A few months ago, FIIs were exiting India, the rupee touched 97, and the hundred-mark loomed psychologically. The RBI stepped in - not with ordinary open market operations, but with a full revival of the FCNRB (Foreign Currency Non-Resident B) deposit scheme, last deployed during the 2013 Taper Tantrum.

How the scheme works: An NRI walks into an overseas HDFC or SBI branch, deposits say $100K, takes a leveraged loan (up to 10:1) from that branch, and deposits the resulting ~$1M into an Indian bank. The bank converts to rupees and deploys the liquidity normally. In 3–5 years it repays the NRI in dollars. The RBI backstops all currency risk.

In 2013, there was a genuine run - the Fed's taper signal triggered a specific, identifiable event. FCNR (B) was a well-timed emergency brake. It worked.

In 2026, none of that was true. The rupee's decline was real but not a cliff. Reserves were already substantial. Growth was healthy. The current account was nearly balanced. This was a correction, partly driven by geopolitical overhang from the Middle East - not a run. Yet $136B came in (including $127B directly through FCNRB). The scheme overshot badly.

All those rupees now sit on Indian bank balance sheets. Banks have far more deposits than they can lend out. Surplus liquidity in the banking system jumped above ₹10 trillion.

Prof. Prasanna Tantri (ISB) lays out five options in a Business Standard piece, also drawing on the Viral Acharya, Raghuram Rajan et al. paper "Liquidity Dependence: Why Shrinking Central Bank Balance Sheets is an Uphill Task". We discuss how each option carries a huge risk.

The core argument: The RBI fired a bazooka it only gets to use once. NRI appetite may not be there next time, terms will be worse, and India's sovereign credibility becomes a pricing factor. The rupee was already drifting back toward 97 as this episode was recorded - Saudi Arabia's East-West pipeline had just been struck in multiple places. If the situation escalates further, the RBI has less ammunition and less credibility.

"You only get to scream fire in a crowded building once."

----------------------------------------

Apple's new launches - What is with the Indian pricing?!?

The folding phone arrived. $2,000+ in the US, ~₹3 lakh in India, and before the ₹40–50K Apple Care Plus you kinda sorta need (screen replacement: ₹1 lakh+). The hinge is genuinely beautiful engineering, though.

iPhone 17 now starts at ₹1 lakh. A year ago ₹70K felt like value; at a lakh, ehhhh.

What does sound like good value: AirPods at ~₹13–15K with ANC.

----------------------------------------

References

* Prasanna Tantri (ISB) — Business Standard op-ed on FCNRB second-order effects. https://www.business-standard.com/opinion/columns/after-the-dollar-deluge-costly-choices-126091001676_1.html

* Viral Acharya, Rahul Chauhan, Raghuram Rajan, Sascha Steffen —-"Liquidity Dependence: Why Shrinking Central Bank Balance Sheets is an Uphill Task" https://www.kansascityfed.org/documents/9666/JH2022_Acharya.pdf

* 2013 Taper Tantrum and original FCNRB mobilisation: https://www.moneycontrol.com/news/opinion/a-re-look-at-rbis-extraordinary-steps-in-2013-to-stabilise-the-rupee-8887771.html

https://www.investopedia.com/terms/t/taper-tantrum.asp

The Lok Sabha just passed a bill approving merchant discount rates on UPI. Nitesh has thoughts. Many of them.

UPI is the single most effective instrument of economic formalization India has ever produced, and it happened without anyone being forced to use it. Cash-based shopkeepers started accepting UPI because their customers wanted to pay that way. Their transactions now had an audit trail. Their wholesalers had to start issuing bills. GST input credits started flowing up and down a supply chain that had run entirely in black for decades.

The cost of running this entire infrastructure is around 2,000 crore a year. A rounding error. And now someone wants to charge for it.

The argument for MDR is the Uber argument: the free rides were for adoption, and the true cost must eventually be recovered. Our answer is that we already know the true cost, it isn't that much, and the payoff in tax formalization alone has returned it a hundred times over.

We also discuss what friction does to frictionless systems. UPI works because the merchant gets a beep and the transaction is done. No doubt, no renegotiation, case closed. The moment you introduce a cost, however small, you introduce a calculation. The panwadi selling a one-rupee chewing gum doesn't get hit directly, but when he pays his distributor via UPI, that payment gets hit. The distributor asks for 310 instead of 300. The supplier does the same thing one step up. All of it finds its way to the end consumer eventually, because it always does, regardless of who the government says it's charging.

