# CBRS — article plan

Last updated: 2026-09-06
MSU: baseline not yet written · U: —

This file is the collaboration surface for humans and agents. Edit the
structure and research bullets. The ticker hub (`index.html`) shows them.
It is not a finished client report.

## Structure

1. Baseline — what M is pricing (EV, implied growth / FCF vs guidance)
2. Baseline — what S is saying (majority BB; material non-consensus)
3. Speed and cost vs NVIDIA — from filed Kimi drop (keep metric, generation, measurer)
4. MTP factor — model/context Pareto frontiers and the conditional GPU-vs-WSE moat impact
5. U on top — unique / special information if found; else label aligned
6. MSU scenario + what would change our mind

## Research to do

- [ ] M: current price / EV and what growth–margin path it implies
- [ ] S: file latest bulge-bracket notes (input, not U)
- [ ] Inventory numbers already in `Kimi-Agent_2026-08-29/` so we do not silent-swap baselines
- [ ] Hunt unique U (capacity, OpenAI/G42, wafer-scale vs GB300) not already in that drop
- [ ] If nothing unique: ship M+S baseline and label U aligned
- [x] Reframe MTP for investors: measured-vs-open setup matrix, S90/S75/S50 decision rule, and GPU-vs-WSE falsifier ([factor note](notes/2026-09-02-mtp-gpu-wse-frontier/index.html))
- [x] Add a second, Kimi-authored MTP narrative centered on per-user speed at matched aggregate throughput, internal batch vs active concurrency, and the unmeasured GPU–WSE frontier ([Kimi V2 factor note](notes/2026-09-02-mtp-gpu-wse-frontier-kimi-v2/index.html))
- [x] MTP Grok1 — first-principles note: no-MTP gap, Next-3 / Deep-16 nets, honest S90, cost knife-edge, 2M–10M ([MTP Grok1](notes/2026-09-03-mtp-gpu-wse-frontier-grok-1/index.html))
- [x] MTP Grok2 — independent parallel note: silicon-fair vs product-fair, Next-16 15.7 vs 14.6, model-specific cost ([MTP Grok2](notes/2026-09-03-mtp-gpu-wse-frontier-grok-2/index.html))

## Astra finishing plan — 6 September 2026

- Independent review and provisional [finishing plan](notes/2026-09-06-mtp-astra-review/index.html): keep the operator question, correct the relative GPU/WSE comparison, and connect latency to customer value and serving economics.
- Owner scope: high-level investor take; roughly 1,200–1,700 words, a few consequential numbers, compact supporting evidence. Compare GB300 NVL72 with CS-3; close with what CS-4 changes and what persists.
- Fable 5.1's original feedback is preserved in `Fable-5.1_2026-09-06/`. The [combined plan](notes/2026-09-06-mtp-astra-finish-work/combined-plan.md) adopts its useful corrections and tests its timing, capacity and economic conclusions.
- Separate article: [MTP - Astra ver.](notes/2026-09-06-mtp-astra/index.html), with [decision closure and Fable dispositions](notes/2026-09-06-mtp-astra-finish-work/synthesis-closure.md). Supporting calculations and review records remain alongside the note.

## Notes for agents

- M/S baseline first, then U. Do not label Street as unique U.
- One mixed client report, not three memos.
- Filing-only drops do not replace this plan.