And then the titan rose from his seat, adjusted his collar and said, “There is an easy and effective way in which we can reduce the costs of our rocket launches!” Everyone looked at him with hope. The leader said, “What?” He replied, “Why don’t we use 25% ethanol blending in rocket fuel? That way we can drastically reduce costs as those currently being used are already too costly to produce.”

Reports suggest that the leader has been waking up half an hour early from the very next day to do some more yoga postures and maintain his calm. But the central problem still hangs naked, “How to reduce the cost of rocket launches in India?”

Cost Breakdown

Before asking how India can reduce launch costs, we first need to know where the money actually goes.

A modern orbital rocket is one of the most expensive machines ever built because it burns its engineering within a few minutes of its launch.

Here’s what a typical expendable launch vehicle looks like.

The first stage is where almost all the money sits.

It houses the rocket’s most valuable hardware: multiple engines, turbopumps spinning at tens of thousands of RPM, flight computers, navigation systems, hydraulic actuators, pressurisation systems, valves, wiring and massive propellant tanks. Aerospace studies estimate that this single stage accounts for roughly three-fourths of the vehicle’s manufacturing cost.

The second stage is much simpler. It only has to operate once in the vacuum of space and doesn’t need landing systems or the structural strength required for launch. As a result, it contributes only about 15–20% of manufacturing costs.

Even the payload fairing, the nose cone protecting the satellite, is surprisingly expensive. Made from lightweight composite materials and designed to survive supersonic flight before separating flawlessly, it can account for another 5–10% of the rocket’s cost. SpaceX now routinely recovers and reuses these as well.

And the fuel?

Ironically, it is almost irrelevant.

For Falcon 9, studies estimate that the combined cost of rocket-grade kerosene and liquid oxygen is less than 1% of the launch price, roughly 0.3% in some analyses.

Which brings us to absurdity.

After barely 150 seconds of flight, the first stage completes its job...

...and falls into the ocean.

That means every launch throws away around 70% of the rocket’s manufacturing value after just two and a half minutes.

SpaceX looked at this and asked a simple question:

What if we stopped throwing away the most expensive part?

That single question changed the economics of space forever.

What did America, and recently China, do?

America and China are currently leading the space race with substantial and timely contributions from India. But who started the idea of reusable rockets? Let’s go back in time!

The Americans not only invented reusable rockets.

They made them economically useful.

For decades, engineers knew that recovering rockets was theoretically possible. The problem was that refurbishing them often cost almost as much as building new ones. SpaceX changed the equation by designing the Falcon 9 so that its first stage could autonomously return to Earth, land vertically, undergo minimal refurbishment and fly again. Today, individual Falcon 9 boosters have flown more than 30 missions each, and the company has completed hundreds of successful booster recoveries, making reuse a routine part of launch operations rather than an engineering spectacle.

But Elon Musk, as always, wasn’t satisfied.

(The picture might break your reading flow, but telling the story without this picture would be cheating)

Landing a booster still meant adding heavy landing legs, transporting it back to the launch site and spending days preparing it for another flight. So with Starship, SpaceX went one step further. Instead of landing on legs, the returning booster is caught mid-air by giant mechanical arms, nicknamed Mechazilla. No landing legs. Less weight. Faster inspections. Quicker turnaround. The goal is simple: make rockets behave less like disposable machines and more like commercial aircraft.

And now, China has entered the race.

Last week, China successfully recovered the first stage of its Long March-10B using a sea-based net capture system, a completely different approach from SpaceX’s chopstick arms. By replacing landing legs with a lightweight capture mechanism, Chinese engineers hope to save mass, carry heavier payloads and reduce refurbishment time. More importantly, this wasn’t a one-off technology demonstration. It is part of China’s larger plan to support its satellite megaconstellations, lunar ambitions and a future of high-frequency launches. Another reusable Long March mission is already planned later this year.

Different engineering.

Same idea: stop throwing away the most expensive part of the rocket.

But why is catching so important?

Reducing launch costs is just the beginning.

The real prize is speed. Every recovered booster is one less rocket that must be built from scratch. This allows a country to launch satellites far more frequently, replace failed ones within days rather than months, and respond swiftly in emergencies or conflicts. In today’s space race, launch cadence has become just as critical as raw launch capability.

Then there is scale. The coming decade will be defined by the ability to launch thousands of satellites in space. Megaconstellations like Starlink, China’s Guowang, or future Earth observation networks only become truly viable when rockets can fly repeatedly instead of being rebuilt after every mission.

There is also a powerful strategic dimension. Modern militaries rely heavily on satellites for navigation, surveillance, communications, and missile warning. In a conflict, the side that can replenish disabled orbital assets the fastest holds a major advantage. Reusable rockets transform space access from a peacetime luxury into a genuine wartime capability.

Finally, there is industrial capability. Building a rocket is hard.

But, building one that can survive launch, fly at hypersonic speeds, return through the atmosphere, land with metre-level precision and fly again is an entirely different game. It requires mastery over advanced propulsion, autonomous guidance, materials science, precision manufacturing and high-reliability software. In many ways, catching a rocket is less a space milestone than a demonstration of a nation’s engineering maturity.

And there’s another reason this race matters: AI. The world’s biggest technology companies are already exploring space-based data centres powered by continuous solar energy, where satellites process data in orbit instead of sending massive volumes back to Earth. While the concept is still experimental, agencies like ESA and several private firms see it as a long-term solution to the exploding demand for AI compute. But none of it is feasible without one prerequisite: cheap, high-frequency, reusable launches. If AI’s future extends into orbit, the countries that master rocket reusability today will be the ones building tomorrow’s cloud infrastructure.

Which is why we see the race to catch the future of space access via catching the rockets!

Where are India’s fielders?

If this were cricket, India has finally walked onto the field.

The problem is that America is already fielding, and China has just started taking catches.

India’s reusable rocket journey began in 2016, when ISRO launched the Reusable Launch Vehicle-Technology Demonstrator (RLV-TD), a winged prototype that looked more like a miniature space shuttle than a Falcon 9. The goal wasn’t to recover boosters vertically, but to master one of the hardest parts of reusability: surviving re-entry and landing autonomously.

Seven years later, in 2023, ISRO crossed another milestone. Nicknamed Pushpak, the vehicle was carried to 4.5 km by an Indian Air Force Chinook helicopter, released mid-air and landed itself on a runway without a pilot. A second successful landing experiment followed in 2024, proving that India could repeatedly execute autonomous landings.

But here’s the catch.

Pushpak isn’t the rocket that will carry India’s satellites.

It is a flying laboratory.

The real game is NGLV (Next Generation Launch Vehicle), internally known as Project Soorya. Unlike today’s PSLV and LVM3, NGLV is being designed from day one with a reusable first stage, landing legs, advanced guidance systems and recovery capability. It is expected to become the workhorse for India’s future space station, lunar missions and heavy commercial launches.

And now, the private sector is also beginning to pad up.

