---
title: "The Dated Tests — The AI Compute Race"
source: https://dryburgh.com/vol1/tests
description: "What would prove the book wrong, and when: 75 dated limbs, one table a year from 2026 to 2030, each with the record it is read on."
---

# The Dated Tests

What would prove the book wrong, and on what day, from 2026 to 2030.

From *The AI Compute Race*. [Read the book](https://dryburgh.com/vol1#buy)

Updated 29 October 2026

**Where the 75 dated tests stand:** 1 fired · 1 on course to fire · 1 weakened · 9 held so far · 3 unscored · 38 open · 22 with a mixed or partial reading.

- Fired
- On course to fire
- Weakened
- Held so far
- Unscored
- Open
- With a mixed or partial reading

·

The row, in full ·

[Subscribe to the days they fall due](webcal://dryburgh.com/vol1/data/tests.ics) in your calendar, kept current; or [download the file](https://dryburgh.com/vol1/data/tests.ics)

## Next due

1. [10 December 2026](https://dryburgh.com/vol1/tests/2026-4) **Chapter 6, the queue is not a forecast; chapter 14, construction is the binding clock.** More than a quarter of the count of ERCOT’s audit, which sorts the requests in Texas’s large-load queue into the real and the rest, found real puts the queue nearer a forecast than chapter 6 allows and weakens its…
2. [31 December 2026](https://dryburgh.com/vol1/tests/2026-5) **Chapter 4, permanent power.** Withdrawn if the records date the second substation’s entry into service on or before 31 December 2025, the end of the season the Chamber named. Weakened if they date it between 1 January and 11 February 2026, before…
3. [31 December 2026](https://dryburgh.com/vol1/tests/2026-6) **Chapter 12, Southaven’s calendar.** The Mississippi permit board’s evidentiary hearing of 10 November 2026 on the Southaven permit. A permit revoked or given new conditions supports the stronger reading, that consent reprices. A permit upheld as issued…
4. [31 December 2026](https://dryburgh.com/vol1/tests/2026-7) **Chapter 12, Southaven’s calendar.** The last of the fourteen units fitted with catalytic controls, the Solar Titan 130s of the testing row above, falls due.

This section gathers in one place the dated tests the chapters print, in one table for each year. To use it, find the day, read the row, and read the status for which side of its line the reading sits. A row gives the day, the claim the limb tests, the condition and what it does to the claim, the record it is read on, and the status, where the reading stood in October 2026. A limb is one part of a test, scored on its own. Each row is one limb, or two limbs read on the same day and the same record. The claims themselves, and the rules several limbs share, are stated once below, before the tables.

The chapters state their own tests in full, and a row is their index, so a chapter’s wording governs where a row’s differs. Chapter 12, the cross-examination, is the exception. It states claims 2 to 5 and the readings beside them and points here, so those tests are stated in full only in this section; the chapters that hand their claims to it are named under it. The wager’s tests are chapter 1’s, read in the tables under chapters 1 and 12. The Dryburgh Compute Race Scorecard at the back keeps its own thresholds, and the three rows of chapter 14’s movements cite them by row. The dates the chapters name as dates on the record, rather than as tests, are not here.

The tests use one ladder of words, and a word means the same thing in every row. Falsified, withdrawn and fails mean the claim, or the channel named, is given up for the scope the test names. Undermined and weakened mean it stands with less behind it. Supported, confirmed and held mean the record so far favors it. Open means the day has not come. Unscored means the reading was not published by its day, which is never a confirmation and never a failure. Met, said of a channel of claim 5, means the channel’s test has fired, which counts against the book. Fired means a limb has gone against the book before its day.

One rule holds for silence, with two exceptions. A limb whose reading has not been published by its day is unscored. The exceptions are both in claim 5. On the paper that rests on OpenAI’s contracts, an unpublished cash burn leaves the burn half unscored and the limb is read on revenue alone. The debt secured on the chips is read on the one fleet reading that can be scored when the others are silent. Both make the book’s claim easier to withdraw, not harder.

## The claims tested, and the rules they are read by

**Chapter 1.** The wager is that a lead in self-improving AI compounds, and chapter 12 names it the uncatchable lead. Its strong form holds that the first state with sufficiently advanced AI gains a decisive, durable advantage over every rival, including a nuclear peer. The book does not assert the wager; it keeps the test because the buildout is priced as if the wager were true. The wager is tested on two routes, split by pronoun. The first route is exclusivity of the model, with “we” as one laboratory. The second is superiority of the machine around the model, with “we” as the leading American-controlled system or the allied stack behind it. On 31 August 2028 four rows read it. Each route can be supported or undermined on its own. The wager itself is falsified only when both of its conditions hold, and the strong form is confirmed only when all three of its conditions hold. The chapter’s calibration, that the buildout is largest in dollars, not in proportion, has two parts, that it runs at telecom scale and that it is not unprecedented in proportion.

**Chapter 2.** The claim is that compute went on scaling in a different place rather than stopping; if its test fires, compute stopped scaling. The chapter’s ladder of cost is a second claim, that the cost of a final training run goes on climbing, tested apart; if that fires, the cost stopped climbing.

**Chapter 3.** The claim is that the outer loop of research, choosing the question and the test of success, is human. Its five conditions are read twice, a provisional reading in 2027 and a last one in 2028, which is why two conditions met weaken the claim in the first and leave it unverified in the second. The chapter also says of OpenAI’s intern benchmark that nobody else can yet check it, and it draws a distinction between “attempt” and “leave with it,” that a model can attempt longer tasks than it can be left with.

**Chapter 4.** The claim is that the builder’s clock does not set the date of permanent power, and the utility’s gates do. It is tested on the second 150-megawatt substation in Memphis. The builder, xAI, said in August 2025 that the substation was “tracking on 90 to 95 days,” and the Greater Memphis Chamber had said in May 2025 that its connection to the grid was “planned for the Fall of 2025.” The board of the Tennessee Valley Authority (TVA) authorized the additional power on 11 February 2026. An entry into service before that authorization would show the builder’s clock set the date; one after it shows the utility’s gates did. The city’s utility is Memphis Light, Gas and Water (MLGW).

**Chapter 5.** The claim has two halves, scored apart. The concentration half is that one island bonds every accelerator package that matters outside China and carries its memory beside the chip, bonding in the packaging sense, the chip and its memory joined into one package. The memory half is that the memory is booked years ahead in a shortage its makers expect to last. Each half stands or weakens alone; the chapter’s words are stands and weakens, and neither half is withdrawn by a reading. Every reading of 31 December 2027 is provisional, and the last reading of both falls on 31 December 2028. A third reading, NVIDIA’s place in the packaging queue, is read apart and scores neither half.

**Chapter 6.** Three of its claims are read at Meta’s Hyperion campus in Richland Parish, Louisiana, where Entergy Louisiana is building three gas plants for the campus, Franklin Farms units 1 and 2, dated for service in 2028, and Waterford 5, dated for 2029. The Louisiana commission’s docket for them is U-37425, and the monitors are its quarterly monitoring reports. The chapter claims that deliverability and firmness bind there, and it puts its emphasis on the transformer; a plant arriving on time and on its estimate weakens that emphasis. Its overrun-gap claim is that the minimum bill trues up for overruns on the plants while the early-termination fee protects only the original estimate. Its claim that the queue is not a forecast is read on Texas, on ERCOT’s audit of which large-load requests are real, and its claim that firm service is the constraint is read nationally.

**Chapter 7.** The finding has two clauses, that denial slows the frontier’s deployment in China and that, where denial is durable, visible to buyers and backed by Beijing’s reported procurement mandates, it accelerates the Chinese ecosystem beneath it. A fired first clause would say denial did not slow the frontier in China; a fired second, that it did not accelerate the industry beneath it. The chapter also claims that, of the two blockades, Beijing’s is the harder to reverse, and that market access for American AI chips in China is governed by instruction rather than by published rule.