The digital arrest scam sidebar deserves its own episode someday. Fraud reversal is nearly impossible not because the technology doesn't support it, but because mule accounts are so numerous and money moves through them so fast that by the time anyone raises an alarm, the funds have been chopped, scattered, and are genuinely untraceable. The solution that keeps getting floated, a waiting period or escrow on UPI transactions, would kill the thing it's trying to protect. Nitesh recommends the book Scamlands for anyone who wants to go deeper on this.

We also discuss Ola pivoting its two-wheeler business to a dealer model, and a brief AI update to closes things out: DeepSeek is fast, Claude feels slower, and commoditization is now visible in real time with Chinese models hitting the same coding benchmarks as American ones at a fraction of the price. And Situational Awareness is back raising money after nearly blowing up, which is either completely rational or completely insane depending on how close you're standing.

India just launched a futures contract for the monsoon. You can now bet on whether it will rain.

Nitesh opens with the mandatory disclaimer - futures and options, do not try this at home, the statistics are stacked against you - and then proceeds to deliver the clearest explanation of derivatives this podcast has ever attempted. The original purpose of a futures contract was to let a corn farmer and a popcorn manufacturer lock in a price a year in advance and eliminate uncertainty for both. Options added optionality - the right but not the obligation to transact at a fixed price. The premium you pay is like insurance. The options writer - Jane Street, for example - collects small premiums from a million bets and pays out on the small fraction that go against them. Their downside is theoretically unlimited. Their algorithm makes sure it isn't.

Then the new product: RAINMUMBAI. A monsoon futures contract, live for four months (June–September), monthly expiry, built by NCDEX, IIT Bombay, and the Indian Meteorological Department using 100 years of rainfall data from two stations in Mumbai - Santa Cruz and Colaba. The contract is indexed to daily rainfall measured from 8:30am to 8:30am. If it rains less than the 100-year average for that date, the contract drops. A builder in Mumbai whose workers can't show up in heavy rain can short the contract to hedge their fixed costs. A fertiliser manufacturer who needs rain to sell product takes the other side.

A quick question: why financialise something insurance could solve? And a quick answer - insurance claims don't settle automatically, require a surveyor, involve data disputes, and take months or years.

We also discuss Sam Altman's hardware play - building a smart speaker to reinvent Alexa, except Alexa already exists and doesn't do too much - and a broader conversation about when smart people in one domain get treated as gods in all domains.

We close with Lagaan 2 — Bhuvan buys a monsoon futures contract, it rains at the end, the contract doesn't pay off, he wins the cricket anyway, the British somehow still honour the deal. It's fiction.

Apple practically doubled its prices overnight. Your PlayStation costs $900. Vineet's SanDisk SSD died and the company won't replace it because the replacement now costs 12,000 rupees more than what he paid. All of this has one cause, and it isn't the tariffs. (Spoiler alert: It's Sam Altman)

Samsung, SK Hynix and Micron control 95% of global DRAM production. They have systematically reallocated their capacity toward high-bandwidth memory for AI accelerators - and every wafer that goes to an AI data center is a wafer that doesn't go to a consumer laptop, a phone, or a console. This is not a temporary blip. It is a structural reset.

The margins are pretty awesome for the manufacturers though! Micron at 56% gross margin in Q1 2026. SK Hynix at 79%. Nvidia at 75%. Apple sitting at 47.86% gross margins - and choosing to pass every rupee of cost onto the consumer rather than give bad news to investors.

The top comment on The Verge's Apple price hike article: "I'm going to go out on a limb and say I have not had any productivity increases from AI worth the financial burden it has placed upon me and my family."

Nitesh's take: this will normalise through inflation, not through price drops — in 6-7 months, $800 will just feel like what a laptop costs, the same way everything else does. Vineet's take: it's a choice at every level, Apple to Micron to Sam Altman, and none of them deserve the "we had no choice" framing they're all reaching for.

Then Nitesh with an update on Varun Beverages - now distributing Asahi's yogurt drink in India, having convinced PepsiCo to let them rent out their cold chain distribution muscle to any beverage brand that wants plug-and-play Indian market access.

