The words of the buildout.

A backlog is the stock of orders a company has accepted but not yet delivered. It is a promise of revenue, not revenue: it converts at the pace of capacity, and the part due within a year is worth more to the quarter than the part due later. Filings split the two, which is why DeepStack reads the near book and the far book separately.

In Part 01, Part 03, Part 04, Part 05, Part 06, Part 07, Part 11, Part 12, Part 16.

A basis point is one hundredth of one percent. Yields and spreads move in basis points because the differences that matter are small: a 10-year Treasury moving from 5.00% to 5.31% has risen 31 basis points.

In Part 07, Part 08, Part 10, Part 15, Part 17.

DeepStack prints the Brier score on the Scorecard from the tenth result whose test carried a number recorded before the date. For each such result the gap is the Call’s number as a fraction minus one when the answer was yes, or minus zero when it was no, squared; the score is the mean of the gaps. A Call that said 85% and was right adds 0.02; one that said 15% and was wrong adds 0.72. Lower is better, 0.25 is what a coin flip earns, and the score punishes confidence the source did not reward. Since 20 October 2026 a reader who signs in and gives a number is scored by the same rule on the same tests.

Calibration sorts a forecaster’s resolved numbers into ten bands of ten points and asks, for each band, how often the answer was yes. A calibrated forecaster’s 70s come true about seven times in ten and their 20s about two in ten, so the bars climb along the diagonal. A reader’s page and the readers’ table draw it; it rewards honest numbers, not bold ones.

Each pre-specified test records the answer the Call expected (yes, no, or either). When the test resolves, the result is read against that expectation: Call held when the answer is the one the Call expected, Call missed when it is the one the Call did not expect. The test’s own answer wears its own chip beside the reading, so a rule that says “yes if” and the seal never share a word. The hit rate counts Calls held among pre-specified tests, from the tenth result.

When a specification records that either answer fits the Call, the result is read as Call informed: it changes what the Call knows without counting as held or missed. The Call’s number on such a test still prints and still enters the Brier score.

The reading against the Call when the answer is the one the Call did not expect. It sits beside the test’s own answer (yes or no) and counts in the hit rate from the tenth result.

In a capacity auction a grid operator such as PJM buys commitments from power plants to be available in a delivery year three years out. The clearing price, set by the last unit needed, is paid on every customer's load. A price cap limits it; whether the December 9, 2026 auction clears below the cap is a test on the Docket.

In Part 13.

Capital spending, or capex, is the cash a company puts into assets that last years: land, buildings, servers, chips, grid connections. It leaves the cash-flow statement at once and reaches the income statement slowly, through depreciation. The AI buildout is measured in capex because that is where the money moves first.

In Part 01, Part 03, Part 04, Part 05, Part 06, Part 07, Part 08, Part 10, Part 11, Part 12, Part 14, Part 15, Part 17.

CCC is the rating tier just above default. The spread of CCC bonds over Treasuries, published daily on FRED, is the price of the riskiest credit; at 1,146 basis points on September 28 (Part 08) it measured how much lenders charged the weakest borrowers while financing the strongest.

In Part 03, Part 08.

A credit default swap pays the buyer if a borrower defaults. Its price, in basis points a year, is the market's running estimate of that risk: Oracle's five-year CDS at a record 227 basis points (Part 08) said lenders had started to charge for the buildout.

Days sales outstanding measures how long customers take to pay: receivables divided by daily sales. When a supplier stretches terms from 90 days to a year, as Nvidia did for its best-rated customers (Part 11), the number rises and the quarter's revenue arrives in cash much later.

In Part 11, Part 16.

Depreciation spreads the cost of a long-lived asset over its assumed useful life. The assumption is a choice: a server depreciated over six years looks more profitable each year than the same server over three. When chips go obsolete faster than the schedule says, reported profits flatter the economics; that is the argument of Parts 03, 05 and 10.