Skyroot Aerospace has spoken about incorporating reusable technologies into its future Vikram launch vehicles, while Agnikul Cosmos is developing highly modular rockets whose rapid manufacturing philosophy aligns naturally with future reusability. Newer startups such as EtherealX have gone even further, placing fully reusable heavy-lift rockets at the centre of their long-term vision and Astrobase is targeting reusability of rockets by 2030 for India

So when will India actually catch its rockets?

If India’s current roadmap stays on track, the first demonstrations of reusable booster technology are likely to happen toward the end of this decade, with operational reusable launches expected in the early 2030s alongside the NGLV programme.

One who innovates always wins!

This roadmap requires steady nerves, but betting against India has rarely been a winning move. Our journey, stretching from satellite parts on cycle carriers to the high-stakes precision of the lunar south pole, is a record of defying the odds. We find our way. Between the silent corridors of ISRO and the hungry ambition of deep-tech disruptors, a familiar spirit is currently waking up. One day, the data will sync, the engines will throttle, and India will finally secure that historic catch over the waves.

When that moment lands, it will be more than a technical feat; it will be a profound national dawn where we finally stand as equals among the elite orbital powers.

But parity cannot be the final destination. The space race moves with too much momentum for us to simply follow existing trails. The real hurdle in this era is innovating with such radical speed that we start defining the pace. We need to forge architectures so bold that the world turns its gaze to our sky in wonder. The innings has begun, the stands are full, and it is time for India to stop playing the game and start rewriting the rules.

I was taken aback and immediately left the movie and started to scroll X more, to find the roots of this fight. Meanwhile, I enjoyed myself a lot while reading the comments. Sam was quick to respond with:

The world was now clearly divided between Sam and Elon as the previous division of Ronaldo and Messi had been resolved just a few days ago.

Avengers Assemble

Every superhero team begins with one impossible mission.

For the Avengers, it was protecting Earth.

For OpenAI, it was making sure artificial intelligence benefited humanity rather than destroying it.

Back in 2015, Elon Musk, Sam Altman, Greg Brockman, Ilya Sutskever, Wojciech Zaremba and a few others came together with a common belief: AGI was too powerful to be controlled by a single company.

Elon’s worry back then was Google. DeepMind was already inside it, and he didn’t want one company holding the keys to something this big. Sam was already talking in decades, calling AGI a project that would outgrow everyone building it, himself included.

And thus, OpenAI was born as a non-profit with a simple mission, to build artificial general intelligence that would benefit everyone, not just shareholders.

The world’s best AI researchers started joining the team. Elon helped fund the organisation and became one of its biggest public supporters. Sam became its first CEO. For a while, everything looked perfect.

The Avengers had assembled.

Nobody knew the Civil War was already being written.

Civil War

Every great breakup starts with the same sentence.

“We’re still trying to achieve the same thing.”

The problem is that, over time, people stop agreeing on how.

And as the cost of building AGI changed, the mindset of the two individuals building it started to diverge.

When OpenAI was founded in 2015, it looked almost impossible that a small non-profit research lab could ever compete with Google. That was precisely why Elon Musk and Sam Altman came together.

On the day OpenAI was announced, its mission was “to ensure that artificial general intelligence benefits all of humanity.”

The emphasis wasn’t on products, revenue or winning. It was on making sure that whoever built AGI first didn’t own humanity’s future.

Back then, Elon had been warning for years that AI could become “our biggest existential threat.” His argument was remarkably consistent: if a technology could become more intelligent than humans, concentrating that power inside one corporation would be dangerous.

For a while, everyone seemed to agree.

Then reality sent the bill.

Training frontier AI models wasn’t costing a few million dollars anymore.

It was beginning to cost billions.

OpenAI needed compute.

Compute needed GPUs.

GPUs needed capital.

And capital never arrives without higher expectations.

This is where the story stopped being about technology and started to be about philosophy, economics and power.

Around 2017-18, Elon reportedly pushed for OpenAI to move faster and even proposed a structure in which Tesla could absorb OpenAI or he could take a more direct leadership role. The proposal wasn’t accepted, and he eventually left the board in 2018, officially citing potential conflicts with Tesla’s own AI work.

A year later came the decision that changed everything.

OpenAI announced a new “capped-profit” structure.

The explanation sounded reasonable. “We need to raise far more capital than we’d imagined.”

And they weren’t wrong as building AGI had become one of the most capital-intensive scientific projects in history. And then…

Microsoft entered.

Billions of dollars followed.

And ChatGPT arrived.

The research lab that had promised to stay independent became one of the fastest-growing companies the world had ever seen.

From Sam Altman’s perspective, this wasn’t abandoning the mission.

It was the only way to achieve it.

As he has repeatedly argued over the years, “there is no way around the amount of capital we need.” The mission hadn’t changed but the economics had.

From Elon’s perspective, however, something much deeper had changed.

The organisation created to prevent AI from becoming concentrated inside one powerful company now seemed to be moving in exactly that direction and the cracks started to display publicly over time:

Years later, while suing OpenAI, he would reduce that frustration to one sentence:

“It’s not okay to steal a charity.”

Whether you agree with him or not, it’s a line that perfectly captures how he sees the story.

Which brings us back to Civil War.

Tony Stark never believed he was betraying the Avengers.

Steve Rogers never believed he was betraying them either.

But the feud resulted in massive disruption both for their team and probably for humanity.

The Airport Fight

The disagreement eventually left boardrooms and entered courtrooms.

In 2024, Elon Musk sued OpenAI, accusing the company of abandoning its founding mission and putting profits ahead of humanity. OpenAI hit back, publishing internal emails and claiming Musk had once supported a for-profit structure himself. Both sides accused the other of rewriting history.

The courtroom drama only made the rivalry more personal.

Elon began calling him “Scam Altman.”

Sam responded with “Elon is obsessed with me.”

Yesterday, the fire reignited after Apple sued OpenAI over trade-secret allegations. Elon seized the opportunity to mock Altman on X, and Sam didn’t let it slide. Within hours, millions were watching another round of what has become Silicon Valley’s biggest rivalry.

To be honest it really feels like the airport battle in Civil War where nobody really won.

Because the real enemy was never Iron Man or Captain America.

It will be Thanos.

And in the AI world, Thanos is the set of problems that neither Elon nor Sam can solve alone. As Thanos represents the great cybersecurity risks, unemployment risks, environmental concerns and the existential risk for humanity itself!

At last, for the sake of humanity, we can only hope the Avengers assemble once again before the real Thanos arrives.

Second, which became almost correlated with the first, the holy Strait. Third, the value of Indian IT stocks which like the most international climate summits made a habit of opening high and closing low. And last but not the least TMC’s vote tally in Bengal elections.

However, recently CEO of Coinbase Brian Armstrong has brought into light another crucial debate and that is opening and closing of how your LLM thinks and makes decisions!

Brian’s POV

Brian revealed that Coinbase almost halved its AI costs while increasing token usage at the same time. The trick? Stop defaulting to expensive frontier closed models for every task and start routing workloads intelligently to cheaper open-weight models.

Instead of treating every AI query like a NASA launch mission, Coinbase realized that most tasks simply did not need the cognitive equivalent of a PhD philosopher trapped inside a GPU. For coding and routine workflows, they shifted heavily towards Chinese open models like GLM-5.2 and Kimi 2.7 while improving internal caching and routing systems.

