I had a vague understanding of how semis tie in with geopolitics: something that went like, “EUV machines are hard to make and are necessary for leading edge chips, which are also really hard to make, and are really important for AI, and AI is kinda a big deal around here.” If you’re already familiar with the industry, this is probably a waste of time for you to read and you can skip it. However, I wanted to get these thoughts on paper, and thought it might be helpful for those who haven’t had the time to think about it.

I’m not a lithography engineer, which should become obvious pretty quickly. I am also not going to fake a full technical explanation of EUV; if you’re interested in that, unfortunately I’m not your guy.

I don’t think I bring forth any particularly new ideas here, but hopefully I synthesize them in a novel and interesting way to bridge micro and macro. I recommend reading, or skimming, or at a MINIMUM having a bot summarize:

• Hiroku Chuma’s “Increasing complexity and limits of organization in the microlithography industry: implications for science-based industries,” Mike Hobday’s “Product complexity, innovation and industrial organisation,” and most directly John VerWey’s history of EUV for CSET for the multi-decade institutional mess that led to the tech and the importance of the “commons”
• obviously Chris Miller’s Chip War for historical / geopol context, but also The Digital Silk Road by Jonathan Hillman to understand why and how China is exporting digital infra

This is a war game of how China might go down the path of shipping an ASML-equivalent system. The question is: if the tech moat is solvable, what happens next?

As with most things, I started with a dumb question and then followed up with more:

• Why is ASML the only company that can make these things?
• Why can’t a competitor attack the machine component by component?
• If that doesn’t work, why not build one vertically integrated mega-company that develops everything together?
• If the problem is mostly coordination, does AI make the Chinese catch-up path more plausible?
• And if China eventually brute-forces a good-enough EUV ecosystem, what happens next?

This piece is basically me working down the list.

Going on the offensive

The lazy ASML bull pitch has basically always been: EUV is magic, ASML owns the spellbook, therefore nobody can compete. I’ve never been entirely comfortable with that type of tech moat, which is probably why I’m a relatively shitty tech investor. If the moat is “we know secret physics,” I get nervous, because “secret physics” has a way of becoming “plain engineering” given enough time, money, and panic. As a washed-up short seller, I’ve seen this happen with capital equipment across a range of industries, everything from textile printing to kitchen equipment.

Why wouldn’t this happen to ASML? It started with idly thinking about what I’d do as the attacker. Just decompose the scanner:

• light source
• optics
• wafer and reticle stages
• vacuum, metrology, software, contamination control, and service

It’s a bunch of widgets, right? Just make it rain, throw money at each component. Hire every relevant scientist you can find. Replace every piece of ASML one at a time until, eventually, you’ve “Ship of Theseus”‘ed it and boom, you’ve got your own ASML.

That was my first intuition as someone who deeply respects capitalism (as an aside: I recall interviewing for an investment firm straight out of college and the interviewer asked me if I was a capitalist. Not gonna lie, it really weirded me out at the time, but in retrospect probably one of the best interview questions I’ve heard for a finance seat). Economic motivation + fear + cash usually solves a lot. No matter how much people insist that like Sinclair Lewis’ novel, It Can’t Happen Here, it usually does. So why not ASML?

The decomposition trap

EUV isn’t 5,100 physics textbook problems, but really closer to one giant co-design problem.¹

ASML reported 5,100 suppliers in 2025. That number matters, not because “large supplier count = moat,” which would be too easy, but because thousands of parties are working on pieces whose requirements are set by the behavior of other pieces. That’s a lot of correlated variables that feed into each other.

As mentioned, I have virtually no technical background, but the technical point I get at is pretty simple: the integrated system is the product. ASML’s own materials make this explicit:

• ASML says independent optimization of separate parameters is no longer sufficient; scanner tuning, computational lithography, metrology, and process control have to work together.
• On High-NA, larger anamorphic optics halve the exposure field, so ASML had to build new, faster wafer and reticle stages to preserve productivity. “Better optics” changed the requirements for another subsystem.
• Imec’s first electrical-yield validation of High-NA did not validate a scanner in isolation. It required resists, underlayers, photomasks, metrology, OPC, integrated patterning, and etch.

