X gave every user 75 custom topic feeds powered by Grok. It's framed as personalization. The real function is supervised training data acquisition at a scale Meta spent $72 billion trying to replicate.
The shift changes the training signal flowing into Grok every second — from a flat unlabeled graph to a live 75-dimensional labeled vector.
The same labeled interest-graph data. Two very different paths to acquiring it. The Meta capex equivalence claim is the sharpest argument in the analysis.
Elon Musk is executing the 2016 Zuckerberg playbook — with one critical upgrade.
| Metric | Value | What It Means |
|---|---|---|
| Daily Active Users | 250M | Scale of the training data engine — more DAU = more labeled signals per day |
| Topics per User | 75 | Each user contributes to 75 independent labeled topic buckets simultaneously |
| Granularity Multiplier | 75× | Per-user signal is 75× more granular overnight vs. the prior flat engagement graph |
| Daily Labeled Data Points | 18.75B | 250M × 75 — estimated voluntarily labeled topic-engagement signals per day |
| Meta Capex Equivalent | $72B | What Meta spent in 2025 to approximate via inference what X acquires via voluntary labeling |
| Marginal Data Cost for X | ≈ $0 | Each labeled signal carries near-zero incremental cost for X |
| Premium User Share | ~0.5% | Of user base — but ~20% of subscription revenue and the highest-engagement training data generators |
The shift from one global For You ranker to 75 per-topic Grok algorithms restructures who wins and loses on X.
X is not the first platform to make this structural move. The Instagram 2022 topic-first shift is the playbook — and the warning.
Seven steps from the HowTo guide in the RDF, informed by the Instagram 2022 precedent.
Each pinned topic timeline generates labeled engagement signals, giving Grok a 75-dimensional interest vector per user — the equivalent of what Meta spent $72B in capex to infer from raw unlabeled behavior.
Before the update Grok had a flat engagement graph. After, it receives a labeled 75-dimensional interest vector per user, refreshed continuously, with users voluntarily tagging which topic each engagement belongs to.
Meta spent $72B on capex in 2025 to build compute to infer interest graphs from raw behavior. X gets equivalent labeled data by having users hand-label it themselves in exchange for a more personalized feed — per the capex equivalence claim.
Generalist accounts that previously won on broad appeal lose reach. Specialists who dominate a single topic bucket gain preferential Grok ranking in that timeline. Every creator now faces 75 algorithms to beat instead of one.
Instagram shifted to a topic-first feed in 2022. Lifestyle generalists collapsed while niche operators compounded reach. The Instagram 2022 precedent is the structural dynamic the analysis predicts will unfold on X.
Premium users are ~0.5% of the base but ~20% of subscription revenue and the highest-engagement cohort. X charges its most valuable training data generators a subscription fee — a business model few companies can architect.
Own the feed. Own the ranking layer. Own the data that trains the ranking layer. The 2016 Zuckerberg playbook is Meta's original strategy. Musk executes it with one upgrade: charging users for the privilege of generating the training data.
It is described as "the cleanest supervised learning setup a foundation model company could design" — users self-select topic buckets, making every tap, dwell, and reply a pre-labeled training signal.
250 million daily active users × 75 pinned topics makes the per-user training signal 75× more granular overnight, generating an estimated 18.75 billion labeled topic-engagement data points per day.
Specialize deeply in one or two topic verticals. Pin the most relevant topic timelines. Optimize engagement signals within those topics. Treat each of the 75 rankers as an independent feedback loop — the same strategy that won after the Instagram 2022 shift.