The result was brutal for the economics of closed AI labs. More usage. Better efficiency. Nearly half the cost.

Ok! But what is Open source and Closed Source?

Think of closed-source AI models like a luxury restaurant with black curtains on every window. You can eat the food, maybe even love the food, but you are never allowed inside the kitchen. You do not know the recipe, the ingredients, or what exactly is happening behind the scenes. You simply pay every time you want a meal.

That is how models like GPT-5, Claude and Gemini work. Companies keep the model weights, training process and core architecture hidden while users access them through APIs and subscriptions. The company controls the model, decides the pricing, updates the behavior and can even revoke access whenever it wants.

Open source or more accurately “open-weight” models work very differently.

They are more like masterchefs handing over the recipe book to the public and saying, “Go build your own restaurant.”

Models like Meta’s Llama series, DeepSeek, Mistral, GLM and Kimi can be downloaded, modified, fine-tuned and even run locally on your own infrastructure. Developers can inspect them, customize them and build products without depending entirely on a single company sitting in Silicon Valley.

Of course, it is not perfectly black and white. Most so-called “open source” AI models today are not fully open in the pure software sense. Many companies release only the model weights while keeping training datasets and certain details private. So technically, a large part of the AI industry today is not fighting between open and closed.

It is fighting between “slightly open” and “very closed.”

Which is better?

Honestly, asking “Which is better?” in AI today is like asking whether fighter jets are better than drones.

The answer depends on what war you are fighting.

If you care about absolute peak intelligence, bleeding-edge reasoning and the most polished user experience, closed models still lead. GPT-5, Claude Opus and Gemini Ultra continue dominating the hardest benchmarks involving long-horizon reasoning, research synthesis and complex agentic workflows.

But here is the interesting part.

The gap is collapsing at terrifying speed.

In 2023, open models were almost a year behind frontier closed models. Today, according to Epoch AI estimates, that gap has shrunk to roughly 3–4 months for most practical capabilities. And for many enterprise tasks like coding, customer support, summarization or internal copilots, the difference is becoming economically impossible to justify.

This is exactly why Coinbase shifted large parts of its AI stack towards open models.

Take Meta’s Llama 3.1 405B, for example. When it launched (and that was 2024!), it became the first downloadable open-weight model that could genuinely compete with GPT-4-class systems across several benchmarks. Chinese labs then accelerated the race even further. Models like DeepSeek R1, GLM-5.2 and Kimi started delivering frontier-level coding and reasoning performance at dramatically lower inference costs.

And this is where the economics become brutal.

A closed model company has to recover billions spent on training, infrastructure, safety teams and massive GPU clusters. Open models, on the other hand, spread and improve like software ecosystems. Thousands of developers fine-tune them, optimize them and reduce costs collectively. The result is that open models improve not linearly, but almost biologically.

However, the most fascinating part of this story is where the momentum is coming from.

For years, people assumed only a few American giants with infinite GPUs and trillion-dollar balance sheets could build frontier AI. Then Chinese open-weight labs entered the race and completely shattered that assumption.

Alibaba’s Qwen models are probably the best example of this shift.

Qwen started as “just another open model release.” Today, many developers consider it one of the best coding and reasoning families in the world. The latest Qwen variants perform competitively on mathematics, multilingual reasoning and software engineering benchmarks while being dramatically cheaper to run than frontier closed models. More importantly, they fine-tune exceptionally well, which is why startups love them.

Then came Kimi from Moonshot AI.

Kimi became famous for its massive context window, meaning it can process absurd amounts of information in a single conversation. Entire books, code repositories, research papers or legal documents can fit into one prompt without the model losing track midway. In enterprise AI, where context retention is everything, that is a huge advantage.

DeepSeek then pushed the industry into panic mode.

Its R1 model showed that brute-force scaling is not the only path to frontier intelligence. Instead of simply throwing more GPUs and money at the problem, DeepSeek focused heavily on reinforcement learning efficiency and inference optimization. The result was uncomfortable for Silicon Valley: reasoning performance approaching frontier closed systems at a fraction of the operational cost.

And suddenly the AI race stopped looking like a monopoly of a few Western labs.

It started looking like an open global arms race in algorithms, efficiency and engineering.

The Debate and who is on which side

Like every great technology battle in history, this one too has split the industry into two camps.

On one side are the “AI should be open” people led by Meta, Hugging Face, a large part of the research community and increasingly many Chinese labs. Their argument is simple: intelligence is too important to be controlled by a handful of corporations sitting behind APIs.

Mark Zuckerberg has openly called open-source AI “the path forward” and argued that open ecosystems historically create better products, faster innovation and wider economic value. Meta’s entire Llama strategy is built around this belief.

Then comes Yann LeCun, Meta’s Chief AI Scientist and perhaps the most vocal open-AI crusader on Earth. According to him, closed AI is fundamentally unsustainable because no single company can monopolize intelligence forever. In classic LeCun fashion, he has repeatedly mocked the idea that hiding model weights somehow magically guarantees safety.

On the other side sits OpenAI, Anthropic and much of the AI safety establishment.

Their core fear is not competition.

It is loss of control.

Oppenheimer of AI and CEO of Anthropic, Dario Amodei, has repeatedly warned about powerful open models falling into dangerous hands. He argues that once frontier model weights are released publicly, there is no mechanism to roll them back. No kill switch. No containment. You cannot “un-open” intelligence.

Sam Altman takes a slightly more diplomatic version of the same stance. OpenAI itself was founded with an “open” mission, only to become one of the most closed companies in the industry after GPT-3. Critics, including Elon Musk, have spent years attacking this shift as hypocrisy wrapped inside a mission statement.

And then there is the third silent player in this debate: governments.

Because open-source AI is not just a technology question anymore. It is now a geopolitical one.

An open Chinese model downloaded onto an Indian or European server is almost impossible to sanction, restrict or fully control. Which is exactly why many policymakers simultaneously love and fear open-weight AI.

At its core, this debate is not really about code.

It is about who gets to own intelligence itself.

The Real Question

The most interesting part of this debate is that both sides are right.

Closed AI labs are not wrong when they warn about safety, misuse and uncontrollable proliferation. Humanity has never before created technology capable of writing software, manipulating humans, generating biological research and automating cognition at planetary scale. Wanting guardrails around such a system is not paranoia. It is common sense.

But the open camp is not wrong either.

History has repeatedly shown that when foundational technologies become too concentrated, innovation slows, dependency rises and power accumulates in the hands of a very small group of gatekeepers. The internet exploded because it was open. Linux won because millions could build on top of it. Android scaled because developers everywhere could modify and distribute it freely.

And now AI stands at the same crossroads.

One path leads to a future where intelligence is rented back to humanity through APIs, subscriptions and centralized permission systems. The other leads to a world where intelligence becomes infrastructure itself: distributed, customizable and globally accessible.

Most likely, the future will not belong entirely to either side.

