Threadbaire/compute-watch

The Compute Watch

A monthly public record of what AI chips (GPUs) rent for, what lenders charge against them, and how long their owners say they last. Kept by Lida Liberopoulou at Threadbaire, independent research on AI economics and infrastructure.

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. This dataset records, once a month, whether the rent supports that assumption, together with the dated events and reasoning behind each reading.

The notes themselves are published at threadbaire.com/blog/posts/compute-watch.html. This repository holds the data and the structured context, not copies of the notes.

Current status

• Latest note: No. 0, The Baseline (6 October 2026). Read it.
• Test: inactive. The shortage is not yet easing (memory index at 100, the baseline).
• Premium (B200 ÷ H100): 2.12× on Silicon Data, 2.99× on Ornn.
• Next note: No. 1, 2 November 2026.

This block is replaced with each note. Earlier states are in observations.csv and in this repository's commit history.

The prediction, pre-registered

Once the chip shortage eases, the newest chip tracked by the planned futures (the B200) loses its rent premium over the older H100, well before the five or six years the owners' accounts assume.

The rules, as published in No. 0 on 6 October 2026:

• Shortage easing: the GDDR6 memory index is at least 15% below its 2 October 2026 baseline for two notes in a row (gddr6_rebased ≤ 85).
• Once easing, right: the premium is below 1.6× on Silicon Data for two notes in a row.
• Once easing, wrong: the premium stays above 1.9× on Silicon Data for two notes in a row.
• In between: too early to say.
• Until easing: B200 price moves prove nothing either way.
• If not eased by No. 6: the prediction stays open, and that is stated publicly.

The rules do not change. Known confounds (new workloads running on older chips; scarcity of powered data-centre space rather than memory) are recorded in the notes, not used to adjust the test.

Files

File	One row per	What it holds
observations.csv	note	The headline readings, test status, and what changed in the indexes themselves
ladder.csv	chip per note	Each chip generation's rent relative to the H100
events.csv	dated event	What happened in the market around each note, in plain words, with a source
lifespans.csv	company per filing	Server lifespans stated in owners' annual reports
loans.csv	disclosed deal	Terms of disclosed chip-backed borrowing
notes/no-XX.md	note	Structured reasoning: what changed, what it might mean, where sources disagree, test status

Column notes

• Premiums and ladder ratios are computed from index prices recorded on a fixed day (the first working day of the month from No. 1; 2 October 2026 for No. 0). Silicon Data is the headline index; Ornn is the cross-check.
• gddr6_rebased is Silicon Data's GDDR6 index set to 100 at the 2 October 2026 baseline. GDDR6 is graphics memory, not the HBM inside data-centre chips; it is used as a public reading of the wider memory shortage.
• index_content_changes records changes in what an index contains (a new chip, new providers, a method change), so they are not mistaken for market moves.
• causal_claim in events.csv is always none. Events are recorded because they might later show up in the numbers, not as explanations.

Terms

• A100, H100, H200, B200, B300: Nvidia AI chips (GPUs), from older to newer. B200 and B300 belong to the Blackwell generation. Vera Rubin is the generation after Blackwell.
• Premium: B200 rent divided by H100 rent. How much more the newest tracked chip earns than an older one.
• Specialist provider (neocloud): a company that mainly rents out AI chips, such as CoreWeave, Nebius or Lambda.
• Hyperscaler: the largest cloud companies: Amazon, Microsoft, Google.
• Shortage easing: the trigger that activates the test, measured by the GDDR6 memory index.
• Server lifespan: how many years an owner's accounts assume a server stays in use (its depreciation period).

Method

• Both index pages are recorded on the same day each month, and a screenshot is kept for every reading.
• Specialist-provider prices are used unless marked. The same H100 rents for about 2.6× more at the largest cloud companies.
• Lifespans and loan terms change only when a new filing or deal is disclosed; rows are added, never edited.
• Corrections are made in a new commit with a note in the relevant notes/ file.

What is not included, and why

Raw index prices are not republished here. They are quoted on the published notes under Silicon Data's Terms of Service. This dataset contains derived figures (ratios, rebased values), figures from public filings, and the author's own analysis.

Reader contributions about what people pay for compute are never included in raw form. Pooled ranges, where they exist, appear only in the published notes.

Sources

The indexes are Silicon Data GPU Rental Price Indices and the Ornn Compute Price Index. Every row in events.csv, lifespans.csv and loans.csv carries its own source link.

Citation

Liberopoulou, L. (2026). The Compute Watch [Dataset]. Threadbaire. https://huggingface.co/datasets/Threadbaire/compute-watch

Licensed CC BY-SA 4.0. Research, not investment advice.

Corrections, counter-arguments and questions: threadbaire.com/contact.html#compute-watch.

Related: the full articles, as raw markdown, are in Threadbaire/research-articles.