Nik_Maestro

Uber has proved it can make transportation capacity more productive. That is not the next question. Waymo wants riders to choose Waymo. Amazon can put Uber Eats behind the Prime relationship. AI agents may eventually compare both on the customer’s behalf. So, availability is no longer enough. Uber needs a reason to be selected. That is where Uber One becomes more interesting. If membership benefits improve the actual offer price, pickup, service quality, rewards then the agent can still choose Uber even when Uber does not own the interface. The next moat may be simple: not being available but being worth choosing. nikhs.substack.com/p/uber-3q…

Palantir’s contracts are getting shorter. That sounds bearish. It may not be. Palantir’s strongest argument is increasingly that the contract itself does not create the lock-in. The customer owns the weights. But the workflows, permissions, evaluations and retraining loops accumulate inside the Ontology. Leaving can therefore become expensive even without a five-year commitment. That is potentially a better business. But it is also harder for investors to measure. The contract is visible. The reconstruction cost is not. nikhs.substack.com/p/palanti…

Lam Research is becoming more important to semiconductor manufacturing. That part of the thesis is working. 3D DRAM, backside power, advanced packaging and more complex NAND all require more etch and deposition. But Q1 forced an important correction. Technical importance is not the same as pricing power. ASML can charge for the scarcity of its machines. Lam still must keep winning more process steps. The toll is real. The question is how much of it Lam gets to keep. nikhs.substack.com/p/lam-res…

Meta’s agent thesis got stronger this quarter. Muse passed 10 million users in six weeks. Shopping is already its most popular use case. But the most important development was somewhere else. Meta created a new non-voting share class after internal forecasts contemplated raising hundreds of billions of dollars over several years. That changes the investment question. The issue is no longer simply whether Meta can build a valuable agent network. It is how much capital it takes to build and how much of the resulting value each existing shareholder gets to keep. nikhs.substack.com/p/meta-3q…

The most interesting thing TSMC said this quarter was that it is buying from another foundry. GlobalFoundries will supply silicon interposers because TSMC does not have enough back-end capacity. At first glance, that sounds like outsourcing. I think it shows something more important. AI chips are no longer just about the transistor. Logic, memory, interposers, substrates, power and cooling increasingly must work as one system. TSMC does not need to manufacture every part. It needs to remain the company responsible for making the whole thing work. That is where the learning accumulates: yield, thermal behavior, materials, packaging and the handoffs between components. The market already understands that TSMC will grow rapidly over the next two years. The bigger question is whether this integration advantage survives the capacity buildout and keeps compounding after the shortage ends. nikhs.substack.com/p/tsmc-3q…

In 2012, ASML’s customers paid because EUV did not yet work. In 2026, they are paying because it works and they cannot get enough of it. Customer advances have surged. Orders require deposits “enough to hurt.” ASML is rationing future capacity. And customers have accepted higher prices on its most critical tools. That is the important change. Scarcity has allowed ASML to renegotiate how much of the value it creates it gets to keep. But ASML is also spending heavily to remove that scarcity. So the investment question is no longer whether demand is strong. It is whether today’s pricing power becomes tomorrow’s margin. nikhs.substack.com/p/asml-3q…

What we said earlier. Meta's muse is aimed at smb not th fortune 500.

MongoDB fell 29% in six weeks. The obvious story is the CEO departure. I think the bigger story is that the market may have MongoDB in the wrong drawer. Most software gets paid per seat. MongoDB Atlas gets paid for reads, writes, storage and compute. If agents do more work, MongoDB can sell more of what it already monetizes. Snowflake showed how powerful that reclassification can be. MongoDB may be next. The key: our base case does not require it. At roughly today’s multiple, we still get to about $640 in three years. AI is not the thesis. It is the mechanism that could force the market to refile the business. MongoDB Is in the Wrong Drawer nikhs.substack.com/p/mongodb…

Memory cycles eventually end. That does not mean all their value has to disappear with them. Micron is signing customers years forward, taking deposits and gaining better visibility into demand before new supply arrives. The important question now is not whether today’s margins last forever. It is what Micron does with the cash while they last. If contracts support disciplined investment and excess cash retires shares at sensible prices, a temporary shortage can leave a permanent benefit per share. My latest piece is about selling the cycle forward. nikhs.substack.com/p/micron-…

Meta already owns much of the conversation. A seller posts on Instagram. A customer discovers her there. They move to WhatsApp. Then both start doing the actual work by hand. That is the opportunity. Give the customer an agent. Give the seller one too. Now the conversation can carry real commitments: stock, price, delivery, payment not just messages. Facebook Home tried to make relationships central by putting them ahead of apps. Agents may finally make the relationship central by getting the work done through it. nikhs.substack.com/p/faceboo…