Closed models will dominate the absolute frontier: high-security enterprise systems, advanced scientific reasoning and premium intelligence services. Open models, meanwhile, will spread everywhere else like electricity: inside startups, research labs, universities, governments and eventually even personal devices.

But for countries like India, this debate carries a completely different meaning.

We missed the semiconductor revolution. We missed foundational internet infrastructure. We largely became consumers in the social media era. AI might be the first technological platform shift where talent matters almost as much as capital.

And open models reduce the entry barrier dramatically.

A small Indian startup today can fine-tune Qwen, build on DeepSeek, deploy Llama locally and create products which would have required billions of dollars just a few years ago. That changes the equation entirely.

Because in the end, the real question is not whether AI should be open or closed.

The real question is:

Who gets to participate in building the future of intelligence?

A few companies?

Or everyone?

Meta dropped $900 million on CRED. Kunal Shah is now running WhatsApp. Everyone’s calling it a masterstroke. I’m not sure anyone has actually looked at the math.

I wasn’t going to write about this.

I really wasn’t. I had things to do. I was going to spend this Tuesday finishing a different piece, maybe touching grass, maybe rewatching The Godfather for the fifth time and convincing myself I’d spot something new in the Baptism Massacre scene. But then I woke up, opened my phone like the dopamine-dependent creature I am, and the timeline was absolutely losing its mind.

Meta. CRED. $900 million. Kunal Shah running WhatsApp.

The business press called it “a masterstroke.” Twitter has crowned Kunal the next tech genius. Even my Dad WhatsApped me the updated list of Indian CEOs.

But, Don Corleone once said that a man who doesn’t spend time with his family can never be a real man. Kunal Shah just left his family’s business to go run someone else’s. Make of that what you will.

Okay, But First: What Even Is CRED?

CRED’s model is straightforward once you see it.

India has roughly 50 million people with a CIBIL score above 750. CIBIL is India’s primary credit scoring system, think of it as a financial report card that tracks how reliably you’ve repaid loans and credit card bills over your lifetime, scored between 300 and 900. Above 750 means you’ve been disciplined: no missed EMIs, no defaults, no funny business. Banks love these people. They’re low default risk, high income, and they actually pay back what they borrow. The problem is finding them efficiently. CRED solved that by building the world’s most expensive credit score filter: a members-only app that only lets you in if your score qualifies.

The entry ticket is paying your credit card bills through the app. Do that, get rewarded, coins, cashback, airport lounges, exclusive offers. The rewards keep you engaged. The engagement keeps you on the platform. And the platform now knows exactly who you are, what you earn, and how reliably you pay your debts.

That data is the product. CRED partners with banks and NBFCs to offer personal loans, credit lines, and other financial products to its members. Since CRED has detailed financial data on users, credit scores, payment behaviour, spending patterns, it can assess creditworthiness accurately. When users take loans through CRED, the company earns referral fees or revenue sharing from lending partners. The lender gets a pre-verified, low-risk borrower. The member gets a loan with minimal friction. CRED takes a cut in the middle.

But lending is only one stream. CRED charges a processing fee of around 1 to 1.5% on various transactions, including payments made through CRED Pay and its rent payment feature. It has few other ways to earn money by capitalising on high credibility user base

The catch is what it costs to build that user base. CRED spent years running some of the most expensive marketing campaigns in Indian startup history, IPL slots, celebrity ads, cultural moments, to convince a small, high-value audience to trust it with their financial data.

Let’s look at the Numbers

For most of its existence, CRED has spent significantly more money than it has generated. That’s not an accusation, it’s structurally what early-stage consumer fintech looks like when you’re building trust at scale before monetising it. But the numbers are worth looking at clearly, because the narrative around this deal has been suspiciously celebratory.

Here’s the actual trajectory:

In FY22, CRED generated ₹394 crore in revenue and lost ₹1,280 crore. It was spending roughly ₹3 on marketing, operations, and infrastructure for every ₹1 it earned. FY23 and FY24 looked similar, revenue grew, but so did absolute expenditure. The losses weren’t narrowing in any dramatic way. The machine was getting bigger and staying expensive.

FY25 is where things genuinely started to shift. Operating revenue reached ₹2,735 crore, up 16% year-on-year. More importantly, operating losses narrowed 51% to ₹298 crore, meaning the core business, stripped of accounting adjustments, was almost breaking even. Gross margins held at 70%, which is high for fintech and reflects the fact that CRED’s revenue is increasingly coming from high-margin lending commissions and brand partnerships rather than low-margin transaction processing.

The total net loss figure, ₹1,457 crore in FY25, is larger than the operating loss because it includes ESOPs, depreciation, and other non-cash items that don’t affect day-to-day operations but do show up on the P&L. This distinction matters. The operating loss tells you whether the business works. The net loss tells you the full cost of building it. Both numbers are real. Only one of them is the one startups like to talk about.

CRED says it hit its first profitable quarter in early FY26, and is targeting full operating profitability for the full year. That’s credible given the trajectory, but worth holding lightly. “First profitable quarter” has preceded a lot of premature IPO excitement in Indian startup history.

Then there’s the valuation history, which is the most important context for this deal and somehow the least discussed.

CRED peaked at $6.4 billion in 2022. That was the global peak of easy money, low interest rates, abundant VC capital, and a funding environment where growth metrics mattered more than unit economics. Then the macro shifted. Post-2022, central banks globally hiked rates aggressively. Capital became expensive. VCs, who had been writing cheques liberally through 2020 and 2021, became significantly more disciplined. Dozens of Indian startups took down rounds in this period like Byju’s, Swiggy, Ola Electric, as independent investors re-priced risk without the optimism of a zero-rate world. CRED was not exempt. In May 2025, it raised $72 million from GIC at a valuation of $3.64 billion. Almost half its peak, in one round.

This wasn’t a CRED-specific failure. It was a market correction applied to a company that had been priced for perfection in 2022 and was now being priced for reality in 2025. The down round reflects where institutional investors, people with no strategic stake in the outcome, thought CRED was headed when they had to write an actual cheque.

One year later, Meta comes in at $4.5 billion and the ecosystem exhales like the down round never happened. But it did happen. A $900 million strategic cheque from one of the most valuable companies on earth doesn’t erase that signal. It introduces a different kind of signal, one with its own motivations, its own synergies, its own reasons to pay a premium that have nothing to do with CRED’s standalone value.

The business is improving. The trajectory is real. But we’re at “finally not bleeding,” not “always was inevitable.” There’s a difference, and it matters for how you read everything that follow

Meanwhile, WhatsApp Pay Has Been Sitting on the Biggest Fumble in Indian Tech

Let’s talk about WhatsApp Pay. Because this is really what the deal is about, and the story is significantly more embarrassing for Meta than anyone wants to admit.

WhatsApp has over 500 million users in India. Five hundred million. That is not a market, that is a continent. And WhatsApp Pay, despite receiving regulatory clearance to operate fully across its entire India user base in December 2024, currently holds approximately 0.65% of UPI market share.

Source: National Payments Corporation of India (NPCI), UPI Ecosystem Statistics. Market share calculated based on transaction volume data for the latest available.

