Updated README with compute-watch-01

README.md

@@ -31,7 +31,7 @@ pretty_name: Threadbaire Research Articles
 # Threadbaire Research Articles
 Independent research on AI economics and infrastructure: where the value in the AI economy is created, who keeps it, and who ends up paying for it. By Lida Liberopoulou ([about](https://huggingface.co/datasets/Threadbaire/research-articles/blob/main/about-lida-liberopoulou.md) · [threadbaire.com](https://threadbaire.com)).
 
-Structural analysis of AI industry dynamics, software value collapse, and open infrastructure. Eighteen articles published between January and October 2026, available as raw markdown for analysis, citation, and AI-readable ingestion.
+Structural analysis of AI industry dynamics, software value collapse, and open infrastructure. Nineteen articles published between January and November 2026, available as raw markdown for analysis, citation, and AI-readable ingestion.
 ## About
 This dataset contains the complete text of the Threadbaire thesis and blog — a body of independent research arguing that AI has already collapsed traditional software value capture mechanisms, and mapping the capture infrastructure forming in its wake.
 All articles are published under CC BY-SA 4.0. Each piece is sourced, dated, and designed to be machine-readable from the ground up — raw markdown endpoints, citation blocks, and structured metadata are part of the publishing philosophy.
@@ -56,6 +56,7 @@ All articles are published under CC BY-SA 4.0. Each piece is sourced, dated, and
 | `follow-the-money-both-ways.md` | Follow the Money Both Ways - Four Predictions for Anthropic's S-1 | 2026-10-23 | ~1,900 |
 | `income-or-promises.md` | Income or Promises - Five Predictions for Microsoft, Alphabet and Amazon | 2026-10-27 | ~4,400 |
 | `income-or-promises-where-the-bill-went.md` | Income or Promises - Where the Bill Went | 2026-10-31 |~4.300 |
+| `compute-watch-01.md` | The Compute Watch - No. 1: Rents Up, Memory Barely Moved | 2026-11-02 | ~3,000 |
 ## Structured Metadata
 `articles.json` contains per-article metadata: title, date, word count, arc position, core claim, evidence sources, tags, and links to both the dataset files and canonical URLs at threadbaire.com.
 ```python
@@ -63,7 +64,7 @@ from datasets import load_dataset
 ds = load_dataset("Threadbaire/research-articles")
 ```
 ## Arc
-The thesis is the root document. Each blog post extends it by moving one layer deeper through the stack: the human cost of the collapse, then a triptych on the agent web — authentication capture, commerce capture, and post-transaction memory capture — followed by the model commodity layer and the software category sorting mechanism. The latest series, The Silicon Industrialists, moves below the software layer entirely. Part 1 diagnoses the industrial economics — arguing that AI's cost structure is industrial, not software, and that the collision with dissolving revenue models pushes the entire economy toward regression. Part 2 traces the infrastructure trap through railroads, electricity, and telephony, then follows the escape through three institutional openings that made the software economy possible, and asks what held the ladder in place. "The Wrong Game" applies the capital allocation stress thesis to a single company — observing that Microsoft voluntarily disclosed an internal margin metric for Xbox after twenty-five years of absorbing worse losses quietly, while AI returns remain visible only through aggregated segment-level signals. "A Tale of Two Earnings" is the first real-time predictive experiment — six falsifiable predictions filed before Microsoft, Apple, and Amazon report, testing what the circular financing mechanism will make visible in the filings and what it will leave aggregated. "What the Filings Cannot Show" scores all six predictions against the actual filings — all held — and introduces three separations (investment/operation, investor/customer, revenue/capital) as the analytical framework for why the circular financing mechanism remains invisible in standard disclosure, and proposes independently financed versus ecosystem-financed demand as a new analytical category. Part 3 of The Silicon Industrialists traces how the three separations that made the software economy possible — separable hardware, ownable software, and open distribution — were gradually reversed through cloud recentralisation, subscription ownership, externalised capability, gated distribution, joined identity, and the absence of successor institutional protections, establishing that AI entered an already-reconstructed bind rather than creating one. The Compute Watch, launched in October 2026, is a standing instrument rather than an argument: a monthly public record of what AI chips rent for, what lenders charge against them, and how long their owners say they last, with one pre-registered prediction checked against the numbers each month. Its figures and structured context are kept in a separate dataset, [Threadbaire/compute-watch](https://huggingface.co/datasets/Threadbaire/compute-watch). "Follow the Money Both Ways" applies the same pre-registration method to Anthropic's S-1, filed before the document was public: four predictions about whether the filing will let a reader follow the money between Anthropic, the companies that fund it and the companies it pays. It is the first piece written in plain language for readers outside finance, and its predictions are scored in the Compute Watch. "Income or Promises" returns the method to the hyperscalers in the same plain language: five predictions and one control, filed before Microsoft, Alphabet and Amazon reported on 28 and 29 October. Together they test whether any improvement in their cash is new income, or a cost moved into later bills, investment gains and promised revenue from AI labs they fund. “Where the Bill Went” scores them against the filings: four held and one could not be tested, with one pass reported alongside a research miss it discloses and a change to the method. It then follows the cost of the build into places the headline figures don’t show: guarantees of other companies’ data-centre rent, loans secured on servers, short leases kept off the balance sheet, and backlog from companies the sellers part-own.