**Chapter 8.** The finding is that sovereignty over a model lives in the mission layer rather than in the license. It can be undermined from two sides. From the state’s side, a statute or an invocation that compels the maker to serve the state undermines it. From the courts’ side, a court letting the state exclude the maker for the limits it builds into its models does. The chapter also claims that the maker, Anthropic, could not reach the deployed instances and that the state switched the access off. The nineteen days ran from 12 to 30 June 2026, when Fable 5 was off for every customer under an export control directive, a letter from the Department of Commerce citing national security authorities, and the model returned on 1 July. On the letter as the Center for Strategic and International Studies reported it, the chapter reads the nineteen days as resting on an authority never before used as the basis for a control. The section of the export regulations it cites reaches only a few adversarial countries, and the state did not publish the directive. In the chapter’s tests, “the Department” is the Department of War.

**Chapter 11.** The claim is that revenue will not arrive on the schedule the paper, the debt sold to finance the campuses, was written to. Each of its tests fires alone, and fires for one laboratory’s one year, because a laboratory meeting its own forecast leaves the industry’s revenue a fraction of what the paper needs. A reported figure means an audited statement in a public filing or, until one exists, a laboratory’s own investor figures as the press reports them, never a run rate.

**Chapter 12.** The cross-examination stands or falls claim by claim, each claim on its own day and on its own definitions, and a claim whose test fires is withdrawn for the scope its test names. Its first claim is chapter 1’s wager. Chapter 4 hands it two claims (items 2 and 4), chapter 9 one (item 3) and chapter 10 one (item 5); chapter 10’s own claim, that the paper keeps the three losses apart and that no lender has yet gone unpaid, is tested under claim 5.

**Claim 2, the moving bottleneck,** is that the binding constraint on compute moves in a predictable order and can be named before it binds rather than after. Either of two findings falsifies it, each alone. The first, read on the 2026 record, is that a constraint named for 2026, memory or permission, did not bind, or that one not named bound instead. It was scored in October 2026 on the record to then, and it is reported in the 2028 table beside the forecast. The second is that the relay the claim assumes is gone, read on 31 December 2028 as a generator step-up transformer’s lead time under a year. For its tests a constraint binds when a dated public record shows it delaying an energization or a delivery, such as the pause, the lead time and the year’s supply already under contract that the chapter lists for 2026. The claim’s forecast for 2027 and 2028 is a third test, and it fails the forecast, not the claim.

**Claim 3, consent as a race,** is 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. Its tests use three words for what consent does. Prevention is a campus stopped. Repricing is a campus made to carry a condition or a cost it did not carry before the challenge, in a permit, an order or an agreement. Conversion is opposition turned into an enforceable schedule, which is repricing’s weaker form. The claim says 2026 moved consent to repricing. A prevention at campus scale would move the finding further against the builders and strengthen the claim. An enforceable bargain kept for a year, dated in the table, would show consent as a cost paid once and falsify it. What 2026 did cannot be undone, so a fall in Data Center Watch’s count can weaken the claim but never falsify it.

**Southaven’s calendar** tests, under claim 3, whether conversion binds, that is, whether the schedule the state’s order wrote is the one the site keeps. A date kept supports conversion as the finding. A date missed, or moved by a later amendment, supports the reading that the schedule was the builder’s to set, and weakens the claim. A date with no public record by its day, in a report or an order, is unscored. The dates are those of the agreed order of 30 July 2026 between the state and MZX Tech LLC, the SpaceX subsidiary that operates the portable turbines at Southaven, Mississippi, for the business then called SpaceXAI; in the rows it is the company. They are read on the reports the order requires of the company and on the orders and amendments the state itself issues. Every two weeks the company must tell the state’s environmental department which turbines it has deactivated, and it must report the results of the emissions testing to the same department. The order does not require the reports to be published, and the department’s online records list permits, orders and inspections but have no place for the reports a company files under an order. The fourteen Solar Titan 130s of the testing row are the fourteen units fitted with catalytic controls, the units whose exhaust is cleaned of nitrogen oxides. One amendment came in September 2026, and it weakens the claim for the two turbines it moved. On 11 September 2026 the state, at the company’s request and “for operational efficiency,” let two turbines run to 5 and 8 May 2027 in exchange for two others that had already stopped.

**Claim 4, announced against energized,** is that announced capacity and energized capacity are different quantities, that the ratio between them is the central fact of the physical race, and that the cohort of announced campuses is not converting on the dates that were promised. Either of two readings falsifies it, each alone. The first is the matured promises, scored on 31 December 2027 against chapter 12’s slippage matrix, forty campuses of which thirty-one carry a dated deadline. The second is the frozen cohort, the forty campuses of the campus ledger of 31 August 2026, scored on the same day. The ratio itself is not the test, because new announcements enter it and refill it even while every builder keeps every date.

**Claim 5, destructive finance,** is that the buildout’s financing has the shape of a leveraged infrastructure mania and that it will destroy capital chiefly if the accelerator-hour becomes cheap or idle before the contracts and the debt have run off. It is withdrawn channel by channel, in four channels (the hyperscalers’ own borrowing, the paper that rests on OpenAI’s contracts, the paper that rests on Anthropic’s, and the debt secured on the chips), and it fails whole when all four are met. The claim expects the hyperscalers’ capital spending to overtake their operating cash flow, as Epoch’s trend of June 2026 had it doing around the third quarter of 2026. That channel is withdrawn only when their cash flow has crossed back above their spending without the spending being cut.

Destructive, wherever the claim uses the word, means a default, a restructuring or a write-down of debt raised for AI infrastructure or of the assets behind it, of $1 billion or more in a single instance, in an issuer’s or a lender’s own filing. Cash burn means negative free cash flow, as the company’s investor figures or an audited statement give it, and if no figure on that measure is published by a limb’s day, the burn half is unscored and the limb is read on revenue alone. A GPU-hour is the rental market’s name for the accelerator-hour. The rent index, the one-year contract index chapter 12 reads, is read on the midpoint of the range it publishes for the last month it has published by the limb’s day, and a midpoint between $1.50 and $2.25 is scored on neither side. A bulk sale, or a fleet, means a single sale of a thousand accelerators or more, and launch price means the $33,600 basis the secondary-market record reconstructs, since NVIDIA has published none. Each fleet reading on the sound side weakens the claim. Each on the other side is the named channel opening, the accelerator-hour cheap or idle before the debt has run off, and it counts as destructive only on the definition above, in a filing.

**Beside the five,** the chapter reads the price of long money as not having repriced for transformative AI. It also holds a hypothesis it does not keep as a claim, that the software developer, the industry’s first and best customer, is also its worst, a customer whose own product deflates the price of what it buys. Confirmed, it would weaken the link the chapters draw between tokens sold and revenue earned.

**Chapter 13.** It claims that mass is money (that a gigawatt in orbit costs more than the same gigawatt on the ground) and that orbit is a change of landlord, not an escape. Each of the three tests of the second claim weakens it on its own, and all three together would falsify it. For the Gulf it claims that the license, not the land, is the constraint. The Gulf is scored on the sunset alone, the day the pathway in Washington’s rule of July 2026 for G42, the Emirati state’s AI company, and Core42, its cloud arm, expires unless renewed.

**Chapter 14.** Its base case for the buildout is a rationed utility, and its reading beside it is that construction is the binding clock. Its verdict for Hyperion is that the campus’s financing is secure, a verdict on the rent Meta owes and not on which future the campus runs in. A slip in the plants weakens it toward a campus short of power rather than moving it to another future. A fired falsifier withdraws the chapter’s own claim; the chapter’s dated falsifier is the base-case row of 31 August 2027. The Scorecard’s verdict, the plurality of its rows graded A to C, the rows read from primary documents and reporting rather than from estimates, is the book’s verdict on which future is arriving, and a chapter’s withdrawal stands beside that verdict, which it does not displace.