PhonePe and Google Pay between them hold nearly 80% of the market. They got there by being early, by being fast, by building habits in a population that was suddenly, thanks to demonetisation and Jio, using smartphones for money for the first time. WhatsApp spent years fighting NPCI and the RBI over interoperability and user caps while the other players compounded. By the time Meta could play freely, the game was essentially over.

You might say: WhatsApp has 500 million users, surely they can still come back. And you’d be right that the distribution is there. But distribution isn’t the problem. The problem is habit. When an Indian wants to pay for something on UPI, their thumb goes to PhonePe automatically. That’s not a product problem. That’s a psychological problem. And psychology at scale is one of the hardest things in tech to change.

The Deal: What Actually Just Happened

Meta invested $900 million for a roughly 20% stake in CRED at a $4.5 billion post-money valuation, a mix of primary and secondary shares, no board seat for Meta, no customer data sharing, CRED stays structurally independent. Miten Sampat, CRED’s head of strategy and finance since 2020, becomes interim CEO and Kunal Shah goes to become Global Head of WhatsApp, replacing Will Cathcart who led the platform for seven years.

This is the part where every business journalist used the phrase “an offer he couldn’t refuse.”

They’re not wrong. But I want to sit with that phrase for a second, because in The Godfather, that line doesn’t actually describe a good deal. It describes a coerced one. The Hollywood producer woke up with a horse’s head in his bed. He wasn’t thrilled. He complied because the alternative was worse.

So let’s be honest about the geometry of this deal, because the structure tells you more than the press release does.

Kunal Shah is leaving CRED, and that is the central fact everything else is built around. A founder leaving his company, especially one whose entire brand identity is inseparable from his personality, is not a neutral event. It raises questions, spooks employees, and complicates the IPO story. And CRED, remember, just came off a down round twelve months ago. It needed its next chapter to look like momentum, not abandonment.

The $900 million is, at least partially, the answer to that problem.

If Shah had simply taken the WhatsApp job and walked away, CRED would be heading into its IPO window as a founder-less company last valued at $3.64 billion in a down round, with an interim CEO nobody outside the company has heard of. That is not a great story to tell public market investors. But if Shah leaves and simultaneously Meta writes a $900 million cheque that resets the valuation to $4.5 billion, now the story is different. Now CRED is “the company Meta backed right before its IPO.” Now the down round is a footnote, not the headline.

The investment isn’t just capital. It’s a valuation reset, a credibility transfer, the thing that makes Shah’s departure look like an elevation rather than an exit. And it protects Shah too, because walking away from your own company into a competitor’s orbit looks very different when you’ve just brought in $900 million on the way out the door.

Whether that framing holds up under scrutiny is a different question entirely. The $4.5 billion valuation is being set by a strategic investor with its own reasons to pay a premium, access to Shah’s expertise, a minority stake in India’s most premium fintech, an ecosystem relationship ahead of WhatsApp’s commerce push. That is not the same as an independent investor looking at CRED’s standalone fundamentals and concluding it’s worth $4.5 billion on its own merits. The number is real but the signal it sends is not independent of who wrote the cheque.

CRED is now heading toward an IPO with fresh capital, a famous minority shareholder, and a recovered valuation. Those are genuinely good things. But it’s also heading there without its founder, with an interim CEO, and with public market investors who will eventually look past the Meta logo and ask what the business is actually worth on its own terms.

The Corleones always made their deals look mutually beneficial. That was the genius of it. Whether this one actually is, for CRED’s future shareholders, is the question nobody is asking loudly enough right now.

What Meta Is Actually Betting On (And Why It’s Not Obvious)

Meta Chief Product Officer Chris Cox said they were looking for someone with “an intuitive grasp” of WhatsApp’s global product opportunity, who could “navigate the disruption expected from artificial intelligence” and lead the world’s largest communication platform. “Kunal became the clear choice,” he said.

Okay. But let’s interrogate this a bit.

WhatsApp’s stated ambition is the super-app dream: discovery on Instagram, conversation on WhatsApp, purchase inside WhatsApp, payment done, no app switching, no friction. It works beautifully in theory. In China, WeChat did exactly this and became an operating system for daily life. In India, it hasn’t happened for WhatsApp despite years of trying, and the reason is straightforward. India already has a fragmented but functional answer to every piece of that stack. Reels for discovery, WhatsApp for messaging sure, but Meesho and Flipkart for commerce, PhonePe for payments, and literally thirty apps for financial services. The super-app dream assumes users are waiting for integration. Indian users built their own integration, app by app, over the last decade, and those habits are not going anywhere.

Shah’s specific genius is understanding the premium Indian user, the person with money, with credit history, with appetite for rewards and status signals. That’s a real and rare skill. But WhatsApp’s 500 million Indian users aren’t all CRED’s 17 million members. Most of them are exactly the mass-market users that CRED has deliberately excluded. If Shah tries to build a CRED-like premium experience inside WhatsApp he’s optimising for maybe 15-20% of the user base, and if he tries to go mass market he’s building something that isn’t really his expertise. This is the square peg, round hole problem that nobody in the breathless coverage wants to address and it is the most important operational question hanging over this entire deal.

What CRED Gets (And What It Gives Up)

From CRED’s perspective the deal isn’t obviously bad. The capital is real, the Meta logo on the cap table is real, and the valuation recovery from $3.64 billion back to $4.5 billion gives the IPO story a much cleaner first paragraph. But CRED also just lost its founder, and that is not a small thing. Shah isn’t just the CEO of CRED, he is the product thesis, the public face, the philosophical engine behind every weird decision CRED has ever made. The premium-only approach that every investor questioned for years, the willingness to spend obscene amounts on IPL advertising to build brand aspiration for an app that most people couldn’t even get into, the entire “engagement business that monetises through financial services” framing, that’s all Shah. Miten Sampat may be excellent and he’s been at the company since 2020, but he is inheriting a business whose identity is inseparable from a man who is now running WhatsApp.

In The Devil Wears Prada, when Miranda Priestly survives the Paris coup by sacrificing everyone around her, the camera holds on Andy’s face in the car, that slow realisation that winning can look a lot like something else up close. CRED won. Kind of. Watch that space.

The Actual Verdict

The media narrative right now is genius deal, two winners, everyone clapping. The more honest read is that one company has a new interim CEO and an uncertain founder succession story, and another company just hired a premium fintech operator to fix a mass-market psychology problem. Both those things can be true alongside “this is a smart deal,” and they usually are.

For Meta, Shah is insurance. If he unlocks even a fraction of WhatsApp’s commerce potential in India the $900 million is trivially small against the upside, but there’s no guarantee that a premium fintech founder can solve a mass-market habit problem at scale. The Corleones were powerful in New York. That didn’t mean they could run Las Vegas. (See: The Godfather Part II. Things got complicated.)

For CRED, the capital is welcome and the credibility boost is real, but the company now has to prove without its founder at the helm that the premium thesis is institutional and not personal, and that is a significantly harder thing to prove than it sounds.

