Section C · DS inside a neocloud

Churn & Retention

Predicting churn in GPU-rental customers, intervening before it happens, and winning back churned accounts. Distinct dynamics for self-serve vs enterprise customers; DS investment matters at both ends.

The shape of churn

Churn at a neocloud takes several forms:

Self-serve drop-off. A user signs up, uses some compute, stops. Like SaaS freemium churn.
Account graduation. A small customer grows large and migrates to a bigger neocloud or hyperscaler that can handle their scale.
Enterprise non-renewal. A reserved-capacity customer doesn't renew at term end.
Workload migration. Customer keeps the account but moves spend to a competitor.
Vertical exit. Customer's AI initiative pauses or ends entirely.

Each requires different DS treatment.

Churn signals

Predictive signals for impending churn:

Usage trend down (week-over-week, month-over-month).
Decrease in unique jobs launched.
Reliability incidents (any pattern of customer-affecting failures).
Support ticket volume changes.
Reservation renewal approaching with no progress signals.
Payment-method changes or billing issues.
Login frequency decline.
Competitor sign-up signals (where observable).

The DE-side telemetry (covered in DE guide) provides most of these signals; DS transforms them into prediction.

Prediction

Standard churn-prediction approaches apply:

Tree-based models (XGBoost, LightGBM) on feature engineering of the above signals.
Logistic regression for interpretable baselines.
Survival analysis for "when will they churn" rather than "will they churn."
Sequence models if you have enough sequential behavior data.

For enterprise churn, the data is sparser — fewer customers, less history. Bayesian methods or expert-elicited priors complement the model.

Intervention

Predictions are useful only if they trigger action:

Account manager outreach. For enterprise, an AE call to high-risk accounts.
Targeted discounts or credits. Retention promotions to at-risk self-serve accounts.
Product improvements. Address common churn causes systemically.
Customer success interventions. Onboarding assists for accounts not getting value.

DS measures intervention effectiveness with hold-out groups when feasible.

Reactivation

Churned customers are still addressable. Reactivation programs:

Win-back campaigns with special offers.
Product changes that address their original reason for leaving.
Relationship reset by sales (for enterprise).

DS analyzes which reactivation programs work; the answer informs ongoing investment.

Enterprise-specific dynamics

Enterprise churn is qualitatively different:

Lead time is long. Decisions to switch take months.
Stakeholders multiply. Multiple buyer-side personas with different concerns.
Competitive forces. Direct competitor displacement requires understanding their offer.
Renewal as the moment of truth. Multi-year contracts come due; the renewal decision is consequential.

DS work for enterprise churn emphasizes account-level deep-dives, not population-level models. Each churn risk is its own narrative.

✓ Check yourself

Answer from memory before opening each one.

Why does one churn model not serve both a self-serve drop-off and an enterprise non-renewal?

Show answer
Self-serve churn has many accounts and short histories, so a population model learns well. Enterprise churn has few accounts, long lead times, many buyers and a renewal date that decides it, so there is too little data and each risk is its own story. Use account-level review plus priors for enterprise. See enterprise-specific dynamics.
You want to know when a customer will leave, not only whether. Why is a classifier a weak fit, and what do you use instead?

Show answer
A classifier answers 'will churn within N days' and treats customers still active as non-churners, which discards the fact that they have not left yet. Survival analysis models time to churn and handles those censored customers properly. See prediction.
The model flags the top 10% at risk, you give them credits, and their churn drops. Did the credits work?

Show answer
You cannot tell. Some of those accounts would have stayed anyway, and without a comparison you do not know the lift. Keep a hold-out of flagged accounts that get no credit and compare. See intervention.
A customer's login frequency drops, but their job count is steady because they launch from CI. What does that say about your signal set?

Show answer
Logins are a weak proxy for use when automation does the work. Prefer signals that track the workload, such as unique jobs launched and usage trend, and read logins alongside them. See churn signals.
A small customer grows large and moves to a hyperscaler. Is that the same problem as a self-serve user who stops?

Show answer
No. The account left because it outgrew you, so a discount does not fix it. The response is product, such as capacity and a reserved tier, and measurement should separate graduation from drop-off. See the shape of churn.

Takeaway

Churn takes five forms: self-serve drop-off, account graduation, enterprise non-renewal, workload migration and vertical exit, and each needs its own treatment.
Usage trend, jobs launched, reliability incidents, support volume and approaching renewals are the main signals, and most come from data-engineering telemetry.
Use logistic regression as an interpretable baseline, tree models for accuracy and survival analysis when the question is when.
A prediction is useful only if it triggers an action, and its effect should be measured against a hold-out group.
Enterprise churn is sparse and slow, so use account-level deep dives with priors rather than a population model.
Reactivation programs are worth funding only when a measured comparison shows they beat natural return.

Retention feeds demand, and demand forecasts drive the largest spend a neocloud makes. → Capacity Forecasting