The base case can move toward one of the three other futures, and a movement needs all of its signals together. Each movement is read at the Scorecard’s reading of 31 August 2029, when the figures for 2028 have been published, on the Scorecard’s rows and frozen definitions where a row measures the signal, and on the filings the chapters cite where none does. One signal alone is an isolated miss, which weakens the base case without withdrawing it. A movement met by its day withdraws the base case in favor of the future it names. A signal with no reading published by that day is unscored, and a movement with an unscored signal is reported as unscored, not as unmet. The chapter also scores one outside forecast, *AI 2027*’s compute model of April 2025, the scenario a group of forecasters published that month, beside the Scorecard’s clocks and not on its rows.

## 2026

7 tests**THE TESTS OF 2026 — each limb, the record it is read on, and its status in October 2026**

| Day | Claim | The test | Read on | Status |
| --- | --- | --- | --- | --- |
| 27 August 2026 | [Chapter 12, Southaven’s calendar](https://dryburgh.com/vol1/tests/2026-1) | Two turbines, Nos. 15 and 16, due to stop by 27 August 2026 under the agreed order as first written. | The state’s amendment of 11 September 2026 | Weakened for those two turbines. The amendment, made at the company’s request, moved their dates to 5 and 8 May 2027, and two others were retired in their place. |
| 25 September 2026 | [Chapter 8, the mission layer](https://dryburgh.com/vol1/tests/2026-2) | The finding is undermined from the courts’ side if, by 31 December 2027, the D.C. Circuit sustains the procurement exclusion under 41 U.S.C. §4713 on the existing record. The limb is filed under the day it fired. | The D.C. Circuit’s decision | Fired at the panel stage. On 25 September 2026 the D.C. Circuit sustained the exclusion on the existing record, two judges to one, with its mandate, the order that puts a decision into effect, withheld until any petition for rehearing is decided. The finding is undermined from the courts’ side for as long as the panel’s decision stands. Anthropic petitioned for rehearing on 13 October 2026, and the government’s response is due 13 November. |
| 28 September 2026 | [Chapter 12, Southaven’s calendar](https://dryburgh.com/vol1/tests/2026-3) | Emissions testing begun on the fourteen Solar Titan 130s, under the agreed order, by 28 September 2026. | The company’s reports under the agreed order, and the state’s own orders and amendments | Unscored. The date has passed with no public record either way: none of the company’s reports under the order has been published, and no report that testing began has been found. |
| 10 December 2026 | [Chapter 6, the queue is not a forecast; chapter 14, construction is the binding clock](https://dryburgh.com/vol1/tests/2026-4) | More than a quarter of the count of ERCOT’s audit, which sorts the requests in Texas’s large-load queue into the real and the rest, found real puts the queue nearer a forecast than chapter 6 allows and weakens its claim, and says the constraint is physical and near, which supports chapter 14’s reading that construction is the binding clock. A quarter or less, a large cull, says the queue was option value, which confirms chapter 6’s claim and weakens chapter 14’s reading. | The survival rate due that day from ERCOT, the Texas grid operator, as a share of the audit’s own count in gigawatts | The first cut of 3 September kept about 192 of about 494 gigawatts, about 39 percent, above the quarter line, which leans against chapter 6’s claim and toward chapter 14’s reading, but a first cut is not yet a survival rate. |
| 31 December 2026 | [Chapter 4, permanent power](https://dryburgh.com/vol1/tests/2026-5) | Withdrawn if the records date the second substation’s entry into service on or before 31 December 2025, the end of the season the Chamber named. Weakened if they date it between 1 January and 11 February 2026, before TVA’s authorization. Confirmed if they date it after 11 February 2026. | MLGW’s and TVA’s posted records | Open. No posted utility minute records the second substation entering service. |
| 31 December 2026 | [Chapter 12, Southaven’s calendar](https://dryburgh.com/vol1/tests/2026-6) | The Mississippi permit board’s evidentiary hearing of 10 November 2026 on the Southaven permit. A permit revoked or given new conditions supports the stronger reading, that consent reprices. A permit upheld as issued leaves conversion as the finding. No decision by that day leaves the hearing unscored. | The permit board’s decision | Open. The hearing was set for 10 November 2026. |
| 31 December 2026 | [Chapter 12, Southaven’s calendar](https://dryburgh.com/vol1/tests/2026-7) | The last of the fourteen units fitted with catalytic controls, the Solar Titan 130s of the testing row above, falls due. | The company’s reports under the agreed order, and the state’s own orders and amendments | Open. The amendment of 11 September left this date where it was. No deactivation report has been published. |

## 2027

32 tests**THE TESTS OF 2027 — each limb, the record it is read on, and its status in October 2026**