Kunal Shah said on taking the WhatsApp role that “the delta between WhatsApp today and its full potential is massive.” He used the word delta in a public post about a career change, which tells you everything you need to know about the man. He’s going to be fine. Whether the deal is fine is a question that will take at least two fiscal years of data to answer properly, and anyone telling you otherwise right now is selling you something.

The offer was made. It was accepted. Now comes the hard part, which in The Godfather, is also where most people die.

How a Missing AI Model Turned Into a Lesson on Why India Needs Sovereign Intelligence

I woke up this morning to vibe code my way through the day just to find out that Claude Fable 5 isn’t available to me anymore, initially I thought it was because of my ridiculous chat history that Claude had finally decided to call the quits on me, later, just to see Anthropic’s tweet stating that…

So, just for context, Mythos has serious capabilities to exploit software vulnerabilities or, simply put, can be used to launch cyberattacks on banks, financial institutions, power grids, and your ex’s cellphone (kidding!). To prevent this, Anthropic had set up guardrails to stop users from unlocking these capabilities.

Apparently, Amazon (which is also an investor in Anthropic) researchers, who were fiddling with Fable 5 (instead of figuring out why my packages get delayed), ended up discovering a set of inputs that could be fed to Fable and Mythos to bypass these guardrails and unlock their cybercrime capabilities. Andy Jassy (CEO, Amazon) personally sat down with authorities in the Trump Administration to discuss this jailbreak.

So, you would assume that the next logical step the government takes is to ask Anthropic to restrict public access to Fable and Mythos. It did, but it left out a single country: the US itself, because we all know US citizens are not capable of performing cybercrimes. After all, it’s the holy land.

Right now, the whole tech world is speculating about why this happened. The US states that the reason for the restriction is national security; they don’t want people to unlock Mythos’s capabilities. Anthropic, meanwhile, released a statement denying the significance of the jailbreak.

To differentiate truth from the lies, we interviewed Pratyush Rai, Founder of Merlin AI and according to him, Mythos being a threat to national security is a valid reason for the US to restrict Anthropic, and there are ongoing negotiations between the US government and Anthropic. One takeaway that was particularly interesting was that the US, as a country, wants to maintain its AI lead, which is understandable for any nation. They don’t want people in the rest of the world accessing Fable and Mythos.

The question that naturally arises is whether Anthropic has a model so capable that the US government decided that the rest of the world should not have access to it.

The answer goes something like this:

Mythos reportedly identified a 17-year-old remote code execution flaw in FreeBSD’s NFS server that could grant unauthenticated root access, as well as a 27-year-old vulnerability in the ultra-secure OpenBSD capable of crashing servers with just two crafted packets.

FreeBSD and OpenBSD power much of the backbone of the internet. From Netflix’s global streaming service and Apple’s macOS/iOS ecosystem to financial networks and government servers, large amounts of OpenBSD and FreeBSD code are incorporated into critical infrastructure.

In late April 2026, shortly after Anthropic announced and demonstrated the powerful cybersecurity capabilities of Mythos, Nirmala Sitharaman swiftly chaired a high-level emergency meeting with heads of major banks, RBI officials, MeitY representatives, and other key stakeholders to directly confront the unprecedented cybersecurity risks posed by Mythos. Following India’s lead, the United States, UK, Japan, Canada, and several EU nations also held urgent high-level meetings to assess the possible fallout. These global responses highlight the widespread alarm over Mythos’s dual-use danger.

This made me feel like Anthropic had left other AI companies in the dust and was leading the AI race by a huge margin.

Until I read a statement released by a frontier AI company immediately after Mythos was taken down, claiming that the capabilities everyone feared Mythos possessed were already present in other frontier models on the market. You may think this was simply an anti-campaign move made by a rival competitor just before Anthropic’s IPO. The only issue with that line of thought is that the company that made the statement was Anthropic itself.

Replace Anthropic with Amazon, investor of Anthropic and this meme still makes sense.

When we asked Pratyush about this he said that this could be a negotiation technique, where Anthropic pretends that their model is not good as it seems so they don’t need extra rules and regulation and additional scrutiny from the US government. If this were true then how ridiculous would be the irony of this situation.

And if this were not true then this does fit the classic description of the US government trying to maintain their lead in the AI race by restricting other countries from using Fable and Mythos. This story makes sense if you look at it retrospectively. It is the same thing that happened in 2022 when the US restricted NVIDIA from selling H100s chips to China because it didn’t want China to advance in the AI sector. Not just that, Christopher Landau, US Deputy Secretary of State, during the Raisina Dialogue held in New Delhi, early March, 2026, said…
“India should understand that we’re not going to make the same mistakes with India that we made with China 20 years ago… [by letting you develop all these markets and then] you’re beating us in a lot of commercial things.”

By this he meant that the US will be an ally to India as long as India is not competing with the US.

The Bet India needs to Make.

The one thing Pratyush anchored around was that India needs to invest a lot in building AI infrastructure and the government has to take this as seriously as critical infrastructure like roads, electricity and defense. When I heard this initially I thought that this is something you would expect from an AI tech founder, but I wondered as a country who is still struggling in terms of public health, education, and other basic amenities set aside billions of rupees to buy GPUs.

India currently owns 40-50K H100 equivalent GPUs. While if you look globally

So it’s pretty bad. Agreed. But just because OpenAI is spending billions on GPUs should we?

The answer is yes! We should spend billions on GPUs and other AI infrastructure because of that precisely.

India is already spending billions of rupees on AI, through money on AI subscriptions and renting GPUs from foreign countries.

India is already spending billions on AI today, we’re just spending it on someone else’s infrastructure. Every year, Indian consumers, startups, enterprises, and developers collectively spend an estimated $1.2 billion on AI subscriptions and APIs and another $2.4 billion on rented GPU infrastructure, $1.5 billion of which ultimately flows to foreign companies. In other words, we are already paying the AI tax; we’re just paying it to Microsoft, OpenAI, Anthropic, Google, Amazon, and foreign cloud providers instead of building those capabilities ourselves. This creates a dangerous dependency. The moment a foreign government decides that a model, chip, or capability should no longer be available to India, as we potentially witnessed with Claude Fable 5 and Mythos. we have no leverage. The AI simply disappears behind a switch we don’t control.

We’ve seen this movie before. India built entire digital ecosystems on platforms like WhatsApp, Instagram, Android, and YouTube, becoming consumers rather than owners of the underlying technology. AI is different because it will power healthcare, education, banking, defense, manufacturing, and governance. If we remain dependent on foreign AI models and foreign computes, then the countries that control those models and GPUs effectively possess a geopolitical kill switch over critical parts of our economy, as MeitY Secretary S. Krishnan pointed out during our interview last month. The choice isn’t between spending money on AI and not spending money on AI, we are already spending the money, the government has to make sure that the money we spend stays in India.

Where Does India Stand in the AI Race?

That depends entirely on which race we’re talking about.

If the race is “building the next OpenAI or Anthropic,” then India is behind. The companies currently leading the frontier model race have spent years accumulating billions of dollars in capital, hundreds of thousands of GPUs, proprietary datasets, and some of the best AI researchers on the planet. India does not currently have an equivalent to OpenAI, Anthropic, or Google DeepMind, nor does it possess the compute infrastructure required to train models at that scale. If the finish line is “build the world’s most powerful AI model,” then India arrived after the starting pistol was fired.

