Who captures the value in India's UPI, when the busiest half of the network is priced at zero?
A reproducible research portfolio on Indian financial services: payments, bank margins, market structure and wealth.
Read the report · Point of view · Diligence memo · Sources and provenance · Data dictionary
India built the world's largest real-time payments network, and priced the busy half at zero.
In 2026 Q2, merchant payments were 63.9% of all transactions but only 23.0% of the rupees moved. Person-to-person transfers were the mirror image: 30.8% of transactions, 71.2% of value. The merchant leg is the only one a merchant discount rate could ever be charged on, and under India's zero-MDR regime it earns nothing: 50.7 million merchants, 487 transactions each per quarter, ₹2.07 lakh of GMV each, ₹0 of payment revenue.
The market that carries it is concentrated at the same time as it is unmonetised. This repository is the analysis behind that claim: a reproducible Python pipeline over seven public sources, seven analysis modules, and a static site. Every figure traces to a committed dataset and a dated source. Nothing is typed in by hand, and that includes the nine exhibits, the key figures table and the machine-readable summary on this page.
| Finding | Figure | Source and period |
|---|---|---|
| Merchant payments: share of transactions against share of value | 63.9% / 23.0% | PhonePe Pulse, 2026Q2 |
| Person to person, the mirror image | 30.8% / 71.2% | PhonePe Pulse, 2026Q2 |
| Payment revenue earned on the merchant leg under zero-MDR | Rs 0 | PhonePe Pulse, 2026Q2 |
| What 30bps would have been worth on that same leg | Rs 3,144 crore per quarter | PhonePe Pulse, 2026Q2 |
| Merchant contribution to all volume growth since 2018Q1 | 64% | PhonePe Pulse, 2018Q1 to 2026Q2 |
| PhonePe and Google Pay share of national UPI volume | 45.9% / 32.3%, both above the 30% cap | NPCI, 2026-07 |
| Transactions that must change app for the cap to bind | 4.3 bn a month | NPCI, 2026-07 |
| Private against public bank margin gap | 114 bps (59 pricing, 56 funding) | Yahoo Finance, FY2026 |
| Five year price return, public against private banks | +197% / +11% median | Yahoo Finance, 2021-11-08 to 2026-11-05 |
| Merchant share of own transactions: most against least, material states | Delhi 68.5% against West Bengal 56.4% | PhonePe Pulse, 2026Q2 |
| UPI transactions per banked adult per month | 14.9, up from 4.0 in 2021 | World Bank Findex and NPCI, 2024 |
| Fund schemes against distinct strategies | 14,430 to 3,398 (4.2x wrappers) | AMFI, 04-Nov-2026 |
Not the payment itself. The value sits with whoever owns the merchant relationship, because the merchant leg is 63.9% of transactions, 64% of all growth since 2018, and 0% of payment revenue. UPI is priced at zero for merchants by regulation, so a payment app's return has to come from what the merchant relationship makes possible next: credit, settlement float, software, distribution. Underwrite the distribution, not the transaction.
Read: Underwrite distribution economics, not transaction economics
Zero-MDR is the rule that forbids charging Indian merchants for UPI acceptance. On PhonePe's
merchant volume alone, a 30bps rate would be worth about ₹3,144 crore a quarter; the rule makes
it ₹0. The scenarios in analysis/03_pe_diligence.py price the same volume at 0, 10, 30 and
50 basis points, so the size of the forgone pool is explicit rather than rhetorical. Per
merchant, 30bps is about ₹2,480 a year against ₹2.07 lakh of GMV.
Read: India's payments network monetises the wrong leg
Not on the current trajectory. PhonePe holds 45.9% of national UPI volume and Google Pay 32.3%, so roughly 4.3 billion transactions a month would have to change app for the cap to bind, and at the observed rate of share drift the leader reaches 30% in about 470 months. Concentration is falling, but through the tail gaining rather than the leaders shedding.
The first exhibit draws the market as area: column width is share of transactions, column height is average ticket, so the area of each column is the rupees it moves. It shows that scale and ticket size are unrelated. CRED is the tallest column in the market and 0.6% of it.
Read: Concentrated and unmonetised at once, which is why the share cap cannot bind
Delhi puts 68.5% of its own UPI transactions through merchants; West Bengal puts 56.4%. India runs two different payments markets, not one. The measure is a state's merchant transactions as a share of its own transactions, taken from the per-state category files rather than derived from average ticket size, which would have produced a tautology.
All 36 states and union territories are on the map. Seven of them render at under 2% of the largest state's area, so they get numbered markers and a key rather than being present but invisible.
Read: India runs two different payments markets, not one
Roughly half pricing and half funding: of the 114bps margin gap in FY2026, 59bps comes from what private banks earn on assets and 56bps from what they pay for deposits. The interesting part is stability. The policy rate travelled 250bps across FY2023 to FY2026 and the gap moved 2bps, which says the advantage is structural rather than cyclical.