This is where you run into a brick wall with the Ship of Theseus approach:

• You can improve a component and make the machine worse.
• You can sacrifice a component’s standalone performance and improve the machine if it relaxes a worse constraint elsewhere.
• The target is not a local optimum for every supplier; it’s closer to one global operating point where the whole scanner produces acceptable wafers per hour, uptime, overlay, defectivity, maintenance burden, and cost.

I initially thought of this as a giant multivariate optimization problem, but “optimization” implies that you know the objective function and the coefficients, which all related parties are incentivized to hide as much as possible. The map itself is produced by the development process: ASML and imec built a joint High-NA lab so customers, suppliers, and partners could access a prototype and accelerate the learning curve before production deployment. You can’t solve for “good” because “good” isn’t in a vacuum here.

Not really covering any new info here for anyone who’s even glanced at this company before, but this is necessary groundwork.

Okay, then why not build Mega-ASML?

If the components have to be developed together, the next question is obvious: why not put everything under one roof? Why isn’t this one giant company?

The actual ASML structure is easy to misunderstand and over-generalize to other manufacturing equipment. It’s closer to a coordinated technology federation²:

• ASML owns the architecture, interfaces, integration, calibration, customer relationship, service loop, and roadmap.
• ZEISS owns the optics learning loop.
• Cymer and TRUMPF sit inside the source chain.
• Other suppliers own narrow but critical subproblems.

The legal separation makes this look more arm’s-length than it is economically. The industry’s own history makes that clear:

• Intel, TSMC, and Samsung agreed to contribute €1.38 billion to next-generation lithography R&D while buying a combined 23% minority stake in ASML.
• ASML bought Cymer to accelerate EUV source development.
• ASML bought 24.9% of Carl Zeiss SMT and agreed to support roughly €760 million of ZEISS R&D, capex, and supply-chain spending for High-NA.

An ASML-specific paper by Tung and Wan calls the Customer Co-Investment Program “organizational investment.” They argue it helped solve the extraordinary financing needs and information asymmetries surrounding EUV development. I don’t really understand the extension from this to their claim that this program alone explains ASML’s success, but it is further evidence that EUV was not financed through anything resembling a normal vendor/customer relationship.

Either way, it goes without saying that all parties work closely together. Immediately, my monkey brain goes: so why wouldn’t we just cut out the redundant opex that undoubtedly exists and mash it all together? I think the distributed structure partly works around human information limits:

• Specialists can run deep search processes in parallel.
• Tacit knowledge stays close to the teams generating it.
• A weird five-year optics experiment does not have to re-win a capital-allocation fight against source uptime, stage throughput, software, service, and whatever fire is currently burning at TSMC every quarter.

Inside ZEISS, mirror metrology is the mission. Inside a theoretical Mega-ASML, it is one cost center among hundreds. Specialization creates parallel search processes.

The second answer is economics. These are tiny markets in unit terms with absurd fixed costs:

• ASML spent €4.7 billion on R&D across its portfolio in 2025 and sold only 48 EUV systems.
• ZEISS says High-NA alone involves about 2,000 employees, more than 1,200 network partners, billions of euros of investment, and over 2,000 patents.

The economic conclusion is mine, but it seems hard to escape: without a winning scanner customer, many of these suppliers are terrible businesses. The bootstrap problem is circular:

• the optics company is uneconomic without a scanner customer;
• the scanner is useless without the optics company;
• the scanner cannot win a fab without the source, stages, metrology, software, and service organization;
• none of those suppliers can justify the spend without confidence that the scanner and fab customers will exist.

Normal capital markets don’t enjoy funding businesses like this, where sure, once it works, several nodes become monopoly-like, but each node is a bad standalone business until the whole system a) has actual functional science behind it and is b) accepted by customers.

