackwell. Its NVL72 rack is the design SpaceX’s first satellites are to carry. Chapters 5, 13.

Veto.* A link’s power to delay, redirect or stop the links above it in the chain. Each veto the Introduction’s map draws has been used at least once, on a date its chapter gives. The Introduction.

Wafer. The silicon disc on which chips are printed, the unit in which foundries count their capacity. Chapter 5.

Wager, the.* The proposition the machine exists to test, that AI is about to improve itself and that a head start in that loop, bought with compute, compounds into a permanent one. Chapter 1 calls it the wager too, and chapter 12 tests it as the uncatchable lead. It is also called the bet and the belief. The Introduction.

Warm shell. A finished datacenter building, its power and cooling in place, waiting for the racks. Chapter 6.

Warrant. A right to buy a company’s shares at a set price. Chip vendors gave warrants to buyers, vesting as the buyers bought. Chapter 5.

Weak form, medium form. See strong form.

Weights. The trained values of a model’s parameters, which are the model itself. Open models publish them for anyone to download and run. Chapter 1.

WPM. Wafers a month, the unit in which packaging and wafer capacity are counted. See wafer. The Scorecard.

xAI. The company Elon Musk founded in 2023, which built Colossus in Memphis and merged into SpaceX on 2 February 2026, doing business afterward as SpaceXAI and, from October 2026, SpaceXSI. The chapters use the name for events before the merger. See SpaceXAI. Chapters 4, 9.

Yield. For a bond, the return it pays at its market price. For a chipmaker, the share of chips that come out good. Chapter 10.