| Day | Claim | The test | Read on | Status |
| --- | --- | --- | --- | --- |
| 6 April 2027 | [Chapter 13, the license, not the land](https://dryburgh.com/vol1/tests/2027-1) | The G42 pathway renewed past its sunset of 6 April 2027 without a change of ownership would weaken the claim. A lapsed pathway would confirm it, and so would a renewal that came only with a change of ownership. A new American vehicle taking the license in G42’s place would count as that change. | The Commerce Department’s record | Open. Stargate UAE’s first 200 megawatts, due energized by 30 September 2026 on its builders’ schedule of December 2025, are read as a schedule indicator, not as a test of this claim, because a missed quarter has other causes on the record, the war with Iran that began on 28 February 2026 among them. No announcement that the phase was energized has been found by the end of October, so the indicator reads late. |
| 14 July 2027 | [Chapter 12, Southaven’s calendar](https://dryburgh.com/vol1/tests/2027-2) | The last of the sixty-eight portable turbines that reached the site falls due. | The company’s reports under the agreed order, and the state’s own orders and amendments | Open. Eleven units are retired by the amendment of 11 September: nine against the first eleven scheduled dates, 18 to 28 August 2026, which the amendment marks retired without saying on which day each stopped, and two in exchange for the two turbines it moved to May 2027. Under the calendar’s rule the nine undated retirements are unscored, neither met nor missed. The amendment left this date where it was, and no deactivation report has been published. |
| 31 August 2027 | [Chapter 1, the calibration](https://dryburgh.com/vol1/tests/2027-3) | That the buildout runs at telecom scale is falsified if American AI investment for 2026 stands above 3 percent of GDP for 2026, a share above the 2.8 percent Goldman Sachs’s own projection gives for 2028. That it is not unprecedented in proportion is falsified only if the same figure stands above the 7.3 percent of Britain’s railway mania in 1847. If Goldman has published no count for 2026 by that day, both parts are unscored. | Investment on Goldman Sachs’s perimeter, and GDP on the Bureau of Economic Analysis’s figure | Held so far. American AI investment in 2026 is projected at about 1.8 percent of GDP on Goldman Sachs’s perimeter, or $581 billion, under the 3 percent line. |
| 31 August 2027 | [Chapter 3, the outer loop](https://dryburgh.com/vol1/tests/2027-4) | Withdrawn if the public record shows a laboratory giving a system a broad mandate rather than a problem, the system choosing the question and the test of success, the result reproduced by an unaffiliated laboratory and adopted into the next production model, and a third party auditing the prompts and the interventions. All five conditions are needed to withdraw it. A broad mandate with the system choosing the question and the test of success weakens it. | The public record | Open. The inner loop of research is automated and the outer loop is human. |
| 31 August 2027 | [Chapter 3, the intern](https://dryburgh.com/vol1/tests/2027-5) | The sentence “nobody else can yet check” is withdrawn if OpenAI publishes the intern benchmark, trajectories and a reliability rate. | OpenAI’s own publication | Open. OpenAI said on 6 September 2026 that the intern goal had been reached, on measurements of its own that it calls preliminary, and as of the end of October 2026 it has published no task-level record anyone else can examine. |
| 31 August 2027 | [Chapter 3, “attempt” and “leave with it”](https://dryburgh.com/vol1/tests/2027-6) | The distinction weakens if METR’s 80-percent horizon passes eight hours on a task suite long enough to measure it. | METR’s own published measurements | The distinction holds: 3.1 hours, the 80-percent figure on METR’s dashboard for Claude Mythos Preview in May 2026, well short of the eight-hour line. GPT-6 Astra had no published METR horizon, and since 8 September 2026 METR’s page has said it is no longer being updated, so the test needs a successor measurement to be scored. |
| 31 August 2027 | [Chapter 11, OpenAI’s 2026](https://dryburgh.com/vol1/tests/2027-7) | The claim fails for OpenAI’s 2026 if OpenAI’s reported revenue for 2026 reaches the $30 billion of its own forecast. The test keeps that forecast, though a later OpenAI presentation, which the Financial Times reported on 18 September 2026, put 2026 revenue at $36 billion, a figure that would also fire it; a later report, the Financial Times of 9 October 2026, put the company’s own forecast revised to $35 billion from $30 billion. | OpenAI’s reported revenue | On course to fire. OpenAI’s reported revenue for the first half of 2026 was $12.4 billion, by this book’s sum of its two quarters, so $30 billion for the year needs about $17.6 billion in the second half, an annual rate of about $35 billion. On the run rates reported since, the pace the second half needs is below the pace the company reports: close to $30 billion in July and approaching $50 billion by the end of September, on the company’s presentation to investors as the Financial Times reported it on 8 October 2026. The claim is held in October 2026. |
| 31 August 2027 | [Chapter 12, claim 5, the paper that rests on OpenAI’s contracts](https://dryburgh.com/vol1/tests/2027-8) | OpenAI’s revenue for 2026 against its own forecast of $30 billion, with its cash burn against the forecast’s $25 billion. One year met alone weakens the claim for that paper and withdraws nothing, since the claim is withdrawn for it only if 2026 and 2027 both hold. | OpenAI’s reported revenue, and its cash burn | The revenue half is on course to reach the forecast, as chapter 11’s test of the same day shows; the burn half is unscored, no figure on that measure having been published. |
| 31 August 2027 | [Chapter 14, the base case](https://dryburgh.com/vol1/tests/2027-9) | Withdrawn if the reading shows construction with fewer than three-quarters of its voting rows for rationing, or capability with a verified research milestone. | The Scorecard’s frozen definitions, at its reading of 31 August 2027 | Held so far. Construction has all eight of its voting rows for rationing, two of its ten rows abstaining. The research milestone is unverified on any public record, and OpenAI’s claim of 6 September 2026, on preliminary measurements of its own, is one on which the Scorecard’s row 5 still abstains. |
| 2 September 2027 | [Chapter 3, the outer loop, on Anthropic’s date](https://dryburgh.com/vol1/tests/2027-10) | The twelve months of Tom Brown, Anthropic’s chief compute officer, who said on 2 September 2026 that “a once-in-a-generation scientist for key scientific fields” could come within twelve months. The claim weakens if Anthropic shows on the public record a system choosing its own question and test of success in a scientific field, with the result reproduced by an unaffiliated laboratory. An account by Anthropic, published by that day, of its most capable research system that shows a person still choosing the question or the test of success confirms it. | The public record | Open. |
| 31 December 2027 | [Chapter 2, the ladder of cost](https://dryburgh.com/vol1/tests/2027-11) | Falsified if no training run exceeds $1 billion by 31 December 2027, on Epoch AI’s reconstruction or a laboratory’s own disclosure. | Epoch AI’s reconstructions of training cost, and the laboratories’ own disclosures | Open in October 2026; the largest reconstructed run is about $490 million. |
| 31 December 2027 | [Chapter 5, the concentration half](https://dryburgh.com/vol1/tests/2027-12) | A provisional reading. It stands if no line outside Taiwan is bonding the chapter’s accelerators, CoWoS-class packages (chip on wafer on substrate) with memory beside the chip, in volume production qualified by their buyers. A weakening in 2027 would have to come by another route than the reported lines, Intel’s packaging taking some of the work TSMC’s chairman said he hoped it would share, so the live test is the last reading. | The packagers’ own statements | Every accelerator that matters outside China and carries its memory beside the chip passes through one island’s lines. The earliest lines reported off the island are Amkor’s in Arizona, targeted for early 2028, and TSMC’s own there, by 2029. On 21 October 2026 NVIDIA and Amkor said they planned to expand advanced packaging in South Korea, targeting more than a threefold increase in interposer wafer capacity by the end of 2028, naming no plant, product or volume. The half stands. |
| 31 December 2027 | [Chapter 5, the memory half](https://dryburgh.com/vol1/tests/2027-13) | A provisional reading. It stands if the memory shortage SK hynix forecast through 2030 is still on its makers’ own calls, and weakens if it is not. | The makers’ own calls | SK hynix’s own forecast was of shortage through 2030, and Micron had agreed price and volume for its entire 2026 supply of high-bandwidth memory by 17 December 2025. |
| 31 December 2027 | [Chapter 5, NVIDIA’s place in the queue, read apart](https://dryburgh.com/vol1/tests/2027-14) | NVIDIA’s majority of the packaging queue weakens if CoWoS capacity reaches the 2027 estimate of about 170,000 wafers a month while NVIDIA’s share of it falls below half. The reading scores neither half, because a falling share would measure concentration among the buyers, not the island’s. | The trade press’s estimates or Epoch AI’s model | NVIDIA held a rising majority of CoWoS capacity across 2024 and 2025, on Epoch AI’s model. |
| 31 December 2027 | [Chapter 7, the frontier’s deployment](https://dryburgh.com/vol1/tests/2027-15) | An audited frontier model trained end to end on domestic Chinese silicon, at a compute budget within a factor of two of the American frontier, weakens the clause that denial slows the frontier’s deployment in China and confirms the clause that durable denial accelerates the ecosystem beneath it. The audit counts only if a party outside the developer publishes it. | The audit, and Epoch AI’s frontier training series for the American frontier | Open. No frontier model has been trained end to end on domestic silicon on the record. |
