Why other countries can't produce tech giants
America suffocates them
(note: this is an appendix for another article on the nature and dangers of the American monopoly on tech)
Lemma #1: Tech is monopolistic by nature
This is axiomatic to any technology investor. After all, this is the source of tech’s outsized returns. There are various types of monopolies.
Some are near malicious. Apple simply does not allow competition: no alternative app stores, no alternative payment processors, no access to the APIs their own products use.
Some are merely anti-competitive - Google pays to be the default search engine in most browsers, pays to be the default app store in most Android distros, and distorts the ad auction marketplace.
Some are natural. Most Software-as-a-Service (SaaS) products are simple economies of scale that amortize expensive development through small subscription fees over millions of users. Network effects mean you use Meta’s products because that’s where everyone is.
All of them suck the oxygen out of the room for other players. In almost every segment - especially anything consumer - there is usually only room for 1 or 2 large players.
All of these large players happen to be American. They got there first. With no perturbations to the equilibrium, they will stay American. Monopolies are called monopolies for a reason. Tech is no exception. IBM, a company we consider a has-been dinosaur today, still dominates mainframes. Only the advent of entirely new computing paradigms creates space for new giants. When the personal computer replaced the mainframe, Microsoft and Apple replaced IBM. As software ate the world, we got Google and Meta. Is there anything to eat software?
Lemma #2: Software-enabled monopolies are “superlinear”
General Electric (GE) held various “monopoly” market positions on household appliances in the 20th century. Yet their dominance slowly crumbled. Other companies simply made better products and people bought them. There was nothing to lock you in. It would be ridiculous if owning a GE toaster meant you couldn’t buy a Samsung plasma TV.
Not so in software. Consider Google Workspace, an email/work productivity suite. All Workspace users were recently forced to “upgrade” to a package with the Gemini AI Assistant. I parenthesize “upgrade” because users are clearly not happy with this change. Yet! With the flip of a boolean switch, they’ve managed to turn a dominant position in an existing market into a dominant position in a burgeoning one. Can you convince your procurement team to buy another (better) assistant tool when you’re forced to pay for one? Probably not. Thus a hundred potential competitors and startups are pre-emptively extinguished.
The success of generative modeling is the biggest technological disruption since the iPhone. But if we zoom out on the Gemini AI Assistant example, we see this story replicated a thousand ways elsewhere. Most of the value has accrued not to the innovators but to those who can distribute it or tack it on to existing products. Adobe has seen more ARR attributable to generative image/video models than any startup other than Midjourney. Microsoft sees more AI-related revenue than OpenAI and Anthropic combined!1
Ah, but the dust hasn’t settled yet, you might say! Generative models haven’t fully embedded themselves in our everyday lives, and it’s too soon to call a winner. I agree. But looking at who’s winning so far - existing BigTech, OpenAI, Anthropic, Midjourney - and looking at those who have already stumbled - Cohere and Mistral - I think we can call a different sort of winner.
The American tech ecosystem.
Lemma #3: The virtuous flywheel of capital and talent
Microsoft invested a billion dollars in OpenAI in 2019. This was not at all an obvious bet back then. Reinforcement learning, what was considered the most promising avenue to general intelligence, had plateaued.2 That year, OpenAI released GPT-2, and I remember thinking “wow, this is dogshit” when I walked into BAIR that day and someone told me to try it.
What country other than America would have financed such a crazy venture back in 2019? None of them did, and none of them would have, because none of them have massive excess capital from monopolistic profits like America does.
Which German executive has $10b from internet advertising at near 100% margins to throw at autonomous vehicles or augmented reality? Which French VC feels confident betting on the biggest, boldest startups because they previously successively funded AOL then eBay then Uber? Money and conviction are the lifeblood of a startup, and both are abundant in America.
Human capital is just as important as financial capital. Because the biggest companies and most cutting-edge startups are in America, American engineers get to work on the biggest, hardest problems. So when a new large-scale, hard problem pops up, it’s no surprise those getting a head-start are somewhere in SF/NYC/Seattle.
When Waymo needed someone to work on ML-based trajectory ranking for its autonomous vehicles, they just took someone from the ML search ranking team at Google. When Anthropic needs someone to optimize inference workloads, there’s someone down the valley who pushed code to the actual CUDA framework at Nvidia.
When the talent and capital flywheel started spinning, the Internet was a much simpler place. You could make billions putting radio on the internet. It was like having training wheels on for engineers, executives, and investors. Those conditions will never be naturally replicated again.
China: the counter-example
When the Communist Party first instituted the Great Firewall in 1998 as a measure to suppress free speech, I doubt they were thinking about monopolies, network effects, and flywheels. Yet, I argue it has inadvertently been the single most effective industrial policy ever. They essentially put on the training wheels for their own tech ecosystem, made sure that the inevitable monopolies were domestic, and put them in pole position to expand into emerging technologies. This industrial policy has led a lower-middle income country to have comparable technological talent, capital, and output as 20+ wealthy liberal democracies combined.
If you have spent any time in China and have had to navigate their ecosystem, the effects of the firewall are glaring. Take, for example, Chinese navigation maps like Amap or Gaode Map. You immediately notice how divergent and baffling their UI/UX patterns are. That might lead you to gloss over how baffling it is they exist at all! Their would-be European and CANZUK equivalents were snuffed out by Google/Apple maps.
There are other explanations for China’s outperformance, and I find them unconvincing. I’ll briefly cover them.
“They have much more talent than we do!”
As a %, China has fewer high-school graduates than Mexico. Even today, only ~50% of Chinese middle-school students test into a high school. Their high school system is way too difficult relative to the content, and their college system is way too lax. Despite the population of 1.4b, the talent density is so low that absolute talent is probably comparable to a (hypothetical) wealthy liberal democracy of 140m people.
“European/Canadian/British/… regulation is too onerous!”
Okay. Every public Chinese company needs to have a “Party Cell” that studies like Xi Jinping’s ramblings or whatever. They put Jack Ma under house arrest for mild criticism. Half of the edtech sector was nuked by Xi Jinping on a whim. Is GDPR as bad as all that?
“They have a massive unified market!”
I concede this, but want to raise some points here. First, the EU market is economically larger on both a per capita and absolute basis. Second, despite the efforts of the Communist Party, China is still very culturally and linguistically diverse. No longer to the degree of Europe, but certainly not as uniform as America. Hundreds of millions still do not speak standard Mandarin, and a similar number who are barely literate in written Chinese. That is to say - you cannot think of the Chinese market as your country’s market scaled to 1.4 billion people. The process of nation-building (or some would say empire-consolidating!) is not yet complete.
Building serious tech in China is just as hard as any place in the free world, except in one big way: your startup will not be suffocated in the cradle by 👊🇺🇸🔥.
They don’t break out what exactly they include as this “AI business”, but presumably it’s dominated by selling compute on Azure and the the various Copilots.
Interestingly, the SOTA in RL for control has still not advanced that much, but using RL for post-training LLMs has been incredibly effective.