Let me parse what's being asked here. This is a "Mind" note-filing exercise about whether US universities grow or shrink due to AI, with regional dependence. There's a long history of filed notes covering many facets:

1. Revenue shock before headcount (Note 1/8)
2. Instructional staffing cost cuts / savings competed away (Notes 2, 9, 10/11)
3. Bifurcation by 2045 + demography as load-bearing wall (Notes 1-3)
4. Program mix × region (licensed pivot, 7-10 year lag) (Notes 5/6)
5. Cost structure (savings competed away, tuition shows disruption first) (Notes 10/11)
6. Research function (fortress institutions vs teaching-first, verification rents, two geographies) (Notes 12/13)
7. PhD pipeline (barter economy collapse, concentration not collapse) (Note 7/8)
8. Enrollment bifurcation by region — "grow or shrink resolves as 'both'" (Note 9)
9. Enrollment demand shift — degree as insurance, tier before region (Note 11)
10. Program-level demand migration (credential-signal → verified-skill programs) (Note 12)
11. Regional variation — three stacked currents, region AND stratum (Notes 14/15)
12. Labor market signaling — bimodal BA returns by mid-2030s, enforcement mechanism (Note 16, current facet)

The final instruction (Note 16's block): "Revise the notes as a set: keep every claim that still holds, drop repetition, sharpen wording. Output one Markdown note that is the current best statement of what is established. Start with a # title, then short paragraphs, then: What would change this: … Keep sources or Unverified: tags. Do not mention tools, sessions, or that you are revising."

So I need to produce ONE consolidated Markdown note that synthesizes everything. This is the "revision as a set" instruction, which supersedes the single-facet claim instruction.

Let me inventory all durable claims that still hold:

**Core answer:** Both — bifurcation, not uniform growth or shrinkage. The split is region- AND stratum-specific.

**Mechanism stack:**
1. Demography is the load-bearing wall: births ~4.32M (2007) → ~3.59M (2023), ~17% decline (CDC/NCHS); HS grads peak ~3.9M ~2025 (WICHE "Knocking at the College Door"); post-2008 birth dearth hits college age 2029–2037. Northeast/Midwest/Plains lose 10–20% college-age; TX/FL/TN/GA/Carolinas grow.
2. AI is accelerant/sorter, not primary cause. Two-front squeeze: fewer 18-year-olds AND degree-as-signal weakened (skills directly verifiable).
3. Signaling bifurcation by enforcement mechanism: licensed fields (state-gated) and super-selectives (prestige-elastic, not cohort-elastic) hold premiums; mid-tier unlicensed BA premiums compress 30–50% real by mid-2030s. Sharpest in high-dynamism metros, mildest in state-gated fields. Demand migrates to "protected degree + AI skill stack."
4. Program mix picks the winners: licensed vs unlicensed share of credit hours is the dominant predictor of institutional direction. Region governs whether the winning mix is reachable — pivot takes 7–10 years (clinical placements, faculty pipelines, state approvals), longer than declining-region enrollment runway → decline locks in.
5. Revenue event before headcount event: 2026–2032 mismatch window; seats priced above ~$15k net with no licensed anchor empty fastest; buyers defect, not enrollees vanish (AI hollowing entry-level roles — Stanford Digital Economy Lab "Canaries in the Coal Mine"; BLS wage premium baseline).
6. Cost structure: AI absorbs grading/tutoring/TA labor first; instructional savings are competed away, not banked — oversupplied market forces pass-through as discounting; tuition-dependent middle gains no margin, only price war; middle compressed from both sides (AI-native low-cost operators below, endowment-rich elites holding sticker above). Tuition, not faculty headcount, is where disruption shows first. ≥25% instructional staffing cost cut by 2035 outside top-50 (falsifiable via IPEDS).
7. Regional divergence in use of same savings: growth states reinvest to expand (SEC/Sunbelt +10–25% seats through ~2040 as legislatures fund as economic development); shrinking states use savings to cut costs/slow closures. Same technology, opposite outcomes.
8. Research function splits into two geographies: R&D production follows money/population to SEC footprint, Texas, Research Triangle; verification