| 31 December 2027 | [Chapter 7, the ecosystem beneath it](https://dryburgh.com/vol1/tests/2027-16) | Huawei’s Ascend shipments for 2027 at or below the 805,000 SemiAnalysis estimated for 2025 weaken the clause that denial, where durable, visible to buyers and backed by reported procurement mandates, accelerates the Chinese ecosystem. | SemiAnalysis’s estimate for 2027, published by 31 December 2027 | Open. |
| 31 December 2027 | [Chapter 7, the harder to reverse, Beijing’s side](https://dryburgh.com/vol1/tests/2027-17) | Beijing’s guidance lifting, with mainland deliveries of licensed H200s or their successors in the calendar year to 31 December 2027 exceeding half of what Washington has licensed to Chinese buyers by that day, withdraws the claim. Both are needed to withdraw it, and either alone weakens it. The threshold is set against what Washington has licensed, so that Washington’s own ceiling cannot be what keeps it from being met. | NVIDIA’s filings or the press reports the chapter reads | Held so far. By August about 10,000 licensed chips each had reached ByteDance and Tencent against the 100,000 each Washington had cleared, a tenth, under the half the test asks for; NVIDIA said the shipments that had reached China were under one percent of its datacenter revenue in the quarter then reported. |
| 31 December 2027 | [Chapter 7, the harder to reverse, Washington’s side](https://dryburgh.com/vol1/tests/2027-18) | No change by Washington, in either direction, to which NVIDIA accelerators may be licensed for China or on what terms, between the H200 rule of January 2026 and 31 December 2027, weakens the claim, which rests in part on Washington’s reversals of 2025. A further reversal by Washington in that time, while Beijing’s guidance holds, confirms it. | The Federal Register and NVIDIA’s filings | Open. |
| 31 December 2027 | [Chapter 7, instruction rather than rule](https://dryburgh.com/vol1/tests/2027-19) | A published central legal text for the customs instruction or the state-datacenter ban withdraws the claim. A provincial or municipal reply describing central guidance is not one. | A published central legal text | No published directive has been found. |
| 31 December 2027 | [Chapter 8, the mission layer](https://dryburgh.com/vol1/tests/2027-20) | The finding is undermined from the state’s side if Congress enacts a continuity or compulsion statute, one that requires a supplier to keep serving the state or to license lawful uses on demand, or if the Defense Production Act is invoked against a model’s maker and upheld; the Act’s authorities terminate on 11 December 2026 unless extended, and if they lapse unrenewed that limb is unscored. | The statute, or the invocation and any ruling on it; the Act’s extension | Open. |
| 31 December 2027 | [Chapter 8, the mission layer](https://dryburgh.com/vol1/tests/2027-21) | The finding is undermined from the courts’ side if the Ninth Circuit reverses the district court’s judgment of 27 August 2026 in California, which set aside the designation and the boycott and permanently enjoined the directives. | The appellate court’s decision | Open. The government gave notice of its appeal to the Ninth Circuit on 26 October 2026. |
| 31 December 2027 | [Chapter 8, the maker could not reach the deployed instances](https://dryburgh.com/vol1/tests/2027-22) | Falsified if the record shows a vendor-side alteration or shutdown of a deployed government instance. | A court record or the Department’s own filings | Open. |
| 31 December 2027 | [Chapter 8, the nineteen days](https://dryburgh.com/vol1/tests/2027-23) | The reading is withdrawn if an open copy of the June directive shows that it cites a published authority that expressly provides for a worldwide license requirement. If no open copy is published by that day, the limb is unscored. | An open copy of the directive | Cannot yet be scored. Neither the government nor Anthropic has published the letter, as far as a search of the record could find; an unofficial transcription open since 16 June 2026 names three authorities, none of which expressly provides for a worldwide license requirement. Bloomberg’s printing of the same text, behind its subscription wall, is neither an open copy nor the letter itself. |
| 31 December 2027 | [Chapter 8, the state switched the access off](https://dryburgh.com/vol1/tests/2027-24) | Withdrawn if the record shows that the withdrawal was the company’s own decision rather than the state’s. | A court record or the Department’s own filings | Open. |
| 31 December 2027 | [Chapter 12, claim 4, the promises](https://dryburgh.com/vol1/tests/2027-25) | Falsified if three in four or more of the dated promises that have matured and can be scored by then were kept on their promised phase dates, scored as the slippage matrix scores them, on evidenced load and not the builder’s word. | The slippage matrix, chapter 12’s second list of forty | Held so far. Of the five dated promises that had matured, one could not be scored; of the four scorable, two were missed outright, one was kept only on the builder’s own account, which the test does not count, and one was missed unless “early 2026” means the first quarter: at most one of four kept, under the three-in-four line. |
| 31 December 2027 | [Chapter 12, claim 4, the frozen cohort](https://dryburgh.com/vol1/tests/2027-26) | Also falsified if the forty campuses of the campus ledger as read at the end of August 2026 have energized, on evidenced load, at least the sum of the phases they had dated for that day, so that conversion kept its promised pace whatever new announcements add. | The campus ledger | The ledger’s ratio stood at eight to one at the end of August 2026, 38.4 gigawatts announced or estimated against 4.8 evidenced, and about eight and a half to one on Epoch’s dataset of 9 October 2026, 38.4 against 4.56. The ratio is not the test; the test is what the forty had dated for 31 December 2027 against what they have energized, read on that day. |
| 31 December 2027 | [Chapter 12, claim 5, the hyperscalers’ own borrowing](https://dryburgh.com/vol1/tests/2027-27) | Withdrawn for the hyperscalers’ own borrowing if the reported operating cash flow of the hyperscalers (Microsoft, Amazon, Alphabet, Meta and Oracle together) has re-crossed their reported capital spending (capex) without capex being cut. Both are summed over the last four quarters each company has reported by that day, and capex counts as cut if that sum is lower than the same companies’ capital spending in the four quarters before. | The five companies’ reported operating cash flow and reported capital spending | Held so far, and moving the claim’s way. In the first quarter of 2026 the five together still took in more from operations than they spent, $157.9 billion against $148.4 billion; in the third quarter two of them crossed, Oracle in its quarter to 31 August and Meta in its quarter to 30 September, each spending more on capital than it took in. The channel is withdrawn only when the five together have crossed back. |
| 31 December 2027 | [Chapter 12, claim 5, the first fleet reading, rent](https://dryburgh.com/vol1/tests/2027-28) | The 2023–24 H100 fleets re-contracting at $2.25 a GPU-hour or more weakens the claim. Term rent below $1.50 is the named channel opening. | The one-year contract index, on the midpoint rule | On the sound side: $2.80 in July 2026, the midpoint of $2.40 to $3.20, a range wholly above the $2.25 line, July being the last month in the index’s data when this book archived it on 1 September 2026. In March 2026 the midpoint was $2.35. The index gives no single point for any month, only the middle half of its prices. |
| 31 December 2027 | [Chapter 12, claim 5, the second fleet reading, utilization](https://dryburgh.com/vol1/tests/2027-29) | Utilization of the 2023–24 H100 fleets holding at 75 percent or more weakens the claim. Utilization below 60 percent is the named channel opening. Without a series that can separate those fleets by 31 December 2027, the reading is set aside as unscored. | A published series that can separate the 2023–24 fleets | Unscored. No such series exists; the one market measure on the record, the rented share of the capacity Ornn tracks, from a marketplace that both prices compute and rents it, has nothing to separate a 2023–24 fleet from any other. |
| 31 December 2027 | [Chapter 12, claim 5, the third fleet reading, resale](https://dryburgh.com/vol1/tests/2027-30) | A bulk sale of the first H100s, delivered in 2022 and five years old by the end of 2027, clearing at a quarter or more of launch price weakens the claim. A fleet clearing below 15 percent is the named channel opening. With no such sale by 31 December 2027, the reading is set aside as unscored. | A bulk sale, against the reconstructed launch price | Unscored. No fleet has cleared in public. Twelve eBay listings of single H100s sold second-hand in the ninety days to 31 August 2026 at a median of 45 percent of a reconstructed launch price. |
| 31 December 2027 | [Chapter 12, claim 5, the debt secured on the chips](https://dryburgh.com/vol1/tests/2027-31) | If two or more readings can be scored and all of those fall on the sound side, the claim is withdrawn for the debt secured on the chips. With fewer than two scorable, the channel is read on the one that can be scored, as the fallback, and it is unscored, not held, only when none can. | The three fleet readings above | Read on rent alone, which sits on the sound side: only the rent reading has a series, so the channel can be met only through the fallback, and so far it is not met. |
| 31 December 2027 | [Chapter 14, AI 2027, an outside forecast](https://dryburgh.com/vol1/tests/2027-32) | *AI 2027*’s compute model of April 2025 put global AI power at about 60 gigawatts by the end of 2027, the leading laboratory at about 10 gigawatts of peak power and 15 to 20 percent of the world’s AI compute, and a thousand-times-GPT-4 training run in 2027. Its four figures are read as three numbers. Hitting them would count as evidence for the capability and construction clocks moving upward together, though no row of the Scorecard reads them. Missing all three by half would falsify that forecast’s timetable. Missing by half, by this book’s arithmetic, means energized load under 30 gigawatts, a leading laboratory under 5 gigawatts and under 7.5 percent of the world’s AI compute, and no run of 500 times GPT-4’s training compute. | Energized load, the leading laboratory’s fleet, and Epoch’s largest run | Open. |