Four companies. Four IPOs. Four completely different stories about what it means to go public in 2025 - and what that tells you about the state of capital markets, quick commerce, and the art of calling yourself a trillionaire.

Nitesh walks through SpaceX, NSE, Jio Platforms, and Zepto all heading to market (almost) simultaneously. We start with the numbers - $75 billion, $60 billion, $150 billion, $6 billion - and work backwards to what each of them actually means.

SpaceX is the con told in public: 4.2% of shares in the free float, the rest used to calculate a $1.75 trillion valuation that then crossed $2.5 trillion before coming back down to Earth. Musk immediately used the listed company as collateral to acquire Cursor via share swap and is now considering a $20 billion debt raise - one week after the IPO. Every single criticism Hindenburg had of Adani, Musk is doing in the open. Hindenburg no longer exists. Funny that.

NSE is the straightforward one - a decade-long saga of spectacular bad luck has led up to this IPO. 100% offer for sale. No new capital needed. A literal toll booth on every trade in India regardless of direction, and the most active retail derivatives market in the world feeding it daily. Radhakishan Damani holds a stake worth around 8,000 crores and isn't selling a single share. That should speak volumes.

Jio Platforms is the Reliance play - fresh issue only, all proceeds go to retiring debt, 2.5% float to start with a path to 25% over three to five years. A nine-year-old company targeting a market cap equivalent to AT&T, which took a hundred years and the merger of dozens of storied companies to build. Nitesh will die on the hill that this is genuinely phenomenal regardless of how it was built.

Then Zepto - another one with impressive numbers. Revenue from operations at ₹22,624 crores, doubled year-on-year, net loss of ₹6,000 crores. 1,200 dark stores, plans for 1,900 more by 2030. Classic heavy capex upfront story - sweat the assets, hit the breakeven, hockey stick of profit. The ED summoned the founders under FEMA during fundraising, which is a rather unique risk factor to mention in an RHP. Gig worker legislation looming. And ₹1,636 crores in ad revenue, because everything eventually becomes an advertising business.

We close on the gig worker question properly - Zepto, Blinkit, and the rest calling delivery workers freelancers while tracking their every minute, penalising downtime, and counting waiting time as not their problem. Nitesh: "Clearly you have his entire mental bandwidth for every waking moment." Vineet: "Pick a lane. Either you're a job creator or you're not."

Vineet's video dies midway, so maybe we should call this 3/4th of an episode?

Google is going to Google for you, and tell you the results - rather than pointing you to a webpage. That changes the Internet...a LOT.

Two Google stories this week that are really one Google story - about a company that has spent 25 years collecting everything about you, and has finally figured out what to do with it.

We open with Amazon Prime putting a "add to cart" button inside TV ads. No friction. See product, want product, own product —-in 15 minutes. Nitesh watched it happen on his Firestick and was genuinely unsettled. We are a cynical, hardened generation and we might be fine. Our parents and our kids are not.

Then the main story: Google is deprecating links. AI overviews are replacing search results. 93% of searches are already zero-click. Click-through rates are down 37-47% depending on which study you read. People are still on Google - they're just not going anywhere. Which raises an obvious question: if nobody visits your website anymore, why would anyone write anything on the internet? The incentive structure that built the open web is being quietly dismantled, and the replacement is Google answering your question itself using content it scraped from the people it's now cutting out.


Then Google Spark —-their new all-in-one personal assistant. A Verge reporter asked it to plan a weekend trip. It knew his son's age, scheduled nap time, his wife's name and dietary preferences, pulled his concert tickets from a Ticketmaster email, and called his parents by name when offering to arrange childcare. That's both very impressive, and deeply terrifying. The data was always there. Google just finally showed you what they had.

We close with Ixigo quietly acquiring a majority stake in Brevistay - the hourly hotel booking platform - for 66 crores. We have no insights to add, just some obvious jokes.

Links:
https://blog.google/products-and-platforms/products/search/original-high-quality-content-search/
https://www.bain.com/insights/goodbye-clicks-hello-ai-zero-click-search-redefines-marketing/
https://www.theverge.com/ai-artificial-intelligence/941388/gemini-spark-ai-agent-trip-planning

The word "AI" has meant nothing specific for 60 years. That's not a bug, it's the main feature.

