Datasets:

Threadbaire
/

compute-watch

@@ -1,82 +0,0 @@
----
-note: 0
-title: "The Compute Watch - No. 0: The Baseline"
-note_date: 2026-10-06
-silicondata_recorded_on: 2026-10-02
-ornn_recorded_on: 2026-09-30
-canonical_url: https://threadbaire.com/blog/posts/compute-watch-00.html
-author: Lida Liberopoulou
-license: CC BY-SA 4.0
-test_status: inactive (shortage not easing)
----
-# No. 0: The Baseline
-## What this is
-The Compute Watch is a monthly public record of what AI chips (GPUs) rent for, what lenders charge against them, and how long their owners say they last. Much of the AI buildout is financed with debt, and chip rent is what repays it. Owners' accounts assume a server stays in use for five or six years, while a better chip arrives every year or two. The watch checks, once a month, whether the rent supports that assumption.
-This file is the structured reasoning behind No. 0, the baseline. It sets the starting values; every later note is compared against these readings.
-- The published note: [No. 0: The Baseline](https://threadbaire.com/blog/posts/compute-watch-00.html)
-- What the watch is, and the prediction it checks: [The Compute Watch](https://threadbaire.com/blog/posts/compute-watch.html)
-- The data for this note: `observations.csv`, `ladder.csv`, `events.csv`, `lifespans.csv` and `loans.csv` in this dataset, rows marked note `0`
-- The dataset: [Threadbaire/compute-watch](https://huggingface.co/datasets/Threadbaire/compute-watch)
-- Kept by Lida Liberopoulou at [Threadbaire](https://threadbaire.com), independent research on AI economics and infrastructure. Research, not investment advice.
-## Readings
-- Premium (how much more the newest tracked chip, the B200, rents for than the older H100): 2.12× (Silicon Data), 2.99× (Ornn).
-- The same H100 rents for 2.60× more at the largest cloud companies than at specialist providers.
-- Memory index (GDDR6): 100 by definition.
-- Owners' server lifespans: 5 to 6 years for most large owners (see `lifespans.csv`).
-- Loan terms: one disclosed chip-backed loan (Nscale, $1.4 billion, secured on GPUs), at 5 points over the benchmark rate (see `loans.csv`).
-## What changed since the last note
-Not applicable: first note.
-## What it might mean
-The readings are consistent with a shortage: rents are high and some list prices were still rising (Nebius, from 1 October). Lenders asking for more protection does not contradict this. They are pricing what happens after the shortage, which is what this watch tracks.
-In the week before the baseline, the Silicon Data premium narrowed slightly because the H100 rose about 5% while the B200 was flat. That is consistent with tight supply for older chips, but one week of an index cannot show why.
-## Where the sources disagree
-Silicon Data and Ornn disagree on the level of the premium (2.1× against 3.0×) and on how much it moves. One likely reason is method: Ornn builds its index only from completed deals, while Silicon Data adjusts and averages across providers. Over the month before the baseline, Ornn had the H100 down about 18% and the B200 up about 18%. Both indexes are tracked; Silicon Data is the headline because it is steadier, and it is the index CME plans to use.
-## Test status
-The prediction: once the chip shortage eases, the B200 loses its rent premium over the H100 well before the five or six years the owners’ accounts assume. Right if the premium is below 1.6× on Silicon Data for two notes in a row; wrong if it stays above 1.9× for two notes in a row.
-Status: inactive. The shortage-easing trigger(memory index at or below 85 for two notes in a row) has not been met. Until it is, movements in the B200 price prove nothing either way.
-## Known confounds, recorded at baseline
-- New workloads on older chips: older-chip rent can rise because of new demand, not only because the shortage eases. A narrowing premium could then look like a pass for the wrong reason.
-- Power, not memory: if scarcity sits with powered data-centre space, the B200 premium partly prices access to power, and could compress because power loosens rather than memory.
-Neither changes the test.
-## What to watch in No. 1
-- Whether Nebius's 1 October price rises appear in the October indexes.
-- When Vera Rubin, the generation after Blackwell, enters the indexes (a content change, not a market move).
-- Whether the two indexes move closer together or further apart.
-- The outcome of the CFTC comment period (closes 20 October), and whether CME or ICE launches.
-## Counter-argument and reader question
-None yet. From No. 1, each note records the strongest counter-argument received, with an answer, and one reader question.
-## Contribute
-Send what you pay for compute, a correction, a counter-argument or a question: [threadbaire.com/contact.html#compute-watch](https://threadbaire.com/contact.html#compute-watch). You don't need to give your name. Your own numbers are never published or put into AI tools, and pooled ranges appear only once at least five people have reported on a chip. Arguments and questions may be discussed with AI tools while drafting an answer, without your name or identifying details, unless you ask for them to stay out. [How contributions work](https://threadbaire.com/blog/posts/compute-watch.html#how-to-contribute).