ASML coordinates the global objective and the interfaces between them to get to the “right answer” with economically incentivized suppliers, and this also derisks the mothership from any individual failure from tech and operational risk.

So then, why not tackle the entire system at once and bypass the “coordination” moat?

AI attacks the coordination constraint

I think this where AI matters most; I don’t think it’s particularly relevant to the business if AI can spit out an EUV machine from first principles (based on the capabilities it’s been showing in technical fields, it probably can). However, it could reduce the organizational tax of running an industrial program with thousands of self-interested entities by:

• mapping dependencies across teams and suppliers
• identifying hidden correlations and propose higher-value tests
• batching operations and allocate reps across the system rather than within one silo

There are some pieces of evidence directionally indicating that the greatest value of AI comes from the organizational level. ASML invested €1.3 billion in Mistral and said the partnership would apply AI across its product portfolio, R&D, and operations to improve time to market and holistic-lithography performance. I also came across an interesting paper by Alex Farach suggesting that organizational structure flattens with AI adoption; whether this is causative or just correlation is difficult for me to parse out from the paper.

That alone makes a Mega-ASML attempt more plausible than it would have been ten or fifteen years ago. AI just has to cut friction and reduce enough dumb iteration to matter.³

Who will train the machine?

If you had a central coordinating system, you still need data. The EUV machine needs reps. A prototype is far from being a production-equivalent tool. ASML’s own development model makes that painfully clear (see: joint High-NA lab with imec gives customers access to the prototype scanner plus surrounding processing and metrology tools specifically to accelerate learning before the machines enter production fabs).

A development customer transfers much more than a purchase order:

• real production requirements and process windows
• comparative data versus ASML
• failure logs, downtime, maintenance, and service experience
• feedback on yield, overlay, throughput, defectivity, and reliability over repeated wafer runs
• proof that the tool works outside a protected lab

It makes zero sense for any of the Western-aligned fabs to participate in this. I don’t have a TSMC memo saying “we will never test Chinese lithography.” This is a game-theoretic inference. But the incentive seems obvious:

• Validating the Chinese tool helps train the Chinese tool
• A better Chinese tool makes Chinese leading edge foundries more viable
• Those foundries become subsidized competitors to the fab that supplied the feedback
• Qualification would create U.S., Dutch, and Japanese export-control risk, threaten allied subsidies and customer trust, and strain relationships with the existing equipment stack

You aren’t “helping the company save money.” You are helping train the ecosystem that may eventually compete for your customers. If you adopt Chinese semicap equipment, as the kids on the internet would say, you’ve committed (dishonorable) sudoku. Any sense of self-preservation and a time horizon of longer than next twelve months kills the idea.

Western firms will obviously tear down and benchmark Chinese machines. That is different from placing them on a production line and becoming a development customer. China will have to train the machine itself rather than relying on the obvious existing commercial partners.

The first captive loop

China having a viable EUV machine is looking increasing less theoretical. Reuters reported that China completed a crude factory-scale EUV prototype in early 2025 that can generate EUV light. Huawei is coordinating a web of companies and state research institutes involving thousands of engineers. Reuters reported in July 2026 that a state-owned company had begun producing immersion DUV machines that still lag ASML and require further testing. The first units are expected to go directly to SMIC, Hua Hong, and CXMT. Whether or not these are commercially viable is still entirely unproven, but the placement is the point.

The tool-level loop is:

state-funded equipment developer → captive domestic fab qualification → production failures and service data → improved tool.

This is why the state matters: a normal company can’t rationally fund the supplier network, scanner integrator, immature product, and captive qualification customer while everything is still worse than the foreign alternative. Shareholders like to ask annoying questions like “why are we lighting money on fire?”

A state can answer “because dependence on a hostile supply chain is worse and this is existential to our hopes for hegemony and contesting other nations.” Everything can lose money; the value exists at the national security level. There’s an institutional route and willingness to generating the required reps.

