This article is based on a LinkedIn post by Shisa AI CTO＆Co-founder Leonard Lin and is republished with permission.
The past few weeks have been unusually eventful for the geopolitical and economic debate around frontier LLMs. In particular: the Fable 5 export-control saga (still unfolding, not over), and the heated discussions on how extractive frontier labs may be toward their clients and users.
At Shisa.AI, we train open-source Japanese language models, build open evals, and care deeply about sovereignty as a practical concern: control over models, data, workflows, deployment, and the feedback loops that improve them. As a business, what we build reflects both our principles and our read of the landscape.
Those in some of the private groups I'm in know that I sometimes share some of the research/analysis that I'm pre-disposed to doing. These are often short-hand/one-off, not really polished enough for public consumption or publishing (weird to say on LinkedIn I know, but in the age of AI slop, it seems more important than ever to actually have your hand on the keyboard, at least when it comes to prose).
Of course, sometimes, there's something topical/interesting enough that's worth sharing more widely, and these days, with AI tooling, it is much easier to turn these into legible and sharable artifacts. So here's:
Frontier Labs, Enterprises, and the AI Value Chain:
Who learns what from whom when AI labs deploy into businesses (and who keeps the value)
This is just one take, by its nature incomplete and contingent, but I think as a motivated exploration on the AI value-capture question, it's worth sharing for those interested in the topic. (It's also doubles as a good mid-2026 snapshot of what AI-assisted analysis looks like atm.)
This document is written as a long-form brief: 12 sections, ~15K words, roughly 35 pages. It's intentionally more explicative than usual so it's hopefully legible to readers who aren't perma-plugged into the AI bubble.
Original LinkedIn Post