The market had already decided. Over five years the public bank cohort returned a median +293% against the private cohort's +18%, and the two cohorts do not overlap at any point: the worst public bank beat the best private one.
Read: The private-bank margin advantage is half pricing, half funding
That a single brand can hold both the best and the worst experience in a market. The neobank cohort scores +47 on making a payment and -47 on getting support. Its overall score, +11.8, shows neither. Measuring satisfaction per episode rather than per brand is what turns a survey into a roadmap, because it names the episode to fix.
The panel below is synthetic and labelled synthetic, in the filename, on the chart, in the memo and in the module docstring. It demonstrates the method; it is not evidence about the market. Replacing it with real fieldwork is the top item in the backlog.
Read: Episode-level NPS reveals what brand-level NPS hides
Most of it. 14,288 listed schemes resolve to 3,353 distinct strategies, about 4.3 wrappers each. Debt is the heaviest at 5.3 wrappers per strategy, and ETFs are the honest exception at 1.0, because an ETF cannot easily be sold in growth, dividend, direct and regular flavours. This is a scheme count and not assets under management, which is stated on the chart, because AMFI's public NAV file carries no AUM.
Read: India's fund shelf is four times more administered than it is diverse
Yes, but not in the payment. The investable object is the merchant relationship and what can be sold through it, in a market where acquisition is subsidised by regulation and the incumbent leaders cannot easily be dislodged by a cap that does not bind. The diligence memo carries an explicit ambiguity register for the parts that public data cannot settle, including the largest one: the state-level merchant mix above is PhonePe's, and no open source publishes a second operator's to check it against.
Read: Diligence memo, with the ambiguity register
| Point of view | India's payments network monetises the wrong leg |
| Sector scan | The private-bank margin advantage is half pricing, half funding |
| PE diligence | Underwrite distribution economics, not transaction economics, with an explicit ambiguity register |
| Survey analytics | Episode-level NPS reveals what brand-level NPS hides (synthetic panel, clearly labelled) |
| Market structure | Concentrated and unmonetised at once, which is why the share cap cannot bind |
| Wealth and asset management | India's fund shelf is four times more administered than it is diverse |
| Gap analysis | India runs two different payments markets, not one |
| Method | Sources and provenance · data dictionary · stack decisions |
The exhibits above are static because a README is. The report itself carries a linked workbench: pick a state on the map or in the table and the other two follow it, switch between the map and an equal-area cartogram, change the ranking metric and every label rewrites, filter the 36 rows as you type. No BI backend, no page reload, and the whole thing is server-rendered markup plus one small script.
pip install -r requirements.txt
python run.py allpython run.py data runs end to end in about two minutes and needs no credentials. Every
source is public and unauthenticated, which is also why the monthly refresh in CI cannot fail
silently on an expired token.
| Command | Does |
|---|---|
python run.py data |
Fetch, validate and transform to KPIs (8 fetchers, 17 processed datasets) |
python run.py analyze |
Regenerate every memo, chart dataset and README exhibit (deterministic) |
python run.py site |
Build the static site |
python docs/build_docs.py |
Regenerate the provenance docs from the fetchers' ledger |
flowchart LR
A["PhonePe Pulse · NPCI<br/>Yahoo Finance · FRED<br/>World Bank · AMFI"] --> B["fetch/<br/>validate schema<br/>record provenance"]
B --> C["transform/<br/>build_kpis.py"]
C --> D["analysis/<br/>7 modules"]
D --> E["insights/*.md<br/>site/src/data/*.json<br/>docs/assets/*.svg"]
E --> F["Astro static site<br/>→ Vercel"]
C --> G["docs/sources.md<br/>generated"]
Stack: Python 3.12+ with pandas · Astro 5 static output · hand-written CSS · ECharts, lazily imported · seven hand-written SVG chart components · Vercel · GitHub Actions monthly refresh. Accessibility, Best Practices and SEO all score 100 in Lighthouse on the production URL.
The README exhibits share no code with the site. They are drawn by
docs/build_readme_charts.py in about 600 lines of stdlib
Python, with no plotting dependency, on the site's own palette from
site/src/styles/tokens.css.
- Never fabricate a number. Not fetched or computed means it stays a visible gap. Four months NPCI does not publish are left empty rather than interpolated.
- Fail loud on shape change. Each fetcher validates schema and plausible range. The bank fetcher refused to run when a margin came back negative, correctly: the company was a payments firm, not a bank.
- Synthetic data announces itself, in the filename, on the chart, and in the memo. One dataset here is synthetic and it says so in four places.
