If UPI Is Free, Who Pays for It? | PhonePe, Razorpay & NPCI | WTF is Payments?
Executive Summary
Money keeps getting lighter - grain, gold, paper, plastic, and in India, a sound. This episode is the definitive anatomy of that sound. UPI is not an app but a six-party relay: your app talks to a sponsor bank, which talks to NPCI, which resolves your virtual address and coordinates the issuer and receiver banks - roughly 10 API calls in milliseconds, at an estimated ecosystem cost of 25 bips, of which the issuer quietly pays the PSP 15-50 paise per transaction. 'Free' is a policy choice: zero MDR in 2020 made UPI explode but froze private investment - a hundred TPAPs came and went - and the ecosystem is now debating how to bring pricing back without killing inclusion. Along the way: the Big Billion Day that failed 9 of 10 million card attempts and taught Sameer Nigam payments; Razorpay's real economics (charge 1.85%, pay 1.75% downstream, keep 10 bips, and carry merchant-fraud liability); why 2FA cut fraud but also success rates; why regulation should punish abuse rather than size; and what comes next - B2B payments, escrowed COD, CBDCs over stablecoins, agentic payments on UPI, and whether the cost of intelligence is really going to zero.
Key Takeaways
UPI is a relay, not an app
A 100-rupee transfer crosses PhonePe (the TPAP), a PSP/sponsor bank, NPCI (which resolves your virtual payment address), the issuer bank and the receiver bank - about 10 API calls in milliseconds. NPCI exists so no one needs bilateral pipes; net settlement runs through RBI's RTGS multiple times a day, so no intermediary ever holds float.
Free UPI costs somebody
The ecosystem cost is estimated at ~25 bips per transaction, with the issuer bank paying the PSP a 15-50 paise switching fee. Zero MDR (and the unique P2PM merchant code) killed the revenue model for third-party apps - 'a hundred TPAPs have come and gone' - and the panel agrees pricing must return in some form for investment to continue.
The 2% card fee is mostly the bank's
On a credit card transaction the issuer takes roughly 1.45% (cost of credit plus rewards), the network around 0.2%, and a gateway like Razorpay keeps about 10 bips on large merchants after paying 1.75% downstream. The fee is ad valorem because the gateway underwrites merchant-disappearance risk.
Big Billion Day was payments' graduation
In 2015 Flipkart asked banks to 100X capacity; 75-80 million logged-in users took down core banking systems anyway. Of ~10 million debit attempts in eight hours, 9 million failed - including 35-40,000 TVs 'deemed successful' that Flipkart honoured by buying stock from retail vans. Today UPI alone does ~783 million transactions a day, roughly 49% of global real-time payment volume.
Regulate abuse, not size - or trust
Nikhil's written lessons: reduce regulation, remove fear, make 'innocent until proven guilty' real - the process has become the punishment. The counterpoint from the panel: intervene on abuse and bundling (the Microsoft/IE precedent), not on scale itself, and remember interoperability is what keeps real choice alive. India's trust deficit and weak contract enforcement are why rules pre-empt rather than punish.
Payments data is an X-ray of India
Tier-2 and beyond contributes 65-70% of digital payments, with village-to-village flows, not just metro-to-hometown. 70% of e-commerce spend comes from about 15 million households. COD is a trust problem, not a payment problem - hence the escrow/double-auth idea. PhonePe's GST study: a 2% rate cut lifted category consumption ~5.5-6%; a 5% cut, ~13%.
The next fights: B2B, cross-border, agents
B2B is 'completely unsolved' (a small retailer juggles 70-80 distributors with invoice-GST workflows; NPCI shipped multi-signatory UPI). RBI clearly prefers CBDC rails and projects like Nexus over stablecoins, whose on/off-ramps would be regulated like banks. NPCI plans agentic payments on UPI (GFF showcase), while founders debate whether token cost goes to zero - Harshil runs infinite tokens on a home DGX Spark for under 1,000 rupees of power a month.
The Conversation Breakdown
Three roads into payments
⏱ Untimed topicSohini Andani: 18 years at Axis Bank doing 'everything non-core', Western Union, then leading growth at NPCI - a Section 8, not-for-profit built by RBI and IBA whose surplus goes back into the rails. Harshil Mathur: Jaipur, IIT Roorkee mechanical, a college software club, Schlumberger, and a crowdfunding side project that could not accept online payments - so he and Shashank built Razorpay. Sameer Nigam: a naval childhood, ShopZilla in LA, Wharton, a digital music startup sold to Flipkart in 2011.