So now you’ve got yourself a lithography machine.

Very cool, but you’re still not at the starting line, let alone the locker room. Closer to JV tryouts. ASML and lithography are one part of a much larger system. Imec’s High-NA work is a useful reality check: even the first electrical-yield demonstration required the scanner plus resists, underlayers, masks, metrology, OPC, pattern transfer, etch, metallization, and electrical testing. Zoom out further and a leading edge foundry also needs:

• deposition, etch, cleaning, implantation, annealing, CMP, inspection, and process control
• masks, pellicles, photoresists, gases, chemicals, and materials
• PDKs, EDA compatibility, standard cells, SRAM compilers, and interface IP
• reliability models, advanced packaging, HBM integration, thermal solutions, and test

You’ve overcome one multivariable optimization problem and stepped into a fractal. Turtles all the way down (and across).

The foundry product is the integrated process and its yield. A change in one module shifts the process window elsewhere; the fab has to identify which step killed the chip, change the recipe, and make sure the fix did not create three new problems. This probably sounds familiar from the earlier section.

At TSMC, the diversity of layouts, workloads, packaging, etc. improve process learning. Last year, TSMC manufactured 12,682 products using 305 technologies for 534 customers. This supports high utilization and profitability, but also funds future investment and is an enormous paid test surface.

There is old but directly relevant semiconductor evidence for the value of reps. Irwin and Klenow studied seven DRAM generations and estimated average learning rates of roughly 20%; more importantly, firms learned about three times more from their own cumulative production than from an equivalent increase in competitors’ production, while learning transferred weakly across generations. The exact numbers clearly shouldn’t be transplanted onto modern semis, but the mechanism is useful: observing the industry is not the same thing as running your own wafers, with the caveat that each generation can require substantial relearning.

The machine needs reps. Then the foundry needs a second, much larger set of reps.

The second captive loop: who will feed the foundry?

The obvious first answer is China itself.

• Huawei and other domestic designers can provide the demand, with telecom operators, the military, state-owned enterprises, government cloud, and consumer-device companies being end customers
• Reuters reported that China ordered state-funded data centers to use domestic AI chips; projects less than 30% complete were told to remove foreign chips or cancel planned purchases. Chinese AI data-center projects had already drawn more than $100 billion in state funding since 2021.

China can force enough domestic utilization to generate learning and produce priority chips while tanking a disgusting operating / financial profile that would probably cut a publicly listed foundry stock in half, if not more. But external demand still matters if you want this to resemble anything close to a commercially viable enterprise. It adds:

• more utilization and scale, alleviating funding strain
• broader deployment experience and product feedback
• legitimacy for the Chinese technology stack

We’ve covered the silicon (in embarrassingly rudimentary detail, but nonetheless), and the necessary circular systems that must be replicated at both the equipment (and many times horizontally across each piece of equipment) and foundry level. How is this reflected in the statecraft?

Who’s gonna step up to the plate?

The “axis of evil” countries ain’t it. Russia, Iran, and North Korea may be strategically motivated customers, but not large or rich enough to transform the economics.

The swing customers are countries with money or scale that don’t want their technological future permissioned by Washington:

• the Gulf, especially Saudi Arabia and the UAE
• parts of Southeast Asia
• Brazil and other autonomy-minded EM countries
• eventually, Africa, whose digital infrastructure China has helped finance and standardize?

The Gulf states are probably both most susceptible and most desirable: Carnegie describes the UAE as a technologically ambitious state seeking to balance relations with both the United States and China. Tin Hinane El Kadi similarly argues that Gulf governments are active participants, using competition among Chinese and Western firms to extract localization, cloud regions, investment, training, and technical capacity rather than passively choosing a flag. China may not need the Gulf to defect; it needs the Chinese stack to become a credible fallback and bargaining chip.