## 2028

20 tests**THE TESTS OF 2028 — each limb, the record it is read on, and its status in October 2026**

| Day | Claim | The test | Read on | Status |
| --- | --- | --- | --- | --- |
| 31 August 2028 | [Chapters 1 and 12, the first route](https://dryburgh.com/vol1/tests/2028-1) | Supported if one laboratory holds an independently measured frontier lead through three or more complete improvement cycles, which at a three-month training run is nine months or more without a break. Undermined if American leadership has gone on rotating between laboratories since 1 January 2027 at a median laboratory tenure under thirteen weeks, less than one such cycle and near the present rate of about ten weeks. Both limbs count laboratories: a laboratory’s consecutive leading models make one tenure. A record that meets neither limb, a median of thirteen weeks or more with no lead held for nine months, scores neither. If Epoch has published no count of lead periods for that window by that day, the limb is unscored. | Epoch AI’s capability index, and Epoch’s count of lead periods, in which any lead counts | Since GPT-4’s 352-day lead the top has passed from model to model at a median of about seven weeks, and from laboratory to laboratory at a median of about ten, with OpenAI on top 80.2 percent of the time, a laboratory median under the thirteen-week line, so the record so far leans to undermined. |
| 31 August 2028 | [Chapters 1 and 12, the second route](https://dryburgh.com/vol1/tests/2028-2) | Supported if the leading American-controlled system, or the allied stack behind it, sustains a lead of more than fourteen months over the strongest Chinese system with the gap widening, both at once. Undermined if the lag is fourteen months or less. The lag is a proxy for the machine, not a measure of it. It stands in because the lead the wager needs is a lead in what the models can do, and a superior machine serves the wager only if it yields one. | The Scorecard’s China-lag thresholds, on Epoch AI’s count | Leaning to undermined: China’s frontier has trailed America’s by seven months on average since 2023, on Epoch’s count of January 2026, in a range of four to fourteen. |
| 31 August 2028 | [Chapters 1 and 12, the wager](https://dryburgh.com/vol1/tests/2028-3) | Falsified if both of two things hold. Rivals are still reproducing each advance within fourteen months, the top of the historical four-to-fourteen-month envelope, so that a faster catch-up counts too. The public record still shows the leader’s generation time, the time one generation of models takes to produce the next, at months rather than weeks, which chapters 1 and 12 read as more than eight weeks between successive frontier releases. The lag alone undermines the second route, and the generation time alone falsifies nothing. | The lag on Epoch AI’s count of the China lag, as the Scorecard’s China-lag row reads it, and the generation time on the interval between successive frontier models from the leading laboratory, by the release dates Epoch AI’s capability index lists | Not established. The lag is inside fourteen months and the generation time is months rather than weeks, so both conditions hold on the record so far, and the test is read on 31 August 2028; no regime change, a lead widening rather than turning over, has been observed. The book does not assert the wager, as the preamble says, and keeps the test because the buildout is priced as if it were true. |
| 31 August 2028 | [Chapters 1 and 12, the strong form](https://dryburgh.com/vol1/tests/2028-4) | Confirmed only if public evidence shows all three of its conditions, that one model generation does most of the research for the next, that the cycle materially shortens and that the resulting advantage converts into a capability rivals cannot effectively counter. Each condition alone weakens nothing and confirms nothing. | Public evidence, verified for the Scorecard by its audited milestone for the automated researcher | No public evidence yet supports it. |
| 31 August 2028 | [Chapter 3, the outer loop](https://dryburgh.com/vol1/tests/2028-5) | The five conditions of the 2027 row (a broad mandate, the system choosing the question and the test of success, reproduction by an unaffiliated laboratory, adoption into the next production model, an outside audit) are set against 31 August 2028, five months after the March 2028 date for the true automated AI researcher. Met by then, they withdraw the claim. The first two met with any of the other three unmet leave it unverified, neither withdrawn nor confirmed. A published account by OpenAI of the researcher it dated to March 2028, or by any laboratory of its most autonomous research system, that shows a person still choosing the question or the test of success confirms it. A record that shows none of these three outcomes is unscored. | The public record | Open. The researcher OpenAI dated to March 2028 has not arrived. |
| 31 August 2028 | [Chapter 11, OpenAI’s 2027](https://dryburgh.com/vol1/tests/2028-6) | The claim fails for OpenAI’s 2027 if its reported revenue for 2027 reaches about $62 billion, the company’s own forecast for the year. | OpenAI’s reported revenue | Open. |
| 31 August 2028 | [Chapter 12, claim 5, the paper that rests on OpenAI’s contracts](https://dryburgh.com/vol1/tests/2028-7) | Withdrawn for that paper if OpenAI’s revenue reaches its own forecasts for both 2026 and 2027 while its cash burn stays within them, $25 billion for 2026 and $57 billion for 2027. The 2027 year is read on this day. Revenue met with the burn above the forecast, or one year alone, weakens the claim for that paper and withdraws nothing. | OpenAI’s reported revenue, and its cash burn | Open. |
| 31 December 2028 | [Chapter 5, the concentration half](https://dryburgh.com/vol1/tests/2028-8) | The last reading. It weakens if advanced packages are bonded outside Taiwan in qualified volume production and the share of the relevant supply made off the island is rising from the end of 2027 to the end of 2028. | Bonding on the packager’s own statement, and the share on the trade press’s estimates | As for the reading of 31 December 2027. |
| 31 December 2028 | [Chapter 5, the memory half](https://dryburgh.com/vol1/tests/2028-9) | The last reading, made on the same calls by the same rule. It stands if the memory shortage SK hynix forecast through 2030 is still on its makers’ own calls, and weakens if it is not. | The makers’ own calls | As for the reading of 31 December 2027. |
| 31 December 2028 | [Chapter 6, the transformer](https://dryburgh.com/vol1/tests/2028-10) | Each Franklin Farms unit that enters service by 31 December 2028 within the May 2026 cost estimate weakens the chapter’s emphasis on the transformer. | Entergy’s filings and the U-37425 quarterly monitors | Open. Franklin Farms 1 and 2 are dated for service in 2028. |
| 31 December 2028 | [Chapter 6, the overrun gap](https://dryburgh.com/vol1/tests/2028-11) | Withdrawn if an unsealed filing shows the early-termination fee tracking the revised rather than the original estimate, and unscored if no such filing is unsealed by that day. The claim stands, with the gap wider, if the monitors show the estimate risen again above $4.421 billion and no filing has shown the fee tracking it; with no filing unsealed and the estimate risen, the fee limb is unscored and the gap limb stands. | Unsealed filings, and the U-37425 quarterly monitors | The estimate had risen from the docket’s original $3.956 billion to $4.421 billion by 15 May 2026, up about 11.8 percent, so the gap limb stands wider and the fee limb is open. |
| 31 December 2028 | [Chapter 6, firm service is the constraint](https://dryburgh.com/vol1/tests/2028-12) | Ten gigawatts of new datacenter load energized in the United States on connection terms that let the system curtail it, counted as the contracted interruptible megawatts at named sites and not as whole campuses, would weaken the claim. The same ten gigawatts energized at least twelve months before the transmission upgrades they were waiting for, on curtailed load rather than on generation or batteries brought to the site, would show that the constraint the chapter names was the wrong one, and withdraw the claim. | The counts that utilities’ filings and state commissions’ dockets publish by that day | Open. The public record holds neither count. The largest published figures are two of about a gigawatt (the demand-response capacity Google wrote into its contracts with five utilities and OpenAI’s up to 1,000 megawatts at its campus in Effingham County, Georgia, for power delivered in phases between 2028 and 2032), and neither is energized curtailable load. |
| 31 December 2028 | [Chapter 12, claim 2, the relay](https://dryburgh.com/vol1/tests/2028-13) | Falsified if the parallel investment of every layer at once has removed the relay the claim assumes. The relay is removed if a generator step-up transformer’s lead time is under a year, the threshold chapter 14 gives for the commodity future. If no such count for 2028 is published by that day, the limb is unscored. | One figure, a generator step-up transformer’s lead time | More than 160 weeks in the first quarter of 2026, up from 143 in 2024, on the count of Ben Boucher, a senior analyst with the consultancy Wood Mackenzie. |
| 31 December 2028 | [Chapter 12, claim 2, the forecast](https://dryburgh.com/vol1/tests/2028-14) | The claim names one constraint for each year: for 2027 permission (the permits and the consent that the 2026 record already shows binding), and for 2028 the bond market’s patience. If over a year the named constraint accounts for no delay on a dated public record while another constraint, turbines or memory for example, accounts for most of what was delayed, the claim’s mechanism survives and its forecast fails; one delayed delivery proves nothing either way. | Dated public records of a delayed energization or delivery, and for the bond market’s patience, how many times over a builder’s bond sale is covered and the spread AI-related financings pay over their benchmark | Held for 2026, on the record to October 2026. The constraint moved where the chapters had placed it, to memory and to permission (the queue and the permit), and the grid’s equipment still bound behind them. |
| 31 December 2028 | [Chapter 12, claim 3, the count](https://dryburgh.com/vol1/tests/2028-15) | Weakens if, in the four quarters to 30 September 2028, Data Center Watch counts fewer projects blocked or delayed than the at least 45, valued at nearly $68 billion, that it counted in the second quarter of 2026 alone. If the tracker has published none for those quarters by that day, the limb is unscored. | Data Center Watch’s count | At least 45 projects in the second quarter of 2026, valued at nearly $68 billion, figures for delay, not prevention. |
| 31 December 2028 | [Chapter 12, claim 3, a bargain kept](https://dryburgh.com/vol1/tests/2028-16) | Falsified by evidence that consent is a cost that can be paid once, an American campus of at least 100 megawatts where a community that opposed it accepted an enforceable bargain, and where a year of operation under that bargain then passed without a breach filing or a renewed challenge. | The record of the bargain, and of any breach filing or challenge | Held so far. No such campus was found on the American record to September 2026, a finding about the record searched, not proof that no such place exists. |
| 31 December 2028 | [Chapter 12, claim 3, prevention](https://dryburgh.com/vol1/tests/2028-17) | A proven prevention above 500 megawatts would move the finding from conversion to prevention at campus scale, which strengthens the claim. | Primary documents, on the most hostile standard of evidence | One prevention proven on the standard’s cause, refusal and act, in Oldham County, Kentucky, at a load the local press put at roughly 100 megawatts. On the hostile standard (an announced load, a final refusal, the developer’s act and its admission of cause, each in a primary document) no campus above 500 megawatts yet qualifies; the open gigawatt case is La Osa in Arizona, refused by Pinal County on 26 August 2026, with no termination found. |
| 31 December 2028 | [Chapter 12, claim 5, the destructive direction](https://dryburgh.com/vol1/tests/2028-18) | If the financing has proved destructive by then by a different mechanism than the claim names, the claim stands in its shape and its named channel is what it got wrong; this direction counts against the named channel and never against the claim whole, an asymmetry the claim accepts. | An issuer’s or a lender’s own filing, on the definition of destructive above | Not found. |
| 31 December 2028 | [Chapter 12, beside the five, the price of long money](https://dryburgh.com/vol1/tests/2028-19) | Withdrawn if the ten-year US real rate averages above 5 percent over any twelve consecutive months ending on or before 31 December 2028, the top of the developed-country history the economists Trevor Chow, Basil Halperin and Zachary Mazlish assembled. A rise short of that, however large, leaves it held, and since a real rate moves for reasons other than AI, a crossing would show that the price moved, not why. | The inflation-indexed ten-year Treasury’s yield | Held: 2.44 percent on 31 August 2026 and 2.83 on 25 September 2026, both far under the 5 percent line. |
| 31 December 2028 | [Chapter 12, beside the five, the coder hypothesis](https://dryburgh.com/vol1/tests/2028-20) | If the coder’s share of tokens falls while the token price falls, the hypothesis is confirmed and the link the chapters draw between tokens sold and revenue earned weakens. The share at the end of 2027 is set against the more than half it reached by the end of 2025, and with no such count published by that day, the test is unscored. | OpenRouter’s programming share of token volume | Open. About 11 percent in May 2025, where the category series chapter 12 reads begins, and more than half by the end of the year. |