The six-party diagram
⏱ Untimed topicThe full P2P flow drawn live: app, PSP, NPCI, issuer, receiver, plus VPA resolution for mobile numbers. Who bears cost where, why the PSP layer exists (RBI didn't want unregulated apps touching the network directly), edge and network-level fraud models, 30,000 servers, and net settlement that makes a massive gross day collapse into small ledger moves.
The night the banks went down
⏱ Untimed topicFlipkart's first Big Billion Sale: placards reading '25X all servers', graveyard-shift testing, 75-80 million people waiting at 8am, core banking systems falling one after another, SBI's ATMs down for the day, and the 'deemed successful' television fiasco - the failure modes that taught a generation of Indian fintech how payments really behave at scale.
Cards, P2M and the fraud budget
⏱ Untimed topicPull vs push transactions, the two-leg OTP dance, MDR splits on a 1.85% charge, why RuPay credit on UPI (with bio-auth) is the panel's favourite experience, merchant-side liability, and the US-India trade: two-factor authentication crushed consumer fraud but dropped success rates - and success-rate engineering became Razorpay's edge.
The regulation debate
⏱ Untimed topicZero MDR and its investment chill; caps and market share (Nikhil's own 15% OI cap at Zerodha cost him the far-options business); monopoly vs abuse of monopoly; banking licences as the real moat (1,300 banks, 35 universal); Flipkart's ED probe; bike-taxi bans; 'the first actor in every category gets burned in India' - and the counter-argument that interoperable, centrally-ruled markets are precisely what India got right.
What the data says about India
⏱ Untimed topicTier-2+ dominance of digital payments; the narrow 15-million-household core of e-commerce; quick commerce decimating offline in dense pin codes while grocery stays 90% offline; assisted commerce (travel agents, insurance brokers) as a trust interface; GST-rate elasticity measured category by category; and why property and inheritance taxes keep coming up in this group.
B2B, borders and agents
⏱ Untimed topicWhy B2B payments need an operating system, not a rail; the tipping problem baked into UPI's push architecture; stablecoins vs CBDCs and why control lives at the on/off-ramps; agentic commerce (Nykaa and ixigo agents, NPCI's interoperability framework, the x402 protocol); the AI bubble argument both ways; and the closing founder panel - Seven's payment rings, Dodo's merchant-of-record model, PayGlocal's cross-border orchestration lifting payment success rates from 50-60% to 80-90%.
Core Frameworks & Ideas
Follow the switching fee
'Free' payments run on a 25-bip ecosystem cost and a 15-50 paise issuer-to-PSP fee. Map who pays whom per hop before believing any 'zero-fee' claim - the cost never disappears, it moves.
Punish abuse, permit scale
India's post-70s stance: monopolies can naturally exist, abuse of monopoly cannot. Interoperability plus conduct rules (not size caps) is the middle path between killing ambition and enabling gatekeepers.
Friction is a fraud budget
Two-factor auth traded success rate for fraud reduction in India; the US did the reverse. Every authentication decision prices an implicit fraud-vs-conversion trade-off - engineering that delta is the actual payments business.
Actionable Next Steps
- ✓Selling online? Treat payment success rate as a revenue line - PayGlocal shows moving PSR from 50-60% to 80-90% is recovered top line, not a tech metric.
- ✓If you run an offline business, keep your flows digital: your UPI data footprint is fast becoming the basis on which lenders price your credit.
- ✓Want to build in fintech? Get a year inside a payments company first - the founder panel's unanimous advice over starting cold.
- ✓Hold off on consumer agentic-payment products for India: the panel's call is the wave is 2-3 years out - build the enabling layers (orchestration, compliance, specs) instead.
Notable Quotes
“The entire history of money has been one move, repeating. Money keeps getting lighter.”
“Unless you go fix this, you cannot onboard new customers.”
“Like, I have a DGX Spark running at my home on which I have a DeepSeek V4-Flash model. I can run infinite tokens on it for zero. The margin cost is zero.”
The One Thing to Remember
If you're not able to make money on payments for payments' sake, you will not stay in payments for too long.