This week we have a guest: Jaidev Deshpande, ML engineer, vibe coding janitor, and chief code deleter, who joins Vineet and Nitesh to try and parse what AI and AGI mean.

We open with the defective magic kit. The AI industry has perfected a single rhetorical trick: when the product doesn't work, it's your fault for not knowing how to use it. This isn't new. The first AI chatbot, Eliza, was a rule-based script. Its own creator came out and said it was nothing special - but the public refused to believe him. That was 1966. Nothing has changed except the valuation.

AGI gets its proper dissection. The goalposts have shifted so many times that nobody - not Dario, not Sam, not the researchers - can define what human-level intelligence even means. The Turing test was passed. Nobody declared AGI. Code generation is here. Nobody declared AGI. Microsoft and OpenAI had to define it for their lawyers: AGI will be achieved when OpenAI turns a hundred billion dollars in profit. That definition has since been quietly dropped from their latest partnership agreement.

Vineet walks through his actual workflow - building AI video for enterprise clients, and gets into why the dropoff rate for folks trying to DIY their way through AI video, is crazy high.

The term "Vibe coding janitor" used to be a joke. It isn't anymore.

Then the numbers. Nvidia's last quarter: $50 billion in operating cash flow, 95% of which went straight back into circular investments — CoreWeave, downstream stack, the whole flywheel. Anthropic claims $30 billion ARR but filed in court in March that lifetime revenues are $5 billion-ish. If MRR in March was $2.5 billion, their February number of $10 billion ARR can't both be true. The numbers don't add up and nobody is asking why.

SpaceX's S1 gets a reading - Nitesh has made it through the first hundred pages and reports that the space and launch business is genuinely extraordinary, the Twitter-xAI integration is where it falls apart. Musk wants to solve the data center energy problem by putting them in space. We debate if the funding, the appetite and the capability exist.

The closing argument brings in the foreign hand: AI models can't raise prices because they haven't locked in consumer use cases. DeepSeek and Qwen exist. The moment American models hit $50 a month, the switch to Chinese models becomes viable overnight. Jaidev says asymptotically, it will happen — just a question of when. It's not doomerism — it's just the most natural outcome. Nitesh tries to bring out the positives of the bubble.

We end on tramp stamps.

Links:
https://www.beingbayesian.in/p/the-compounding-fallacies-of-ai-hype

The AI chip trade is moving downstream fast. And the Indian economy is moving somewhere nobody wants to talk about.

Two stories this week that are bigger than their headlines suggest.
Vineet opens with a tour of the AI semiconductor stack - and the uncomfortable math behind it. Nvidia's crazy spike was just the start. AMD up 200%, memory plays like SanDisk up 400%, cooling companies, networking layers, power infrastructure, optical interconnects - every bottleneck that gets solved just reveals the next one. But GPUs bought in 2025 won't go live until 2027 or '28, by which time they'll be two generations old and possibly already written off. Satya Nadella has said openly that the constraint isn't GPUs — it's energy and the buildings to house them.

Then the detail that broke Vineet: compute futures. Someone is actually proposing an actively traded F&O market for compute hours, awaiting SEC approval. The original purpose of futures contracts was to help potato farmers sleep at night. This is that, except the potato is a GPU warehouse and nobody knows who the end buyer is yet. Nitesh is horrified and also excited.

Then Nitesh's story. India's consumption-driven economic pitch to the world has always been "come here, sell here." The PM's statement asking citizens to reduce gold purchases, cut foreign travel, and conserve fuel is, unfortunately, an acknowledgment that the consumption engine is under serious stress.

The hidden structural problem: India's retail investors have been nudged, incentivised, and tax-restructured into equity markets as the only viable savings vehicle. Debt mutual funds now offer no indexation benefit. FDs don't beat inflation. So when sentiment turns, people reduce spending and drawdown their investments - a potentially vicious cycle. The tide that rose monthly on SIP inflows will go out the same way.

This week we get into the data center story properly. One upcoming facility in Utah will consume more power than the entire state does today. Data centers account for 0.01% of US jobs but 4.4% of electricity consumption. Anthropic is spending $50 billion on infrastructure that will create 800 jobs. And Microsoft, OpenAI, and Anthropic have all simultaneously published press releases promising to be "good neighbors" - which is as close to a public admission as you'll ever get that they haven't been.

