If you need an investor that can put on their big boy pants and lead your round, call me.

We're so glad to have doubled down on our conviction in Rio through this round! Medicine on qcomm is a different beast - the consumer behaviour, the frequency of need and the sensitivity of what's being delivered all raise the difficulty bar well beyond other categories. Their efficient use of AI, strong customer obsession and bottoms-up execution have already helped the team scale to ~30k orders/month, with 80% coming from repeat customers. Super bullish on Amit, Ankur and the team!

We Built the Product Apple Gave Up On...

Amazon is now blocking Meta's Muse agent from buying anything on its site. I think that's the first real shot in a fight that's been coming for a while. 1/ Any commerce site is typically the one owning the relationship. The app has always been the place where you browse, get nudged and build the habit of coming back. If you put an agent in the middle of this equation, the app is just reduced to shipping whatever the agent decides to buy (that too if your app is offering the best price / fastest shipping / whatever the consumer demands) and so loses out on that relationship. 2/ A lot of the money on these platforms isn't the sale but the ad revenue. That whole business assumes a human is looking at a ranked screen and can be nudged into buying the sponsored product. The ad stops working because you can't sell placement to an agent that sorts by spec and never sees the placement. For Amazon, that ad business is one of the fattest margins it has. If agents continue to win - which seems likely - most of the commerce stack will either wall the agents out or race to build their own that's on par with Muse and the others.

India now makes close to a quarter of the world's iPhones. Apple exported iPhones worth a record ₹2 lakh crore from India last year, making it the country's single largest branded export. So the manufacturing prowess question feels settled. We've shown we can build to that standard, at that scale, for the most demanding customer. But the numbers that show "made in India" are really just "assembled in India." The design, chip, R&D, and IP still sit in Cupertino despite India possessing the capability. CMF going independent here is at least a bet that we finally own this lucrative layer too.

We've been seeing Indian brands lean heavily into their "Indianness." Usually, it shows up in uniquely Indian nomenclature or in maximalist packaging. While this is used as a branding exercise, there's another way founders are approaching their Indianness - through insight. India offers a set of vantages that are genuinely uncommon. Indian markets sit at intersections that don't exist elsewhere: deep informality alongside digital public infrastructure, low institutional trust alongside dense personal networks, small physical spaces alongside high aspiration. A founder living inside this can see problems that a founder outside it structurally cannot, and that is the resource. Nuuk, a company we've backed, is a good example of what I mean. There's nothing obviously Indian about the way it looks. But the insight the founders had was that the Indian household is changing from the inside: more nuclear families in the metros, higher disposable income, a growing set of the demographic "urban DINKs." Meanwhile, the incumbent appliance brands like Bajaj, Usha & Havells have kept their design language and product architecture largely unchanged for 3 decades. Nothing on the shelf was actually built for the household Nuuk's founders were watching form around them. It's the kind of problem you can only really see from inside the Indian context. I feel this version of displayed Indianness will go a longer way.

MDR will not reach the end consumer as inflation only if merchants steer people into cash payments or get them to split up bills. Both will leave a trace in the UPI data in a few years: splitting could drag the ticket sizes down & cash would eventually thin the volume tracked.

I see often this. Jevon's paradox applied to the attention economy without tuning. Professor Herbert Simon (Nobel in economics) famously argued : abundant information creates scarce attention and the effect in unbounded. This explains why you can't remember the specific reel you promised yourself you will remember. Its impact on entertainment is even more insidious, as the amount of content become infinite, you stop engaging with content beyond your shortest attention span. in 2006, Prof. Lanham (UCLA proffesor or rhetoric) wrote the book The economics of attention, where he argued once information is abundant and attention scarce, all value accrues to Style. His used this term to mean the packaging, presentation, and rhetorical craft that makes something noticeable. Trouble with Style is that you need novelty and variation. ((Yes!!! the term TASTE is not that new )) Prof Lanham explained why the marvel universe suddenly started delivering flops even as the costs went up.

One of the more interesting uses of AI in production i've seen is PocketFM. AI has cut their production cost by 80x, but the lagging metric that increased was retention (44% to 76% in ~2 years). This is because cheaper production let them make stories for the long tail of listeners that are too niche to justify a human writer's time. So scaling with AI made even serving smaller audiences economically viable.

One of the reasons Dario gives in his essay for why the AI industry needs to slow down is recursive self-improvement: AI systems are now doing a meaningful share of the work that produces the next generation of AI. Models help design, train, and evaluate their successors, and capability compounds on itself. However, there's a second recursive loop running at the same time. In a 2024 Nature paper, researchers showed that when AI models are trained on data produced by earlier AI models, they progressively lose information about the underlying distribution. Over gens, the model converges toward a narrower version of reality. The authors call this model collapse and prove it is mathematically inevitable, even under ideal conditions. Both recursive processes are running inside the industry at once, even though they work in opposite directions. One makes AI more capable and the other makes the data it learns from drift away from reality. It's hard to say what they compound into. The assumption is that systems get more capable while their grip on the world they're modelling gets weaker. Faster capability, weaker ground truth. If both loops run unchecked at the same time, the models become more powerful at acting on a reality they understand less well. That's the case for pacing.

Mass A.I training is happening in India. It is hard to imagine what the world will look like in the next 5-10 years.

I do not support pacing frontier models. If I am going to die at the hands of killer AI, I want it to be American, not Chinese. Buy American, Die American.

The CAC gap here is not just huge but it's been widening. Meesho's cost to acquire a customer has roughly halved over 3 years, from ~₹350 to ~₹170. Nykaa's has gone the other way, climbing past ₹2,800. At the mass end (Meesho), scale eventually lowers your acquisition cost; network effects and reseller pull do the work. This is because the long-tail supply that few want to formalise becomes the moat that attracts mass demand. At the premium end (Nykaa), scale raises the acquisition cost because you exhaust the easy customers early and then pay up for a smaller share of the market through more curation, so higher ARPU can justify the economics.

Was wondering who the market for this is and apparently foldable phones are still a fairly niche category, even 7 years after Samsung started producing them. They're just 2% of all phones sold globally.