If we look to emerging markets and take Africa as an example: if China supports Africa in industrializing, there still probably won’t be any major African fabs anytime soon, but if China can support demand, they can buy the output:

• Chinese accelerators, servers, and networking
• cloud capacity and models
• telecom, smart-city, and industrial systems
• data centers, power infrastructure, and Chinese financing

China’s Digital Silk Road already packages and exports infrastructure, vendors, standards, and financing together for digital infrastructure. That’s the closest existing analogue to “sell the stack” as opposed to actual wafers: the semiconductor layer sits inside a broader export system rather than being a standalone foundry service.

The stack will be way shittier, but you are more likely to be able to afford it, and they’ll actually let you have it. End demand then flows backward into Chinese chip-design and foundry volume.

The export-control own goal

The policy question gets a bit awkward here:

• Export controls slow China’s access to frontier tools and chips
• They also force Chinese customers to use worse domestic alternatives, creating the reps those alternatives were missing
• If neutral countries believe Western access is arbitrary or revocable, they gain a reason to fund or adopt the Chinese fallback

Obviously the U.S. shouldn’t hand China frontier chips and equipment, but they need to deny capability to China without making every neutral country feel that it is renting technological permission from Washington.

While effective in the short term, the underlying mechanism of “weaponized interdependence”, with states controlling hubs that they can deny access to for political power, probably also raises the target’s willingness to spend buckets of money building an alternative network.

CSIS estimates Chinese semiconductor-equipment penetration rose from 25% to 35% between 2024 and 2025, in part because export controls coordinated customer demand with domestic suppliers’ roadmaps.

The United States’ July 2026 move to give approved UAE entities license-free access to advanced computing items is a decent example of the opposite approach: use access to the superior stack as geopolitical glue. Carrot and stick.

If the West keeps capital-rich and scale-rich hedgers inside the allied ecosystem, China may build sovereignty but remain commercially boxed in. If we piss neutral parties off, China gets the volume, legitimacy, and deployment experience it needs to make the parallel stack much more credible.

Show me the money — what does this mean for the stocks?

Sorry if you’ve made it all the way down here for a stock pitch, I don’t have a particularly clean twelve-month trade from a ten- or twenty-year thought exercise. But some high-level thoughts which are likely to be useless to anybody who actually knows what they’re doing in semis:

• ASML: Chinese EUV success is virtually irrelevant to the business, which might be a hot take. China already can’t buy EUV (legally, but what’s a little bit of sanctions violation between friends). The more realistic bear case is that ASML’s customers lose volume to upstart domestic fabs, not TSM cutting orders.
• TSMC: the strategic risk sits one layer downstream. China threatens TSMC if it creates a self-reinforcing foundry learning loop, domestic design ecosystem, and external customer base. TSMC’s moat is recipes, but also the customer base.
• NVIDIA and other chip designers: a self-contained Chinese foundry and AI stack can permanently remove Chinese demand from the Western merchant ecosystem and compete for sovereign deployments elsewhere.
• Semicap and materials broadly: bifurcation can increase gross capex because the world builds two versions of the same stack. More fabs, tools, capacity, and duplication can be good for unit demand while being terrible for returns on capital.

Outside of these though, EM is… surprisingly interesting? I mean, that’s kind of the main conclusion of all this, right? China (or anybody else for that matter) who wants to go down this decoupling road needs to recruit other polities to their side. They seem like the most likely parties to capture the pressure from two sumo wrestlers shoving against each other.

Back to me writing this. Thanks for taking the time to read; I’m hungry and kinda phoning it in here at the end, so maybe there will be forthcoming edits. I will be putting out more directly single stock-focused analyses in the future, but thought this was a helpful exercise personally to get back into writing investment-related material, and ideally it made you consider something new too.

