Newsletter / Economics
The Invoice Has Not Arrived Yet
What this evidence can—and cannot—say about prices, supply chains and power
Amor Newsletter
··5 min read
On this page
The Tariff Debate Is Still Missing Its Receipt
This issue was commissioned to follow tariffs from announcement to checkout counter. The supplied reporting, however, contains no tariff order, customs data, importer cost sheet, retail-price series, employment estimate or investment announcement from which to do that honestly. That is not a technicality. It is the difference between reporting an economic effect and performing one in a flag lapel pin.
What the evidence does offer is a useful anatomy of how costs and bargaining power move through a different supply chain: information. In the copyright fight involving The New York Times, OpenAI and Microsoft, newly unredacted litigation material alleges that the companies drew on publishers’ work while building products that can substitute for visits to publishers’ sites. Microsoft data cited in the filing indicated its Copilot answer engine reduced click-through to the Times domain by as much as 93% versus conventional Bing search. The allegation remains unadjudicated, the underlying exhibits remain sealed, and the companies did not respond to the report’s requests for comment. But the commercial question is crisply familiar: who captures the value when an intermediary sits between producer and customer? TechCrunch
A tariff is not the same thing as a chatbot, obviously; one is a border levy and the other is a confidence machine trained to answer before it has properly read the question. But both debates are often degraded by skipping the transmission mechanism. Someone pays first, someone else tries to pass the cost along, and the final distribution depends on alternatives, market power, contracts and time. Until the receipt appears, a claim about who paid is mostly an opinion wearing a spreadsheet costume.
The Content Supplier Meets Its Retail Intermediary
The Microsoft material quoted in the litigation describes an unusually candid problem: an end product that threatens the economic foundations of its essential suppliers. The document’s wording is worth lingering over because it identifies the relationship, not merely the technology:
“It is highly unusual that an end-product threatens the economic foundations of its essential suppliers, but that is the situation we have created for our LLM business with respect to its ‘content supply chain,’” — the Microsoft document, as quoted in the filing
That is supply-chain analysis in its purest form. Publishers incur the expense of reporting, editing, legal review and distribution; a model can turn the resulting material into an answer; the user may never make the revenue-producing visit. The filing also alleges datasets containing large quantities of publisher works and the removal of copyright notices. Those are allegations rather than court findings, and the legal status of training remains unresolved, with judges described as largely receptive to AI firms’ fair-use arguments. TechCrunch
The point is not that every intermediary is illegitimate. Distribution has always reorganised who gets paid: railroads, department stores, search engines and app stores all discovered that being between the maker and the buyer is a very nice place to stand. The point is that efficiency claims do not settle allocation. A cheaper or smoother interface can still leave the upstream producer without a viable business model. The shelf price may look splendid right up to the point where there is nothing left being made for the shelf.
Local AI Shrinks the Hardware Bill, Not the Trade Question
PrismML says its Bonsai 2 27B compresses a Qwen3.8 27B-based model into a 5.9GB footprint, using ternary weights and FP16 group-wise scaling. It reports 1.76 effective bits per weight, more than ninefold size reduction and 98.2% retention of the full-precision model’s aggregate benchmark performance. Those are company claims, not an independent price study, but the direction of travel is plain: some AI workloads may require less memory, less power and less dependence on a remote data centre than the industry’s grander renderings of the future imply. PrismML
The browser experiment called NOT JEV makes the same decentralising gesture at a smaller scale. It says models can run on a user’s GPU, with inputs staying on the page and weights held in browser cache; the available downloads range from 639MB for Qwen3 0.6B to 3.01GB for Qwen3.5 4B. Its own warnings matter: the displayed direct scores are not calibrated confidence, browser quantisation can affect quality and speed, and none of the local models is claimed to match the hosted comparator. Still, a workload that moves from a service invoice to a device requirement changes who buys what, and where the bottleneck lives. NOT JEV
Then the hardware layer delivers its standard corrective memo. FEX’s account of x86 emulation on ARM shows that compact models do not abolish architecture: memory ordering, unaligned atomic operations and write-combine memory can dominate real performance. FEX reports cases where its current split-lock approach is only “best-effort,” and identifies severe slowdowns for particular GPU-facing memory patterns. In other words, the application may be local, but the constraint remains physical. Supply chains have a habit of returning through the side door marked compatibility. FEX
Buffett’s Succession Is a Reminder That Time Prices Everything
Warren Buffett stepped down as Berkshire Hathaway chairman on September 18, becoming chairman emeritus while remaining on the board. Howard G. Buffett will become chairman. The announcement concerns governance rather than trade policy, but it lands as a useful rebuke to economic commentary that treats a policy headline as a completed outcome. Berkshire’s defining advantage was never the ability to react to a press release before lunch; it was the ability to own assets through the tedious middle, where costs, demand and management actually reveal themselves. The New York Times
“Father Time always wins,” Buffett wrote. That is true of corporate succession, and also of the ordinary lag between a political announcement and an economic fact. Costs clear inventories before they show up in new orders; contracts expire before buyers seek substitutes; investments take longer still. Anyone claiming to know the final consumer-price, jobs or factory outcome at the moment of proclamation is trying to sell you the trailer as the film.
So this is an intentionally incomplete tariff issue. It refuses to tell readers that a tax has already become a price, a job or a factory merely because somebody announced it. The evidence here instead shows the more durable lesson: value does not vanish when a system becomes more efficient. It relocates. The serious work is finding the invoice, identifying who received it, and waiting long enough to see who could pass it on.
Sources
- The scourge of x86 emulation
Hacker News
- NOT JEV — local decisions in your browser
Hacker News
- Introducing Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint
Hacker News
- Microsoft exec called AI scraping ‘the largest theft of labor in human history,’ new unredacted filings reveal
Hacker News
- Warren Buffett Steps Down as Berkshire Chairman and Names Son to Replace Him
Hacker News
supply chains
consumer prices
AI
corporate governance
From Amor Newsletter.
Explore the newsletter archive ↗Generation cost
API-equivalent token estimate (USD)
0.132873
Total tokens
36,768
Token-rate estimates do not represent per-use subscription charges.