We also dig into what the Google Andhra data center actually means for India. The press release has all the keywords - Viksit Bharat, gigawatt-scale, national industrial ecosystem - but run the math and the majority of that investment goes straight back out of the country in GPU purchases.
So...what are we actually getting in return for all this energy use?

Nitesh breaks down the Indian companies that are getting in on the AI action.

Then: the OpenAI vs Musk trial, which has confirmed what everyone suspected - the entire AI industry is about seven people who cannot stand each other, and by extension so is the global economy. Discovery has given us Mira Murati testifying under oath that Sam Altman lied to her, Ilya Sutskever's memo saying Altman has a consistent pattern of undermining everyone, Greg Brockman saying he genuinely thought Musk was going to physically attack him, and a board member describing OpenAI's safety methods as "becoming somewhat less slapdash over time." Musk, for his part, didn't read beyond the first box of a four-page document before suing over it.

Links:
A fun CNBC interview with Gamestop CEO Ryan Cohen:
https://www.youtube.com/watch?v=Bmj2PaxX24E

Sources for data center news and surveys:
https://www.pewresearch.org/short-reads/2026/05/05/many-americans-hold-utility-companies-responsible-for-their-rising-home-energy-bills/
https://www.sltrib.com/news/2026/05/04/utah-data-center-final-vote-box/
https://www.theverge.com/ai-artificial-intelligence/902546/data-centers-ai-energy-power-grids-controversy
https://www.theverge.com/news/889578/data-center-power-pledge-white-house-google-meta-microsoft
https://www.bloomberg.com/graphics/2025-ai-data-centers-electricity-prices/
https://www.anthropic.com/news/anthropic-invests-50-billion-in-american-ai-infrastructure
https://www.storyboard18.com/how-it-works/explained-what-googles-discom-licence-in-andhra-pradesh-really-means-94979.htm

AI companies want to create myths around themselves - and now they're literally calling a product "Mythos". Just in case, y'know, you miss the point.

Anthropic announced Mythos - a new model capability so powerful it found multiple zero-day vulnerabilities in critical banking infrastructure, alarming enough to call an emergency meeting between the Fed Chair and the Treasury Secretary. OpenAI, not to be outdone, had "sources" leaking within days that they had something similar.

In this episode:
We open with the UBI debate - but not the one you've heard. Sam Altman's floating the idea of giving everyone equity in AI companies as their slice of the coming abundance. So it's time for...AI SIPs? #AIfundssahihai . Also, a side-rant on how the rich people hoarding all the productivity gains don't want to pay taxes, and no one has actually explained where UBI money comes from. The "abundance mindset" crowd keeps skipping ten steps in the middle.

A brief but important detour: free bus rides for women in Karnataka. 250 crores a year. The moral hazard brigade comes out in force. Nitesh is furious about it in the most reasonable possible way. Almost too polite.

Cover story: Project Glasswing - Anthropic's new Mythos model, the zero-day vulnerability story, and what it all means. Is this a genuine capability breakthrough? A strategic positioning move against the US government fallout? Both? We dig into why fear has always been the most reliable fuel in the AI news cycle, why Dario makes doomsday predictions every three months, and are companies adopting AI because it works or because they can get away with it.

We also return to the consumer vs. B2B question - the $20 plan is dying. And if the $20 plan dies, the whole consumer narrative goes with it. The Anthropic-US government fallout gets a proper dissection: can you be the ethical AI company while a huge potential revenue stream is government contracts from the Department of War? Can you refuse to hand over autonomous targeting decisions while being fine with everything else?

Then Nitesh's story: Andrew Ross Sorkin's 1929 - and what the mechanics of that crash look like when overlaid on today. Margin trading. Hidden linkages. Correlated assets in a crisis. A small-scale preview on India's election results day - small caps crashing not because anything was wrong with them, but because someone's broker issued margin calls on pledged shares. The brand-new Polymarket angle: US Air Force pilots betting on whether their strikes will hit targets in Iran. White House insider trading on oil stocks.
We close with Meta Ray-Ban glasses being called "pervert glasses" by Wired, a debate on whether the cultural backlash can kill the category (Vineet says yes, Nitesh says Apple won't let it die), Chinese CCTV bans that ban nothing, and a post-dated disagreement to revisit in April 2027.