Footnotes:

1. Herbert Simon’s useful term for this is “near-decomposability”: complex systems can be split into subsystems, but the boundaries are imperfect and cross-system feedback still matters. Hiroyuki Chuma applies a closely related idea directly to microlithography, calling ASML’s approach “interim modularity”: provisional interfaces that let dispersed specialists work in parallel during trial-and-error development even though the final product is not actually modular. Breaking the machine into boxes is an organizational convenience, not proof that the boxes can be optimized independently. ↩︎
2. Brusoni, Prencipe, and Pavitt’s formulation is that complex systems integrators must “know more than they make”: outsourcing production does not let the integrator outsource architectural knowledge, because component technologies progress at different rates and interact unpredictably at the product level. Mike Hobday’s “complex product systems” literature adds that for high-cost, low-volume, engineering-intensive products with deep customer participation, the relevant unit of analysis is often the project or system of firms rather than the nominal lead company, a useful way to think about ASML + ZEISS + the source ecosystem + the fabs. ↩︎
3. Gary Pisano provides a useful distinction between “learning before doing” and “learning by doing.” Where the underlying process is scientifically mature, laboratory work can materially compress development. Where the process remains partly an “art,” more lab work does not necessarily eliminate the learning that occurs only in the final production environment. His evidence is from pharmaceuticals, not semiconductors, so this is an analogue, but it is exactly the right AI frame: AI should make learning-before-doing much better; doesn’t follow that it removes learning-by-doing ↩︎

First of all, hope the back half of 2025 and the start of 2026 have been excellent to you! Secondly, I apologize for being so delinquent with my posts; it’s been a minute since we last corresponded, hasn’t it? For those of you who are somehow still keeping up with this blog, some updates!

The Runescape Saga

I detailed the beginning of this project in my previous post, “Day 1: Backtesting Fish, Fighting Python”. I thought I owe y’all an update, even though details are admittedly hazy here. Overall, the project became somewhat more trivial / unexciting than I’d thought — the big hurdle was pulling pricing and volume data (which had granularity issues of its own, but more on that later), but once you overcame that, it was relatively easy to implement basic strategies:

• relative value was probably the easiest: arbing iron ore vs. iron bars for example, or food items used in combat — there was a period in time where lobsters traded more richly than swordfish, which made no sense because lobsters restore 12 hp vs swordfish at 14. There was no reason for the lobster trade to have been available: these items have no additional use that I’m aware of outside of weird burnt food collectors, and inventory space is a premium during combat.
• moving averages and seasonality — there were several items that had pronounced daily seasonality, like soft clay if I remember correctly. Going off of memory, but if you bought prior to 10 am CT and monetized after 8 pm, there was persistent alpha.
• macro trades — unfortunately, macro isn’t particularly lucrative in OSRS, given the long time frames required. The big event I was positioned for was the release of Sailing, which was the first new skill added to Runescape in over 15 years. There was a Reddit post (I believe this one) that discussed the raw materials (planks, nails, logs). I bought those, sat on them, and produced >200% returns over 6 months. Sounds great, but unfortunately this was relatively unimpressive compared to the alternatives above. I didn’t bother tracking performance, but I’m fairly confident it’s possible to do 5% daily returns using the stat arb strategies above.

While it was fun to watch the gold stack up, I gradually lost interest over time because of difficulty scaling: you only have three trading slots (i.e., you can only place three buy OR sell orders at a time, so it’s impossible to leave multiple standing orders on both sides of the trade with only one account) and OSRS is pretty strict on automated activity / botting. I did make multiple accounts but managing all of them kinda became a pain in the ass at the end, and I’m a cheap bastard so I didn’t feel like paying for membership, which expands available slots to six. I considered expanding via setting up a clan and basically giving people a cut of trading profits, but I couldn’t think of a way to a) prevent alpha leakage and b) figure out good enough incentives (I wrote on the post-scarcity nature of OSRS in my Medium post here: RuneScape as a Post-Scarcity Model: A Closed-System Preview of Late-Stage Capitalism). After racking up more gold in a few months than I had in YEARS of playing OSRS as a child, I lost interest and Gielinor’s preeminent one-man multistrat hedge fund*, Zaros Capital, returned all external capital (seed capital of 60K gold pieces from a friend who had played OSRS years ago; returned many multiples of this!), and wound down. Given no personnel, liquid holdings, and no physical assets associated with the firm, no meaningful expenses to the investor were incurred as a result of the closing.