## 2029

11 tests**THE TESTS OF 2029 — each limb, the record it is read on, and its status in October 2026**

| Day | Claim | The test | Read on | Status |
| --- | --- | --- | --- | --- |
| 31 March 2029 | [Chapter 2, compute moved rather than stopped](https://dryburgh.com/vol1/tests/2029-1) | Falsified if the frontier of Epoch’s capability index keeps its slope in calendar 2027 and 2028, advancing at least 14 points in each year, the pace Epoch reported on 1 September 2026, while the growth of compute slows sharply everywhere at once in both years. That means Epoch’s frontier training-compute series growing less than twofold in each year, no offsetting rise in any post-training or inference estimate Epoch publishes for those years, and accelerator demand flat, NVIDIA’s reported Data Center revenue no higher in either of its fiscal years 2028 and 2029 than in the year before it. Every one of those conditions must hold together, and any one alone, frontier training compute slowing on its own included, leaves the claim standing. If Epoch has not published its index or its training-compute series for those years by the reading date, the test is unscored. | Epoch’s capability index, its frontier training-compute series and its post-training and inference estimates, and NVIDIA’s reported Data Center revenue | Open. Since the first reasoning models of September 2024 the index’s frontier has advanced 14 points a year, Epoch’s reading of 1 September 2026. |
| 1 June 2029 | [Chapter 6, deliverability and firmness bind](https://dryburgh.com/vol1/tests/2029-2) | The claim stands if either Franklin Farms unit slips past the 1 June 2029 rent date. Waterford 5’s own schedule already sits after it. | The U-37425 quarterly monitors | Open. Franklin Farms 1 and 2 are dated for service in 2028, and Meta’s rent on the campus’s buildings starts on 1 June 2029. |
| 1 June 2029 | [Chapter 14, the campus’s financing](https://dryburgh.com/vol1/tests/2029-3) | The verdict that the campus’s financing is secure stands while the monitors show the plants on schedule and S&P’s reports show the project’s coverage above 1.10 times. It is weakened, toward a campus short of power, if either Franklin Farms unit slips past 1 June 2029. | Hyperion’s 2028 energization notices and the U-37425 quarterly monitors | Open. S&P’s case for Hyperion’s notes, issued by Beignet Investor LLC, rests on a flat 1.12-times coverage to 2049, two hundredths above the 1.10 below which it says it could take a negative action. |
| 1 June 2029 | [Chapter 14, the hour, on a proxy](https://dryburgh.com/vol1/tests/2029-4) | No filing reports Hyperion’s utilization or what an hour is worth to Meta, so the chapter’s question, and its reading that the campus can run as a strategic engine, are read on a published rent against the break-points of chapter 14’s model of Hyperion’s four futures, on SemiAnalysis’s $3.1 million price for the rack-scale server. Above $3.01 an hour the reading stands, with the strategic engine first on this model. Between $2.66 and $3.01 the strategic engine yields first place to the rationed utility, and the reading is weakened. Below $2.66 a strategic hour no longer covers its rent on this model, and the reading is withdrawn. A market rent is not Meta’s own value, which is why this limb is a proxy, and if neither series is published on the day, the limb is unscored. | Silicon Data’s B200 rental index, or the CME contract on it if one has listed | The reading stands so far: Silicon Data’s B200 index entered September 2026 at $5.62 an hour, above the $3.01 line; CME’s contracts have not listed. |
| 31 August 2029 | [Chapters 11 and 12, Anthropic’s 2028](https://dryburgh.com/vol1/tests/2029-5) | Chapter 11’s claim fails for Anthropic’s 2028 if Anthropic’s reported revenue for 2028 reaches $190 billion, the low end of the forecast Reuters’s sources described in August 2026. The same figure withdraws chapter 12’s fifth claim for the paper that rests on Anthropic’s contracts. | Anthropic’s reported revenue | Open. Anthropic’s run rate was $65 billion a year in July 2026, and by 19 September 2026 its backers forecast that it would exceed $120 billion by the end of the year, figures the test does not score, given as the direction of travel against the $190 billion line. |
| 31 August 2029 | [Chapter 11, revenue against what was committed](https://dryburgh.com/vol1/tests/2029-6) | The claim fails against what was committed if OpenAI’s reported revenue for 2028 reaches $175 billion, one-eighth of the about $1.4 trillion of commitments over eight years that Sam Altman, OpenAI’s chief executive, restated in writing in November 2025, by this book’s arithmetic. | OpenAI’s reported revenue | Open. |
| 31 August 2029 | [Chapter 12, claim 5, the claim whole](https://dryburgh.com/vol1/tests/2029-7) | The claim fails whole when all four channels are met, which cannot be known before this day. | The four channels’ readings above | Open. Only the rent reading had a series, so the debt secured on the chips can be met only through the fallback. |
| 31 August 2029 | [Chapter 14, the movement upward, toward a strategic engine](https://dryburgh.com/vol1/tests/2029-8) | By 31 December 2028, a verified autonomous-research result meeting the chapter’s definition, and queues turning into load without credit or consent giving way, with 2028 a second year, after 2027, in which energized additions rise while spreads stay tight and the permits flow. | The research-milestone row (row 5) voting strategic engine. The gigawatts newly evidenced in the frozen cohort of forty campuses (row 12) larger in 2027 than in 2026, and larger again in 2028. Oracle’s five-year credit protection under 200 basis points and CoreWeave’s borrowing spreads under 500, out of the distressed-asset bands of the two borrowers’ rows (25 and 26). The local-restrictions row (31) voting strategic engine, restrictions falling while negotiated approvals rise. | Open. The Scorecard’s reading of October 2026 was rationed utility; Oracle’s protection stood at 288 basis points on 15 October 2026 and CoreWeave’s spreads past 1,000, both on the far side of the movement’s lines. |
| 31 August 2029 | [Chapter 14, the movement downward, toward distress](https://dryburgh.com/vol1/tests/2029-9) | By 31 December 2028, two consecutive years of falling energized additions (2027 and 2028), cancellations of financed blocks, downgrades below investment grade for a hyperscaler or a neocloud default, and shutdowns ordered rather than negotiated. | The frozen cohort’s (row 12) newly evidenced gigawatts smaller in 2027 than in 2026, and smaller again in 2028. The cancellations row (30) voting distressed asset, cancellations dominating additions. A rating below investment grade from S&P or Moody’s on one of the five hyperscalers whose cash cover row 21 reads, or a missed payment on the bonds or loans of a neocloud, a cloud that rents out accelerators. A regulator’s or a court’s order to stop a datacenter’s operation, read on the order itself. | Open. Oracle sits one notch above the investment-grade floor after S&P’s cut of 9 July 2026, so no hyperscaler is yet below it. |
| 31 August 2029 | [Chapter 14, the movement sideways, toward commodity](https://dryburgh.com/vol1/tests/2029-10) | By 31 December 2028, three or more suppliers crossing the capability thresholds while transformer lead times fall below a year and credit spreads converge. | A plurality of the capability clock’s voting rows for commodity infrastructure, whose limbs ask for three or more suppliers. The transformer lead time (row 15) under 52 weeks. Oracle’s credit protection (row 25) under 100 basis points. | Open. |
| 31 December 2029 | [Chapter 6, the transformer](https://dryburgh.com/vol1/tests/2029-11) | The emphasis is shown overstated only if three conditions hold together, Franklin Farms 1 and 2 in service by 31 December 2028, Waterford 5 in service by 31 December 2029, both within the May 2026 cost estimate, and the monitors showing no further rise by 31 December 2029. | Entergy’s filings and the U-37425 quarterly monitors | Open. Waterford 5 is dated for service in 2029. |