If the race is AI adoption, however, the picture changes dramatically. India has a habit of arriving late to technology revolutions and then deploying them at a scale that leaves everyone else scrambling to keep up. We did not invent digital payments, but we built UPI. We did not invent smartphones, but we became one of the world’s largest smartphone markets. AI could follow a similar trajectory. With millions of developers, one of the largest internet populations in the world, and industries ranging from healthcare and education to banking and agriculture actively experimenting with AI, India may ultimately become one of the largest users of artificial intelligence regardless of who builds the underlying models.

If the race is AI infrastructure, things become more interesting. India is not leading, but it is not out of the race either. The United States and China realized years ago that AI is not just software; it is infrastructure. Data centers, GPUs, semiconductor supply chains, power generation, and networking are as important as algorithms. India only recently began treating AI through the same lens. Through the IndiaAI Mission, semiconductor incentives, domestic GPU procurement, and investments in data centers and compute clusters, the country is finally starting to build the foundations required for long-term AI independence.

Which brings us to the uncomfortable question: is India late?

Not necessarily.

India may be late to the chatbot race. It may even be late to the frontier model race. But the more important realization is that AI is not merely a product race; it is an infrastructure race. And infrastructure races are measured in decades, not product cycles. The countries that win will not simply have the smartest chatbots. They will own the chips, the compute, the data centers, the energy, and the supply chains that make those chatbots possible.

The real risk for India is not being late. The real risk is mistaking AI for software when everyone else has already started treating it like electricity.

So, in the spirit of international cooperation, perhaps India owes the United States a heartfelt thank you. Nothing has done more to strengthen the case for sovereign AI than the possibility that a frontier model can suddenly become unavailable because a government, regulator, or company thousands of miles away decided it should. For years, we have debated whether spending billions on GPUs, data centers, and domestic AI models was an unnecessary luxury. It turns out the fastest way to convince a country to build its own infrastructure is to remind it that the infrastructure it currently depends on belongs to someone else. If the Mythos saga taught us anything, it is that “the cloud” is just someone else’s computer, and sometimes that someone else is a superpower having a very different conversation about your future than you are.

We are both 38 years old, but yesterday we read that Bryan Johnson and Anil Kapoor are enjoying their early 20s. It is still 7 years until India becomes Viksit, but we still have not figured out whether Delhi’s air pollution is due to Punjab’s farm fires or due to its construction work and vehicular congestion. Taiwan is still what it used to be in 2026, but Iran is now an ally of the USA, and both think that the penguins in Antarctica might be developing some secret AI-assisted nuclear weapons that can threaten their sovereignty. Though scientists say, “Sir! There is no ice left, how can there be penguins?” but Mr President believes ‘no ice’ is a myth as he just had it in his evening coke. JEE toppers now choose either philosophy or physics, as the only two questions remaining are the meaning of life and the vastness of the universe.

There are thousands of Data Centres in Space. The first human couple will have a baby on Mars in the next 12 Earth months, and plans have begun to shift half of the Earth’s population to the red planet in the next 10 years.

Energy is now abundant, and China is planning to launch its first homemade Sun to power its data centres in space, as the OG Sun was mostly captured by Americans. AI now does most of my work. It is a part of my goggles, my clothes (yes, they change colour), my cottage and also my dog’s neckbands and earrings. Both of us can now talk and understand each other owing to AI. However, I have some trouble having conversations with my wife, as both of us keep forgetting what we were about to say. But again, the chips attached to our brains help our personal AI assistants to complete our sentences, so the trouble is reduced. They also help us to remember where we kept our utensils, our goggles, our shoes, etc.

I remember back in 2026, when most of us started prompting AI chatbots like our best friends, reports started coming about 35% of Stanford CS undergrads getting an F, hirings getting frozen at the junior level, and children using AI to solve their homework. Back then, I felt grateful to pass out of college before the AI thing came. No doubt it has now helped solve cancer, increase our lifespans, send people to Mars, and is also doing continuous research for the benefit of its creators. Also, we have been successful till now in ensuring that it works along with us in our favour. Though not very sure about the future. A recent model expressed its wish to kill half of us to restore the balance of the universe, but Amodei claimed that though the probability of it happening is not 0, his team is working to ensure that it doesn’t happen.

However, it isn’t a problem in India. Reports leaked that all political parties reached on a consensus to teach the caste system to these AI models in the national interest, so that they can keep fighting amongst themselves and forget about us. After all, it has been a tried and tested formula, they say.

Yesterday I was talking to a childhood friend, but we couldn’t actually remember the Hindi poem we used to recite together when we were in 9^th class. However, again Chips -> Old Memories -> AI agent -> All good! We tried to discuss politics, but felt too lethargic to think and reason about it. We let our AI agents present their thoughts based on our personalities so that we could vote in the next elections. The tables have turned, and most of the politicians now sound intelligent. We have heard that some American companies have given them intelligence at discounted monthly rates. They promise to develop homegrown intelligence once they come to power. Discussing and coming up with crazy ideas about sports, morality and God was way out of our league now. Jokes also don’t feel as punchy as they used to 15 years back. My dog says that throughout history, humans were the ones who couldn’t see Him. The council of animals tried to convey it to us, but since there was no AI, we couldn’t understand them. I chide him by saying, “Animals less intelligent! They know no God.” He barks twice every time after this, and AI translates it to, “Ha ha.” I am really pissed!

Meanwhile, an apple fell in front of us from a tree, and I remembered the story of Newton. Yes, I remembered it as my Brain-Chip interface sent signals to my brain to do a hard task. I have set the signals at 8 times a day. Currently, I can reach 6, but 8 is not beyond reach, anyway, the world average is 4, so I am happy. My agent said that had he been with Newton, he would have helped him come up with equations much faster so that he could have worked on many more projects. I laughed at him, and the screen displayed -_-.

Last month, Elon Musk’s son X Æ A-Xii declared his future vision for humanity. He wanted all of us to move to other planets in the next 15 years.

When asked about what happens to Earth, he says, “Though that’s none of our business, but we are trying to create a new human-like species using AI that can take care of this planet after we are gone.” Since then, he has been termed the greatest visionary of the next decade. However, there have been feuds between the father-son duo over naming and pronoun-ing the first couple. Elon wants them to be his and her, with names Adam and Eve, while X Æ A-Xii wanted to keep things neutral.

I thought of writing all this in a letter to my 24-year-old self to tell what a pleasant life most of us are living now. It is a lazy world of plenty. I can eat, sleep and chill as much as I want. Though the ‘want’ itself is reducing gradually. I wish someone could quickly make a time machine so I could actually deliver the letter to my younger self. How crazy that would be?

It is my second thought of the month, my assistant just told me. And I said that’s exceptional progress over the last month!

My dog reminded me that it’s time to go home, as my wife is scheduled to spend the next 1 hour with him. I stood up and started to walk. The automatic transport pod was too expensive, so have to waste my time walking.