The same discipline applies to this page. The nine exhibits, the key figures table and the
machine-readable summary are all written by docs/build_readme_charts.py from the committed
data, so the README cannot drift from the pipeline.
Access dates and per-dataset terms are recorded for every dataset in
docs/sources.md, which is generated from the fetchers' own provenance
ledger, so it cannot drift from what was actually fetched.
| Source | Coverage | Access |
|---|---|---|
| PhonePe Pulse | 2018 Q1 to 2026 Q2 | open, no auth |
| NPCI per-app statistics | 2023-12 to 2026-07 | transcribed (why) |
| NPCI monthly statistics | 2016-07 to 2026-07 | open mirror to 2023-08, then transcribed (why) |
| Yahoo Finance, 11 NSE tickers | 5 years | open |
| FRED, India call money rate | 1968-01 to 2026-06 | keyless CSV endpoint |
| World Bank Open Data | to 2025 | open |
| AMFI scheme universe | daily | open |
The NPCI series joins two sources. The overlap month agrees to the rupee across both, so there is no level shift at the seam, and the site colours the two halves differently rather than merging them silently.
Why PhonePe Pulse and not NPCI's headline series: Pulse is open, needs no auth, is current to 2026 Q2, and is the only feed that splits person-to-person from merchant, which is the monetisation question. The NPCI CKAN mirror everyone reaches for is frozen at 2023-08.
name: India FS Pulse
url: https://india-fs-pulse.vercel.app
repository: https://github.com/DogInfantry/india-fs-pulse
license: Apache-2.0
author: Anklesh Rawat
question: >-
India built the world's largest real time payments network and charges
nothing for the merchant leg. Who captures the value, and is there an
investable business model?
method: reproducible Python pipeline, then analysis, then a static site
publishers: 10
fetchers: 12
analysis_modules: 13
processed_datasets: 25
credentials_required: none
findings:
- metric: merchant share of UPI transactions
value: 0.6386
unit: share of transactions
period: 2026Q2
source: PhonePe Pulse
- metric: merchant share of UPI value
value: 0.2304
unit: share of rupees
period: 2026Q2
source: PhonePe Pulse
- metric: person to person share of UPI value
value: 0.7119
unit: share of rupees
period: 2026Q2
source: PhonePe Pulse
- metric: payment revenue on the merchant leg under zero-MDR
value: 0.0
unit: INR crore per quarter
period: 2026Q2
source: PhonePe Pulse
- metric: registered merchants
value: 50705249
unit: count
period: 2026Q2
source: PhonePe Pulse
- metric: PhonePe share of national UPI volume
value: 0.4589
unit: share of national volume
period: 2026-07
source: NPCI
- metric: Google Pay share of national UPI volume
value: 0.3233
unit: share of national volume
period: 2026-07
source: NPCI
- metric: NPCI per app volume share cap
value: 0.3
unit: share of national volume
period: due December 2026
source: NPCI
- metric: transactions that must change app for the cap to bind
value: 4312.7
unit: millions per month
period: 2026-07
source: NPCI
- metric: private against public bank net interest margin gap
value: 114
unit: basis points
period: FY2026
source: Yahoo Finance fundamentals
- metric: median five year price return, public banks
value: 1.9725
unit: price return
period: 2021-11-08 to 2026-11-05
source: Yahoo Finance prices
- metric: median five year price return, private banks
value: 0.108
unit: price return
period: 2021-11-08 to 2026-11-05
source: Yahoo Finance prices
- metric: highest merchant share of own transactions, material states
value: Delhi 0.6851
unit: share of own transactions
period: 2026Q2
source: PhonePe Pulse
- metric: lowest merchant share of own transactions, material states
value: West Bengal 0.5643
unit: share of own transactions
period: 2026Q2
source: PhonePe Pulse
- metric: mutual fund schemes against distinct strategies
value: 14430 to 3398
unit: count
period: 04-Nov-2026
source: AMFI
caveats:
- PhonePe Pulse covers PhonePe's own transactions, not all of UPI.
- The NIM figure is a proxy: net interest income over average total assets.
- The NPS panel is synthetic and labelled synthetic everywhere it appears.
- Four months NPCI does not publish are left as gaps, never interpolated.@software{rawat_india_fs_pulse_2026,
author = {Rawat, Anklesh},
title = {India FS Pulse: who captures the value in India's UPI under zero-MDR},
year = {2026},
url = {https://github.com/DogInfantry/india-fs-pulse},
license = {Apache-2.0}
}Code in this repository is licensed under the Apache License 2.0. Each dataset
retains its own publisher's terms, recorded per dataset in
docs/sources.md.
Original work. Not affiliated with, endorsed by, or using the trademarks or proprietary data of any consulting firm. Net Promoter Score is a public method; no proprietary benchmark data is used. Nothing here is investment advice.
See docs/stack-decisions.md for what was rejected and why.