## 2030

5 tests**THE TESTS OF 2030 — each limb, the record it is read on, and its status in October 2026**

| Day | Claim | The test | Read on | Status |
| --- | --- | --- | --- | --- |
| 31 December 2030 | [Chapter 13, mass is money](https://dryburgh.com/vol1/tests/2030-1) | Three conditions together withdraw the claim. A gigawatt of usable orbital compute, one year’s deployment at the first rung of SpaceX’s own filed ramp, a gigawatt a year, with ten a year at the second, is in service by 31 December 2029, the year SpaceX’s finance chief named for “meaningful amounts of compute in space,” and runs through the twelve months of 2030 at 90 percent availability, for an all-in cost at or below $51 billion, which is what remains of Wood Mackenzie’s estimated $170 billion after the 70 percent reduction it says parity with the ground would require. A gigawatt in orbit by 31 December 2029 without the availability or the cost weakens the claim without withdrawing it. If the filings have not disclosed the availability or the all-in cost by 31 December 2030, those two conditions are unscored, never met and never failed, and the gigawatt in service is read alone. | SpaceX’s own filings | Open. |
| 31 December 2030 | [Chapter 13, orbit is a change of landlord](https://dryburgh.com/vol1/tests/2030-2) | Half the accelerators, memory and packages in a ten-gigawatt orbital fleet, one year’s deployment at the ramp’s second rung, made outside Taiwan under contracts good for ten gigawatts a year, would weaken the claim. | SpaceX’s own filings | Open. Of the three parts, the memory was already made mostly outside Taiwan, since more than four-fifths of high-bandwidth memory was Korean in the second quarter of 2026 on Counterpoint Research’s count. |
| 31 December 2030 | [Chapter 13, orbit is a change of landlord](https://dryburgh.com/vol1/tests/2030-3) | A route through the FCC, the FAA’s launch licenses and the ITU’s coordination that has licensed a hundred thousand satellites, a tenth of what SpaceX alone has applied for, within a year of a complete application, would weaken the claim. | The FCC’s docket | Open. The FCC’s decision on a million satellites had no date. |
| 31 December 2030 | [Chapter 13, orbit is a change of landlord](https://dryburgh.com/vol1/tests/2030-4) | Orbital compute sold from SpaceX’s constellation in ten national markets with no court’s stay or remand against it and no national regulator’s authorization required of it would weaken the claim. | The national regulators’ own records | Open. |
| 31 December 2030 | [Chapter 14, the campus’s financing](https://dryburgh.com/vol1/tests/2030-5) | The verdict is also weakened, toward a campus short of power, if S&P reports the project’s downside-case coverage below 1.10 times on a sustained basis, in two consecutive reports on Hyperion’s notes. | S&P’s reports on the notes | Open. The same reading as the row of 1 June 2029: a flat 1.12-times coverage, two hundredths above the trigger. |

The 75 tests as data: [CSV](https://dryburgh.com/vol1/data/tests.csv) · [JSON](https://dryburgh.com/vol1/data/tests.json) · [a calendar of the days they fall due](https://dryburgh.com/vol1/data/tests.ics)

## Questions

What are the dated tests?This section gathers in one place the dated tests the chapters print, in one table for each year. Each test is stated in full in its chapter, a chapter’s wording governs where a row’s differs.

What does a row of the tables show?A row gives the day, the claim the limb tests, the condition and what it does to the claim, the record it is read on, and the status, where the reading stood in October 2026.

What happens when a reading is not published by its day?A limb whose reading is not published by its day is unscored, neither met nor failed.

Are the Scorecard’s thresholds among the tests?No. The Dryburgh Compute Race Scorecard at the back holds a reading. Its thresholds are its own, printed with it and not repeated in the Dated Tests.

- [Campuses](https://dryburgh.com/vol1/ledger)
- [Scorecard](https://dryburgh.com/vol1/scorecard)
- [Data](https://dryburgh.com/vol1/data/releases)
- [Chapter by chapter](https://dryburgh.com/vol1/companion)
- [The record at a glance](https://dryburgh.com/vol1/observatory)
- [The machine](https://dryburgh.com/vol1/machine)
- [One campus, four futures](https://dryburgh.com/vol1/hyperion)
- [Glossary](https://dryburgh.com/vol1/glossary)
- [People](https://dryburgh.com/vol1/people)
- [Sources](https://dryburgh.com/vol1/sources)
- [Notes](https://dryburgh.com/vol1/notes)
- [Where the book was wrong](https://dryburgh.com/vol1/corrections)
- [Where its sources were wrong](https://dryburgh.com/corrections?book=vol1)
- [The Paper](https://dryburgh.com/vol1/paper)
- [Consent](https://dryburgh.com/vol1/consent)
- [Who builds and pays](https://dryburgh.com/vol1/parties)

*The AI Compute Race: Five Races, One Machine* is a book by Lee S. Dryburgh, the first volume of THE COMPUTE ECONOMY, published by Dryburgh, an imprint of Contestra OÜ, Tallinn, Estonia; its record closes on 8 November 2026.

[About the book](https://dryburgh.com/vol1)[Tell me when it is out](https://dryburgh.com/vol1#notify)