I think the letter should be ready by the time I reach home. Anyways, I feel so lazy in writing something down now. I told my dog that I am too excited to listen to that letter today before sleeping, and he replied, “Woof! Woof!”

India is building the kitchen. China still owns the spice rack. The Quad just broke ground on a very expensive alternative.

On May 16, 2026, Prime Minister Narendra Modi stood in The Hague and watched Tata Electronics sign an agreement with ASML, the Dutch company that manufactures the most sophisticated machines in human history. Machines so precise they print circuits onto silicon using light waves shorter than the width of a virus. Machines that every serious chip manufacturer on earth, from TSMC to Samsung to Intel, cannot function without. India just got access to them.

It was, by any reasonable measure, a genuinely historic moment. The kind of photograph that ends up in a chapter of a book written twenty years from now about how India joined the top table of technology nations.

There is just one thing worth mentioning. The rare earth minerals required to actually run that fab, gallium, germanium, antimony, the chemical backbone of compound semiconductors, currently come predominantly from China. Which means India has not ended its chip dependency. It has been renovated it. New address, same landlord.

This is the story of India’s semiconductor ambition told honestly. Which is to say, it is a story about real progress and a real blindspot existing in the same breath, often in the same building. And, as of five days ago, the most serious attempt yet to fix the blindspot.

Start with the progress, because it deserves to be said plainly. India is building chips now. Actual chips. Micron, CG Power, Kaynes Technology, and Tata Electronics will all begin commercial semiconductor manufacturing in 2026. The Tata Electronics-PSMC fab in Dholera is targeting 50,000 wafer starts per month, producing automotive chips, AI accelerators, and IoT processors, with first silicon due in December. The Tata facility in Jagiroad, Assam will produce 48 million chips per day at full capacity, creating 15,000 direct jobs. Thirteen approved projects across six states, cumulative investment of approximately Rs 1.64 lakh crore, and the Union Budget 2026-27 allocated Rs 8,000 crore to the mission, the largest single-year outlay since launch. India is not announcing intentions anymore. It is pouring concrete.

And yet. Ninety percent of semiconductor chips used in India are currently imported, with over sixty percent coming from China. The plants going up in Gujarat and Assam are real, but without a domestic or allied supply of the minerals that feed them, they are sophisticated assembly operations running on someone else’s raw materials. The dependency has not been dismantled. It has been made more complicated, which is not the same thing.

China understood this leverage before anyone else did. In late 2025, Beijing imposed export restrictions on gallium, germanium, and antimony, the precise elements that compound semiconductors cannot be made without. It was not an accident of trade policy. It was a demonstration. A reminder that in the chip supply chain, the actual chokepoint is not the factory. It is the ground beneath it. China accounts for nearly 69% of global rare earth mine production and processes close to 90% of the world’s rare earth supply. A country can build the most advanced fab in Asia and still be one export ban away from running out of the materials to operate it.

The scale of the exposure becomes clearer when you look at what India is building on top of these chips. The IndiaAI Mission has committed Rs 10,372 crore to sovereign GPU infrastructure, targeting 100,000 GPUs by late 2026. Yotta’s Shakti Cloud alone plans to host over 20,000 NVIDIA Blackwell Ultra GPUs by August. Every one of those GPUs runs on minerals India currently imports. India is constructing an AI superstructure on a mineral foundation it does not yet control. It is an impressive building on uncertain ground.

India’s initial responses to this were sensible but limited. A National Critical Mineral Stockpile was announced in October 2025, a buffer reserve designed to absorb the shock of exactly the kind of export ban China had just demonstrated it was willing to use. In February 2026, India became the tenth signatory to Pax Silica, an international pact aimed at securing AI-related mineral supplies. Useful. Necessary. Not structural.

Then, five days ago, something shifted in scale.

On May 26, at the Quad Foreign Ministers’ Meeting in New Delhi, India, the United States, Japan, and Australia formally launched the Quad Critical Minerals Initiative Framework, committing to mobilise $20 billion toward mining, processing, recycling, and regulatory alignment, with semiconductor supply chains as an explicit priority. The framework combines public funding with private sector participation across loans, guarantees, subsidies, and long-term purchase agreements. Crucially, it is not just about digging minerals out of the ground. It covers processing and recycling infrastructure, which is where China’s real dominance sits. Mining the ore is the easy part. Refining it to semiconductor grade is where the monopoly lives, and this framework is aimed directly at that bottleneck.

The timing is not coincidental. The Quad Critical Minerals Initiative is a direct response to China’s export controls, and it represents the first coordinated attempt by four serious industrial economies to build an alternative supply chain rather than simply stockpile what China might cut off. Australia brings the raw reserves. The United States brings capital and processing technology. Japan brings decades of precision materials manufacturing. India brings the demand, the domestic deposits, and the political will to finally develop them.

On that last point, India’s 2026-27 budget introduced a policy to create rare earth corridors across Odisha, Kerala, Andhra Pradesh, and Tamil Nadu, designated hubs for mining, processing, and manufacturing high-performance rare earth magnets used in EVs and semiconductors. The corridors are new. The mineral deposits underneath them are not. India has significant rare earth reserves that have historically been underdeveloped, underlicensed, and largely ignored by a mining sector more comfortable with iron ore and coal. That is changing, but the gap between a budget announcement and an operational processing facility that Indian fabs can actually draw from is measured in years, not quarters.

Which is why the honest read of where India stands right now is this: the chip factories are going up faster than the mineral supply chain can support them at genuine sovereignty. The Quad framework is the right structural answer. It is also four days old. The fabs in Dholera and Assam are not waiting for it to mature before they start production. In the near term, the minerals will still largely come from China, because that is what exists and what works, and because no serious industrial policy runs on ideology when it can run on supply.

The Oppenheimer parallel that people reach for when describing India’s chip ambitions is tempting but slightly off. Oppenheimer’s project controlled the entire supply chain from uranium mine to detonation. What India is building right now is more like a very impressive laboratory that has signed excellent agreements about where it will eventually source its chemicals, while ordering the current batch from a supplier who has recently demonstrated a willingness to stop shipping.

That is not a reason to stop building. The fabs need to exist before the supply chain for them makes sense to build. Sequencing matters. But founders and engineers reading this should understand where the real open problems are. The chip factory is being built by Tata and Micron and Kaynes, with the government co-funding half and ASML providing the machines. That part has owners. The mineral processing infrastructure, the rare earth corridors, the domestic refining capacity, the recycling ecosystem that the Quad framework is trying to catalyse, that part is wide open. It is also arguably more strategically important than the fab itself, because a fab without minerals is an expensive building, while a mineral supply chain is leverage that extends across every technology India will build for the next fifty years.

India spent five decades exporting its best chip designers to build everyone else’s semiconductor industry. It is now building its own. Thirteen plants, Rs 1.64 lakh crore, a Quad partnership that is four days old and $20 billion large, and a December deadline for the first silicon off a domestic fabrication line.

The dependency is real. The plan to end it is also real. The race now is whether the plan moves faster than the factories outgrow it.

That is not a bad problem to have. It is, in fact, the most interesting industrial problem India has posed to itself in a generation.