OpenAI dropped the world’s first-ever ‘reasoning’ AI chatbot back in September, called ‘o1’. The Silicon Valley-based pioneer spent vast amounts of cash and time to release this AI bot that can ‘reason’ its way through more complex problems. And notwithstanding OpenAI’s name and founding philosophy, it again kept its coding secret.
But last week, China-based AI startup DeepSeek dropped its own rival called ‘R1’, which apparently took only 60 days to build, less than $6M to train (<10% what US firms spend), and now costs ~98% less than US rivals to run. Oh, and the entire code is now free online.
That’s all rattled not only Silicon Valley, but also Wall St and DC. So let’s take a quick tour:
Stay on top of your world from inside your inbox.
Subscribe for free today and receive way much more insights.
Trusted by 122,000+ subscribers
No spam. No noise. Unsubscribe any time.
- Why Silicon Valley is rattled
First, DeepSeek basically just released a legit Ferrari at the price of a Daihatsu — building off existing US open-source models, it used sweet software tricks to radically reduce AI’s reliance on expensive hardware (chips). So, who’s gonna pay for a US-based OpenAI Ferrari now there’s a China-based equivalent for a tiny fraction of the cost?
Second, there are now real questions around Silicon Valley’s moat strategy — i.e. relying on sheer cost and complexity to keep your competitors at bay. Sure, you can try to keep the ‘how‘ a secret, but once you announce the ‘what‘ (AI reasoning, in this case), the DeepSeek story suggests your moat might only last days. That might mean…
Third, by lowering the cost barrier to entry, DeepSeek has now opened the field for countless smaller firms to start driving more AI-based disruption across more industries, and at a faster pace. That potentially means the Valley’s reward leans less on any moat, and more on the ‘final mile’ of connecting the underlying AI to end users.
But fourth, this also raises the prospect of a China-based firm now setting the standards for how AI progresses — DeepSeek’s open-source model is twice as big as Meta’s, potentially entrenching its role in AI’s future by making it twice as attractive for developers. And that takes us to…
- Why Wall St is rattled
This was all made possible because DeepSeek figured out how to weaken the link between AI progress and AI chips. And yet right now, America’s ‘magnificent 7’ (Google, Amazon, Meta, Microsoft, Apple, Tesla, NVIDIA) are variously investing staggering amounts of cash in those very same world-leading AI chips from Nvidia.
Considering these ‘Mag 7’ now make up almost a third of the entire S&P 500 market cap, a chunk of the US market now arguably rests on a big bet on Nvidia chips. And yet by reducing AI’s reliance on those same chips, DeepSeek might‘ve just popped a bubble.
Now, Jevons paradox suggests that as tech efficiency improves and costs collapse, usage actually goes up — ie, longer term, maybe DeepSeek merely just delayed rather than destroyed part of America’s AI moat. But it only takes a flash of uncertainty to spook markets, who’ve already shaved 5% off Nvidia’s share price in just the fast few hours.
And then there’s…
- Why DC is rattled
First, maybe US export controls simply don’t work — when Biden first started tightening US rules in 2022, DeepSeek didn’t even exist. Now, even with shrinking access to the top US tech, DeepSeek has still closed an AI gap in days.
Second, even factoring in DeepSeek’s pre-existing stash of chips from its CEO’s quant hedge fund, he’s been open (🇨🇳) about the fact that the US cutting off China’s access to chips just motivated him to cut China’s dependence on them.
Third, this might also prompt a DC rethink around the role of tech monopolies — a top US official long argued they harm US competitiveness by reducing innovation, but that puts Trump in a bind: truly helping his Silicon Valley allies might mean cutting them loose.
And finally, this will all revive US fears about China-based tech, a la TikTok: ask DeepSeek about Taiwan and you’ll get Communist Party propaganda. Ask it about Xi Jinping, and it’ll refuse to even acknowledge his existence. You can even watch the bot type out answers, then suddenly hit delete once it realises it’s crossed a Party line.
Sure, someone will eventually just tweak DeepSeek’s open-source code to ditch this kind of censorship. But it’s still a reminder of what’s ultimately at stake here: AI is arguably a race to shape what humans see, think, feel, and believe.
INTRIGUE’S TAKE
There are so many ways to think about this. For example, the sheer shock has spawned theories, like maybe DeepSeek has secretly and illegally used a huge stash of America’s top Nvidia chips. But the beauty of open-source coding is that an independent lab can now just check DeepSeek’s low-chip claims itself, though most US competitors are now too busy adopting DeepSeek’s advancements.
Another way to think about this story is as a modern Sputnik moment for the US — ie, a shock to its complacency around any technological, economic, or military edge, while also highlighting China’s perceived advantages (huge datasets, state support, and an open-source tech culture). So maybe the US will now just redouble its own AI efforts.
But here’s a third way to think about this: DeepSeek licensed its impressive gains via America’s MIT, published its findings on US-based academic platforms, and did so in English, after building on US open-source coding, which all still sits atop a hardware tripod that’s still dominated by the US and its partners: chip design (US), manufacturing (Taiwan), and extreme ultraviolet lithography (the Netherlands).
Anyway, we’d love to hear your thoughts in today’s poll below.
Also worth noting:
- DeepSeek’s CEO reportedly just told China’s Premier Li that US chip controls are still a bottleneck for China’s AI progress.
- Meanwhile, Microsoft’s CEO just quoted the Jevons paradox in justifying his big bet on new AI chips, while Meta’s Mark Zuckerberg has now promised a forthcoming version of his own open-source AI model will soon retake the lead.
- Ditto, OpenAI’s Sam Altman has promised his own new version in the coming weeks, after tweeting that, “It is (relatively) easy to copy something that you know works. It is extremely hard to do something new, risky, and difficult when you don’t know if it will work.”