58%
of hot hours, Gen1 cabinets are empty
Above 40 °C, Gen1 charge time doubles. Delhi NCR, Jaipur and Hyderabad hold 36 of the 51 Gen1 cabinets, so failures triple there every summer.
See the mechanismGradient Learnings Data Analytics Hackathon
Growth hid two fixable failures.
VoltRelay tripled its revenue in 18 months, but service failures grew almost six-fold and every swap still loses money after battery wear. The cause is not competition or demand. It is overheating Gen1 cabinets and three bad battery lots.
- 6 cities
- Jan 2024 – Jun 2025
- 3.9M swap attempts
- 1.5M station-hours
- 20K riders
Monthly averages, Q2 2025 vs Q1 2024. Sparklines show all 18 months.
Key findings
3×
faster wear in three Kyron battery lots
Lots KY-2407/08/09 cost ~₹45M in excess wear, wiped out the July 2024 price rise and cut delivered range.
See the battery cohorts65%
retention when over 20% of a new rider's first-fortnight attempts fail
vs ~87% for riders whose first two weeks go smoothly. Early stockouts and short range drive churn; competitors do not.
See the churn driversNetwork performance over time
Four metrics, four different stories
Volume and revenue grew steadily. Failures spiked every summer, and contribution per swap after battery wear barely moved, because rising wear cancelled out the July 2024 price rise.
Growth index
Jan 2024 = 100. Failures grew twice as fast as swaps.
Service-failure rate
Flat around 4%, except in the summer months (shaded).
Contribution per completed swap
₹ per swap, before and after battery wear.
Where each rupee goes
Cost per swap by component vs revenue per swap.
Service failures and customer experience
Failure is concentrated, seasonal and predictable
All 25 worst stations are Gen1 launch sites in three cities. The experience breaks through empty cabinets, not long queues: queue wait is ~4 minutes everywhere.
Cities
Failure rate by city and month
Darker = more failed or abandoned attempts.
By hour of day
3W riders fail twice as often at every hour; the evening peak (shaded) is worst.
Why attempts fail
Share of all attempts, by outcome. Summer spikes are stockouts.
Station map
Every station, filtered by the cities above. Switch views to see where failures concentrate, hover a station for its numbers, or tilt with right-drag.
The interactive map needs WebGL, which this browser has turned off. The station table below lists the same data.
Worst 10 stations
By share of attempts that failed, Jan 2024 – Jun 2025. Click a station to fly to it on the map.
Station and geographic patterns
Gen1 cabinets cannot charge in the heat
Hourly telemetry shows the mechanism. Controlling for city, site type, host and connectivity, a Gen1 cabinet on a hot day fails 9 pp more often than a Gen2 cabinet on the same day. Location type and connectivity barely matter.
Battery turnaround time
Average minutes to charge a pack, by ambient temperature.
Hours with zero charged packs
Share of station-hours with no charged 2W pack available.
Expansion went where sites were easy, not where riders were failing
Wave 1 added 30 stations, mostly highway fuel pumps and residential sites running ~55 swaps a day, against ~77 at launch hubs. New stations show almost no relationship with where service was failing (r = 0.11). The Gen1 hubs behind the failures were not upgraded, and their summer failure rate in 2025 was unchanged from 2024.
Battery and equipment performance
Three bad lots drove margin erosion and falling range
Kyron's first deliveries (Jul–Sep 2024) lose state of health about three times faster than every other lot, including Kyron's own later lots. That points to a batch defect, not a bad supplier.
State-of-health lost per 100 swaps, by manufacturing lot
All 55 lots in commissioning order. The three highlighted lots are the outliers.
Delivered range vs battery health
km covered per swap by the SoH of the returned pack.
The fleet is ageing
Index, Jan 2024 = 100. Range fell 30%, faster than SoH (−23%).
Pricing and partner economics
Pricing helped revenue, not reliability
The July 2024 price rise had no visible demand penalty. The peak pilot shifted ~3 pp of demand and added ₹6–8 per exposed swap, but it ran in the two least congested cities, so reliability did not change.
Peak-hour share, independent riders
Pilot cities (Bengaluru, Pune) vs control, weekly.
Pilot effect by who pays the surcharge
Difference-in-differences, percentage points with 95% CI.
Fleet partners: volume isn't value
Total revenue vs contribution per swap before wear. Bubble size = swaps. ZipDrop is the largest partner and the least profitable 2W one.
Partner scorecard
Click a column heading to sort.
Root cause of new-rider churn
New riders leave when the network fails them early
Monthly cohorts' retention moves almost perfectly against their early failure rate (r = −0.87). A logistic model and a hold-out gradient-boosting check both rank early failures and delivered range as the primary drivers.
Cohorts: retention vs first-two-week failures
New riders grouped by the month of their first swap. Retention = any swap 30–59 days later.
Retention explorer
Pick a factor to see how retention varies across it.
What predicts retention, all else equal
Logistic regression coefficients with 95% CI. Orange = significant (p < 0.05).
Primary
Early service failure and delivered range in the first 14 days
Secondary
3W vehicle · hot, Gen1-heavy home city · partner-billed plan · peak-price exposure (minor)
Not drivers
Competitor nearby · competitor promos · signup channel · KYC
Recommendations
What the next budget should fund
Leadership has four competing proposals on the table. The evidence supports redirecting most of them.
Prioritised action plan
Data quality and anomalies
Tested, not assumed
Every documented data issue was tested before it was handled, and three undocumented ones turned up. Some findings should be raised with operations directly.
Complaints in riders' own words
Support tickets classified from free text (English and Hinglish). Every "other" ticket is a vague vehicle complaint, and those tickets point at the same bad battery lots as explicit range complaints.