+The thesis is the root document. Each blog post extends it by moving one layer deeper through the stack: the human cost of the collapse, then a triptych on the agent web — authentication capture, commerce capture, and post-transaction memory capture — followed by the model commodity layer and the software category sorting mechanism. The latest series, The Silicon Industrialists, moves below the software layer entirely. Part 1 diagnoses the industrial economics — arguing that AI's cost structure is industrial, not software, and that the collision with dissolving revenue models pushes the entire economy toward regression. Part 2 traces the infrastructure trap through railroads, electricity, and telephony, then follows the escape through three institutional openings that made the software economy possible, and asks what held the ladder in place. "The Wrong Game" applies the capital allocation stress thesis to a single company — observing that Microsoft voluntarily disclosed an internal margin metric for Xbox after twenty-five years of absorbing worse losses quietly, while AI returns remain visible only through aggregated segment-level signals. "A Tale of Two Earnings" is the first real-time predictive experiment — six falsifiable predictions filed before Microsoft, Apple, and Amazon report, testing what the circular financing mechanism will make visible in the filings and what it will leave aggregated. "What the Filings Cannot Show" scores all six predictions against the actual filings — all held — and introduces three separations (investment/operation, investor/customer, revenue/capital) as the analytical framework for why the circular financing mechanism remains invisible in standard disclosure, and proposes independently financed versus ecosystem-financed demand as a new analytical category. Part 3 of The Silicon Industrialists traces how the three separations that made the software economy possible — separable hardware, ownable software, and open distribution — were gradually reversed through cloud recentralisation, subscription ownership, externalised capability, gated distribution, joined identity, and the absence of successor institutional protections, establishing that AI entered an already-reconstructed bind rather than creating one. The Compute Watch, launched in October 2026, is a standing instrument rather than an argument: a monthly public record of what AI chips rent for, what lenders charge against them, and how long their owners say they last, with one pre-registered prediction checked against the numbers each month. Its figures and structured context are kept in a separate dataset, [Threadbaire/compute-watch](https://huggingface.co/datasets/Threadbaire/compute-watch). "Follow the Money Both Ways" applies the same pre-registration method to Anthropic's S-1, filed before the document was public: four predictions about whether the filing will let a reader follow the money between Anthropic, the companies that fund it and the companies it pays. It is the first piece written in plain language for readers outside finance, and its predictions were scored in Compute Watch No. 1: all four held. "Income or Promises" returns the method to the hyperscalers in the same plain language: five predictions and one control, filed before Microsoft, Alphabet and Amazon reported on 28 and 29 October. Together they test whether any improvement in their cash is new income, or a cost moved into later bills, investment gains and promised revenue from AI labs they fund. “Where the Bill Went” scores them against the filings: four held and one could not be tested, with one pass reported alongside a research miss it discloses and a change to the method. It then follows the cost of the build into places the headline figures don’t show: guarantees of other companies’ data-centre rent, loans secured on servers, short leases kept off the balance sheet, and backlog from companies the sellers part-own. Compute Watch No. 1, the first monthly reading against the baseline, records rents rising fastest on the newest chips while the memory price barely moved, and supply held back by credit, power and memory at once; the prediction's test has not yet started.
 ## Research Directions
 The Silicon Industrialists series identified a dual-force mechanism — the double bind between industrial-scale infrastructure costs and value dissolution — that produces the same economic sequence across every shared infrastructure technology examined (railroads, electricity, telephony, computing). The series also identified the institutional conditions (consent decrees, mandated unbundling, conditioned privatisation) that broke this sequence for the computing industry specifically.
 Current research is investigating whether this mechanism provides the missing causal explanation for Kondratieff long waves (~50-60 year economic cycles associated with successive infrastructure technologies). The existing long-wave literature (Schumpeter, Perez, Freeman & Louçã) documents the empirical pattern and maps the installation-deployment sequence but has not converged on a structural mechanism that explains why every infrastructure technology follows the same financial shape. The double bind — industrial costs that cannot be avoided combined with value dissolution that erodes the revenue base funding the buildout — may be that mechanism, with the institutional conditions explaining the transition from crisis to deployment.