Meta's enterprise business won't be won in a Fortune 500 procurement meeting. It'll be won in a shop in Jakarta. The Desai hire has everyone framing Meta Enterprise Platform as a run at US CIOs : Meta vs. Microsoft vs. Google Cloud for the corporate IT budget. That's the fight Meta is most likely to lose, and we've already watched it play out once. Google Cloud nearly lost a decade trying to become an enterprise vendor. It had better technology than almost anyone, and it didn't matter. Enterprise is won top-down (sales teams, contracts, compliance reviews, relationships with IT departments), and that is Microsoft's home turf. Meta would be walking into the same trap with even less enterprise muscle. But Meta has something rarer than a cloud business: a consumer product people pull toward them. That's the key point. Meta has never won by selling to gatekeepers; it wins by going around them. So the question isn't whether Meta can sell to a CIO. It's where there is no CIO to sell to. The answer is emerging-market small business. WhatsApp is already the operating system of small business in India, Brazil, Indonesia, and much of Southeast Asia. Orders come in over WhatsApp. Customer questions come in over WhatsApp. Payments increasingly settle there. Microsoft has little presence in these markets, and no IT department stands in the way. In the US, Meta has to convince a buyer. In Jakarta, the customer is already inside the product. Which means the enterprise upgrade isn't a sale at all. It's an agent. A small business running its customer service, orders, and payments through a WhatsApp agent doesn't "adopt enterprise software." It keeps using WhatsApp, and WhatsApp starts doing the work. The Reliance joint venture in India points exactly this way. Desai is the headline. That shop in Jakarta is the prize. (Note the timing, too: nobody recruits a sitting public-company CEO in three weeks. This pillar was being built before Muse proved anything; it's Meta's answer to Wall Street's capex question. That's fine, as long as Meta remembers where its leverage actually is.) The risk is that it doesn't. If Meta chases US CIOs, builds a Google Cloud-style sales org, and pulls compute away from Muse to serve enterprise contracts, it trades its rarest asset for its weakest position. How to tell which path Meta is on: look at where its first enterprise customers come from. If the first disclosed revenue is mostly US large enterprises, Meta is running Google's play. If it's a million small businesses on WhatsApp, Meta is running its own.

Every fast-growing footwear brand eventually seems to pay the same bill. Too much inventory. Then markdowns. Then outlets. Then customers learn to wait. On Holding has somehow scaled to CHF3 billion of revenue while largely avoiding that cycle. That may be the most important fact in the story. The question is whether On built a better machine or simply grew during the best weather a footwear challenger has ever seen. 2027 should tell us. My latest piece is about The Missing Discount. nikhs.substack.com/p/on-hold…

three proposals in amodei's essay global coordination with china domestic industry-wide coordination embedded evaluators anthropic pre-commits to today the first two are impossible that's the point they make institutional capture look like restraint

the easy work is now trivially easy & the hard work is easy. so what separates ppl is no longer execution. it's the ambition behind what they're willing to attempt. the old unicorn hire was the product thinker who could also engineer & design valuable cuz they collapsed the translation layers between functions. everyone has that now for $20 a month. most ppl still use ai to do their old job faster. the ones pulling away use it to attempt things that were never on their list at all. the constraint moved from what you can execute to what you can imagine. & the uncomfortable part is that models can do all the executing but there's no model for wanting more. ambition is fully manual.

from the outside the last 2 weeks look like every lab coordinating a regulatory capture play. from the inside it's much simpler: they're staring at their scaling plots & the plots have never lied. there are only 2 ways ai capabilities have advanced in the last decade, scale farther on an existing law, or discover a new one. gpt2/3/4 were pre-training scale-ups. o1/o3 opened a new axis (test-time compute). the real question was never "how much farther can we scale." it was "how many unsaturated axes do we know about." & the answer inside the labs right now is prolly more than you think. models went from no advanced math ability to solving a millennium prize problem in under 2 years. same in coding. same, quietly, in hacking which happens to be the gate to the entire internet & most of the physical infrastructure on the planet. that's not overhype. it's what happens when you look at the curve & it still doesn't bend.

the framing on personal agents is off. everyone judges them by the errands they replace. the real opportunity is the work that becomes viable once organizing it costs ~nothing. the biggest competitor to a personal agent isn't openai's next model. it's the small business, community project, or research thing you never started cuz the coordination overhead wasn't worth it. lower that overhead & suddenly the smallest viable business the smallest viable anything just got a lot smaller. that's not automation. that's market creation.

Oracle spent years looking like a software company trying to finance an infrastructure build. Q1 showed something different. Customers prepaid $11.4 billion toward that expansion. That matters. Oracle is not becoming capital-light. It is becoming cash-timing-light: customers fund more of the build upfront, while Oracle still owes the capacity, service and economics later. The next question is the one that matters for shareholders: After customers, creditors and new shareholders take their share, what does Oracle get to keep? nikhs.substack.com/p/oracles…