In Part 01, Part 03, Part 05, Part 06, Part 10, Part 14.

Duration measures how far out a stream of payments sits and therefore how much its present value moves when rates change. A fifteen-year data-center lease, a far backlog and a growth stock are all long-duration promises; a higher 10-year yield taxes each of them.

In Part 03, Part 04, Part 07, Part 11, Part 15.

The earnings yield is expected earnings divided by price, the forward multiple turned upside down. At 19 times forward earnings the yield is about 5.3%. It is compared with the 10-year Treasury as a valuation sensitivity, not as a like-for-like return: earnings are not contractual, and the two assets differ in duration, growth and risk.

In Part 10.

Form EIA-861M is the U.S. Energy Information Administration's monthly survey of electricity sales, revenue and customers by state and customer class. Revenue divided by sales gives the average price per kilowatt-hour; DeepStack computes the household price from it, as first published, for the tests of Part 13.

The forward multiple divides today's price by the earnings analysts expect over the next twelve months. It is a valuation, not a forecast: a lower multiple can mean a cheaper market or earnings expectations that have run ahead of what will be delivered, which is the denominator question of Part 10.

In Part 10, Part 11.

Free cash flow is the cash from operations minus capital spending. A company with strong profits and negative free cash flow is funding its buildout from outside: debt, leases, partners or investors. Alphabet's first negative quarter since its IPO (Part 04) is the example DeepStack keeps coming back to.

In Part 01, Part 03, Part 04, Part 05, Part 08, Part 09, Part 11, Part 14, Part 16, Part 17.

A hyperscaler is one of the handful of companies, Microsoft, Alphabet, Amazon, Meta and Oracle among them, that build and run data centers at a scale measured in gigawatts and tens of billions of dollars a quarter. Their capital budgets, backlogs and cash flows are the largest single inputs to the AI economy.

In Part 01, Part 03, Part 04, Part 05, Part 06, Part 07, Part 08, Part 10, Part 11, Part 12, Part 14, Part 15, Part 16, Part 17.

Interconnection is the connection of a load or a generator to the transmission grid, with the studies, upgrades and approvals it requires. A contracted megawatt is not a connected megawatt until the interconnection is energized, which is why Parts 06 and 13 separate the two.

In Part 06.

A large load is a single customer whose electricity demand is large enough that grid operators and regulators write rules for it: deposits, batch studies, requirements to bring its own capacity. Data centers are the large loads of this cycle.

In Part 13.

A megawatt is a million watts of power; a gigawatt is a thousand megawatts. Data centers are sized by the power they draw because electricity is the binding constraint: a gigawatt campus needs its own substation, its own interconnection and, increasingly, its own generation.

In Part 01, Part 02, Part 03, Part 04, Part 06, Part 10, Part 13, Part 16.

A new-issue concession is the extra yield a borrower pays on a new bond relative to its outstanding bonds, so that buyers absorb the supply. Widening concessions mean lenders are charging more scrutiny; DeepStack reads them as evidence, and settles tests on public spread series instead.

In Part 01, Part 03, Part 04.

An open-weight model publishes its trained parameters. Anyone can download and run it, which puts a ceiling on what a closed model can charge for comparable capability. The gap in months between the best open model and the frontier is one of the numbers DeepStack tracks.

In Part 05, Part 12.

The option-adjusted spread is the spread of a bond or index over the Treasury curve after removing the value of embedded options such as early redemption. ICE BofA publishes it daily for its indices, and FRED carries the series, which is why the Docket uses it to settle questions about investment-grade credit.

PJM Interconnection runs the wholesale electricity market and the transmission grid for about 65 million people across thirteen states and the District of Columbia. Its capacity auctions set the price every household in the region pays for reliability, which is the mechanism of Part 13.

In Part 06, Part 13.