*Not independently verified, but also not contested as far as I know.

Disclaimer: This material is for informational purposes only and does not constitute an offer to sell or a solicitation of an offer to buy any securities or investment products. Past performance is not indicative of future results. All returns are hypothetical, unaudited, and denominated in an unstable medieval currency.

So, now that you’ve finished your sage as an MMORPG hedge fund manager, what’s next?

At a Crossroads

For those of you who don’t know me at a personal level, I love picking stocks. I consider it core to be who I am, which probably sounds weird to people who have occupations they don’t love (which I’m not knocking by the way; from what I see with my peers, unfortunately 90% of the time, it’s very difficult to connect work with play / real interests). I thought “solving the infinite puzzle” of financial markets is the only real place that someone like me (total nerd, obsession with figuring out how things work, desire for a high degree of agency, deep critical thinker, not particularly talented in technical fields like math or science nor good with my hands) could achieve professional fulfillment. I really love this shit. I could do it until the day I die or the robots drag me away from the desk (“beep boop inefficient capital allocator detected”?).

However, for the first time, I’m seriously considering other options for the long term.

It’s… Time to Build?

During the Zaros Capital saga, I used Cursor fairly heavily to handle things like pulling data, running quick backtests, and producing automated signals. This was back in May, and honestly I was already pretty impressed with capabilities at the time. The fact that I could verbalize what I was looking for and have a functional script produced in a matter of minutes, if not seconds, was mindblowing to me — this would probably have taken hours, if not days, for me to do myself (as a matter of fact, I KNOW it would have, because the last time I tried to create a script myself for an investment project I was working on, it took comfortably over a week). Although OSRS was ultimately meaningless, it did help open my eyes to possibilities I hadn’t previously considered.

I genuinely think now is one of the best times in history to start a business — especially a small, software-driven one.

For the first time, the constraint is no longer “can I build this?” but “should this exist?” With tools like Cursor, Claude Code, and other agentic workflows, the marginal cost of turning an idea into something functional has collapsed. Things that would have required a small team a few years ago can now be done by one person with a decent AI IDE subscription, a clear mental model, and a willingness to iterate.

Even the non-technical parts of building a business — marketing, positioning, customer discovery, pricing — feel more tractable than they used to. Not because they’re easy, but because you’re no longer starting from a blank page. You can pressure-test ideas, sanity-check assumptions, and learn best practices without spending months reinventing the wheel.

In practical terms, it feels plausible to me that you could create a real, revenue-generating online business today with:

• a handful of AI tooling subscriptions,
• a laptop,
• and well under $100,000 in capital.

That simply wasn’t true for most of my career.

What makes this moment particularly interesting is that these paths don’t feel mutually exclusive in the way they once did. The skills that make someone a good investor (synthesis, judgment under uncertainty, comfort with incomplete information, systems thinking, ability to deal with the slog and general bullshit) map surprisingly well onto building products. And the tools now exist to let one person move fast enough that the experiment doesn’t have to be all-or-nothing.

Given my nature (skeptical, risk-averse for the most part (my love for games of chance notwithstanding), and dislike of “bitch work,” the fact that I’m even considering should be a sign to both you and me of how real the opportunities are. Fulfillment is out there for all of us. Sorry to wax poetic, but I’ll close out with a16z’s same closing paragraph from their article which shares the same name as this section:

“Our nation and our civilization were built on production, on building. Our forefathers and foremothers built roads and trains, farms and factories, then the computer, the microchip, the smartphone, and uncounted thousands of other things that we now take for granted, that are all around us, that define our lives and provide for our well-being. There is only one way to honor their legacy and to create the future we want for our own children and grandchildren, and that’s to build.”

See you next time, and hopefully it wont be more than 6 months in between posts!