Home / Articles / A supply chain of vaults / Versions / v1.1.0
A supply chain of vaults: what changed in v1.1.0
From v1.0.0 (2026-09-27, e783118b3) to v1.1.0 (2026-09-27, 4030f39f7), paragraph by paragraph.
3 paragraphs added, 1 removed, 2 changed in place, 79 unchanged. About 563 words added and 205 removed. Insertions are marked like this, deletions like this; unchanged runs are folded to one line; figures appear as their file names.
1 unchanged paragraph
Summary: The conversation about food security and the cost of living keeps asking for new ideas and keeps producing the same ones, and the loudest thing said about AI in that conversation is that it is dangerous. This article argues the opposite case, with the evidence it could find. The food chain from a field to a shelf is a series of hops that each keep their own records, mostly in spreadsheets, and share as little as they can; the one party with real systems is the big buyer, and once it holds a large share of a farm's output it names the price, which is the mechanism Giblin and Doctorow call a chokepoint. All of that is logistics, and logistics is what generative AI, used the way this site uses it, is good at: capture everything, structure it, and generate the small, custom tool each piece of the chain needs, then run production without a model in the line. A supply chain of encrypted vaults, one per party, joined by append lanes and a typed graph, is described piece by piece, with what exists today and what is proposed kept apart. The hypothesis that this lowers the price of goods is set against the evidence: two thirds of supply chains on spreadsheets, 13% of food lost before retail, and the gains early adopters of AI planning report. It then takes on two dogmas, that falling prices are always bad, which the BIS's own history of deflations does not support, and that sharing is giving things away, when the uncounted cost is the cost of not sharing. It closes with the second memo's case for openness: open source and Creative Commons for supply chain workflows, open-weight models forthat everyrun industryinside a company's own environment and language,can be built on, the under-reported advantage of the economies already using them, and sharing the journey rather than the curated success story.
12 unchanged paragraphs, under In short
• Openness is the other half. Open source and Creative Commons for the workflows,workflows; open-weight models forthat everyrun industryinside your own environment and language,that you can build on, which is the under-reported advantage the economies using them already have; and sharing the journey rather than the curated success story.
38 unchanged paragraphs, under What the food pound says, The chokepoint, All of it is logistics…
The same argument applies to the models. Companies operating on open-weight models, and today that largely means Chinese ones, all have the equivalent of a frontier model, and the constraints they were placed under, especially on compute, produced efficiency. DeepSeek's technical report says its V3 model needed 2.788 million GPU hours on export-restricted H800 chips for its full training, a run that, as CSIS notes, the company priced at about USD 5.6 million; and by one analyst's count, Chinese open-weight models were about 61% of the tokens served through OpenRouter by May 2026. Whatever the exact figures, the consequence in the West is that most companies pay for AI access when they could have the option not to. And the same openness that gave everybody a near-frontier model makes it possible to have models per industry, per language and per culture: Switzerland's Apertus, released in September 2025 with its weights, data and method all open, covers more than a thousand languages with 40% of its training data in languages other than English. A model that speaks a grower's language and knows a sector's vocabulary is not a luxury for the supply chain. It is the thing that makes the brief in the first figure possible.
The same argument applies to the models, and here is a point that is usually under-reported. China, and the other economies that build on open-weight models, have a large economic advantage that has nothing to do with benchmark scores: their companies can run a near-frontier model inside their own environment and innovate on top of it. The model is one small part of a working solution, a very important part, but a small one; the rest is the workflow, the data, the tooling and the people. When the model is open, all of that can be built and owned locally. The US-China Economic and Security Review Commission's Two Loops paper (March 2026) describes exactly this: most Chinese labs publish weights and charge far less, which "has resulted in the acceleration of global uptake of Chinese AI and created a feedback loop where widespread adoption drives iteration, then further adoption", with Alibaba's Qwen models alone having over 100,000 derivatives on Hugging Face. The paper's own conclusion is that this open ecosystem "enables China to innovate close to the frontier despite significant compute constraints" and that it lets AI be deployed cheaply "across factories, logistics networks, and robotics". That is the supply chain, and it is being done.
I know a good number of startups that have built their products, and their own custom models, on top of open weights, most of them Chinese. Two objections come up every time and both have plain answers. The first is safety: these models run inside the company's own environment, often air-gapped, so there is no call home and no data leaving unless the company sends it, and any attempt to do either is caught by the same cyber security controls that govern every other piece of software the company runs. The second is bias: the harness and the workflow around the model manage it, as they must for every model, because the closed frontier models carry bias too, only less visibly. DeepSeek's technical report says its V3 model needed 2.788 million GPU hours on export-restricted H800 chips for its full training, a run that, as CSIS notes, the company priced at about USD 5.6 million; by one analyst's count, Chinese open-weight models were about 61% of the tokens served through OpenRouter by May 2026. Whatever the exact figures, the consequence in the West, where the conversation is about closed models and the fear around them, is that most companies pay for AI access, per call, to a provider they do not control, when they could have the option not to. The same openness makes it possible to have models per industry, per language and per culture: Switzerland's Apertus, released in September 2025 with its weights, data and method all open, covers more than a thousand languages with 40% of its training data in languages other than English. A model that speaks a grower's language and knows a sector's vocabulary is not a luxury for the supply chain. It is the thing that makes the brief in the first figure possible.
21 unchanged paragraphs, under What exists today, and what does not, Sources
7 unchanged paragraphs