Private credit is lending by investment funds outside banks and public bond markets, often secured on specific assets such as chips or data centers. It financed the compute platforms of Part 01 and the facilities of Part 03; its terms are rarely public, which is why DeepStack reads them from filings and dated reports.

In Part 01, Part 03, Part 16.

A ratepayer is any customer billed under a regulated tariff. The word matters because a data center that pays for its own wires still raises the capacity price every other ratepayer pays, which is the argument of Part 13.

In Part 06, Part 13.

A re-let is the second lease of the same capacity after the first contract ends. The re-let rate, as a share of the original price, says whether the hardware kept its value: near 95% supports durable economics, and a fast decay exposes the depreciation argument.

In Part 03.

A real yield is the yield on Treasury Inflation-Protected Securities, whose principal follows the consumer price index. It is the cost of borrowing after inflation, the number a lender financing a data center locks in. The nominal 10-year yield minus the real yield is the breakeven inflation the bond market prices.

In Part 07, Part 10, Part 15.

Remaining performance obligations are the revenue a company is contracted to deliver but has not yet recognized. Filings split them into the part due within twelve months and the part due later; the split is the near book and the far book of Part 11.

In Part 11, Part 14.

A residual-value guarantee is a commitment to cover the gap if a leased or partnered asset is worth less at the end of the term than the contract assumed. Meta's $28 billion guarantee on its data-center venture (Part 11) is a liability that lives outside the balance sheet until the asset's value is tested.

In Part 08, Part 11, Part 17.

A split Call is one of the three tones a DeepStack Call can carry, with constructive and skeptical. The tone says how the Call leans on the trade; the side says what the Call holds on the question. A split Call still takes a side on the question (Part 13 holds that the ratepayer pays first) while staking no direction on the market. The Scorecard counts the tones of every Call.

A credit spread is the yield on a bond minus the yield on a Treasury of the same maturity. It is the market's price for the risk that the borrower does not pay. Spreads are quoted in basis points; a wider spread means dearer credit. DeepStack reads spreads on FRED series and in filings, never from a dealer's note.

In Part 03, Part 06, Part 07, Part 08, Part 10, Part 13, Part 14, Part 15, Part 17.

The term premium is the part of a long-dated yield that compensates for holding duration rather than rolling short bills. When the Fed raises short rates and long yields rise more, the market is charging for time: supply, deficits and a thinner buyer base, as Part 07 argues.

In Part 07, Part 15.

Every pre-specified test on the Docket can carry the Call’s number: the chance the house gives the yes answer to the test’s question, written into the record with the words of the test and before its date, so it cannot be set in hindsight. It prints on the Docket, in the Part’s list of tests, on the home strip and on the Result, on one line from 0 (no) to 100 (yes), with the readers’ median beside it once 25 readers have given a number. It is DeepStack’s reading, not a market price and not advice; from the tenth result it feeds the Brier score.

Every signed-in reader can give a pending test a number from 1 to 99, the chance of the yes answer. The readers’ median is the middle of each reader’s latest number on the test: half gave more, half gave less. It prints on the Docket, under “What to watch” and on the Result from 25 readers, always with the count, and it is scored as one forecaster on the Scorecard beside the Call.

A token is the unit a language model reads and writes, about three quarters of an English word. Model prices are quoted per million tokens, input and output priced separately. When the price per token falls while usage rises, the question is whether the lab is paid for the token or for something else: Part 12.

In Part 03, Part 04, Part 05, Part 08, Part 09, Part 10, Part 12.

The Treasury International Capital system reports, with a lag of about six weeks, how much of U.S. securities foreign holders own, country by country. Its table of major foreign holders settles the question of whether Japan bought or sold Treasuries in a given month (test p07-t02).

Your number is the control under a pending test on the Docket and under “What to watch”: a slider from 1 to 99 and a Save button, open to a reader who signed in by email. It can change as often as the reader likes until 00:00 UTC on the test’s date; the last number before then is the one the Brier score reads when the test resolves.