01 / What 25–35% actually meant.
In Chapter 15, 25–35% is the author’s subjective estimate of NVIDIA successfully navigating the specified dilemma, not a probability of its demise. The earlier satirical site description reversed the meaning. This companion site no longer repeats that reversal.
The figure is not a measured failure rate, and its complement is not a bankruptcy probability. The original definition also bundles success with adoption of the proposed autonomous-unit structure. Actual success must remain recognisable when a company finds a different response.
Original argument · Chapter 15 →
02 / A response that preceded publication.
The book’s account of NVIDIA rationally ignoring Groq omitted material context. Groq announced a non-exclusive inference-technology licensing agreement with NVIDIA on 24 December 2025. Its founder, president and other team members would join NVIDIA; Groq would remain independent.
That announcement predates the April edition. It is a correction to the account, not evidence that the company only responded after the book. It does not, by itself, settle the wider question of inference competition.
Primary source · Groq, 24 December 2025 ↗
03 / A reversed sentence in Chapter 15.
Under “On AI factory complexity”, the sentence assigning a five-year arrival of modular substitutes to the bull case and a fifteen-year arrival to the bear case reverses the surrounding argument. Within that argument, faster arrival favours the challenge to the integration premium; slower arrival favours its durability.
This corrects the direction of the sentence, not the date of a technological transition. The original sentence remains visible in the reading text.
04 / Organisational diagrams are proposals.
The five directions reflect the author’s work and interests. Proposed units, budgets, locations and deadlines are concrete scenarios for discussion, not demonstrated necessary conditions for corporate survival.
Statements about what every meeting assumes, which projects always win funding, or which talent NVIDIA cannot attract are external interpretations, not access to internal decisions. An initiative is not “innovation theatre” simply because it differs from the author’s preferred structure.
Chapter 15 and the Coda themselves admit that the founder and organisational culture complicate a deterministic reading. That resistance belongs in the interpretation of the whole book.
05 / Keep the distinctions visible.
Framework adoption is not automatically hardware displacement. Portability depends on the workload and implementation. A new interface may still be used on the same underlying platform.
Gross margin is not net profit or the margin of every product. Periods, corporate and segment figures, actuals and forecasts should not be interchanged. The new August figure cited in the afterword is explicitly a company-wide quarterly GAAP gross margin.
New demand is not automatically a lost sale elsewhere. A local experiment can replace an existing purchase, enable previously unaffordable work, or lead to later demand. The book’s one-for-one cannibalisation language is too strong without a specified scenario.
A model token is not a certified economic outcome. The Token Standard is a monetary thought experiment, not proof that tokens are interchangeable money, correct thoughts or legally authorised actions.
06 / Ross King’s influence is not a hardware claim.
The book’s Coda places automated science and biological computing in a long-horizon discussion. Those research programmes should not be collapsed into a fixed timetable for silicon’s replacement.
King’s influence on AuthorityClaw came through conversations about formal logical approaches to AI. The experiment does not use DNA computing. Its later first-order/Z3 checks are separately documented from the recorded competition demonstration.
07 / The experiment and the award.
AuthorityClaw’s recorded run processed fictional invoices and simulated payments. It used a persistent Python harness and a local model service. An OpenClaw-compatible skill is supplied; the recorded run does not claim NemoClaw/OpenShell runtime integration.
The post-challenge formal model produced counterexamples that were replayed against the implementation and retained as regression tests after fixes. The model has stated assumptions and bounds: this is not an automatic proof of all Python code or a real institution’s legal policy.
First Place in NVIDIA’s Claw Agent Challenge: London is recognition of AuthorityClaw, not endorsement of this book, regulatory approval, or proof that every implementation detail is correct.
Project record ↗ · Formal results and limits ↗
08 / The tests are still dated in the future.
Chapter 15’s original falsification criteria are largely dated Q4 2027. October 2026 developments can update an assessment; they are not the final outcome of every scheduled test.
Original thresholds should stay visible. A revised test needs its own date and operational definition rather than replacing the old one retrospectively.