I was following DeepSeek before they exploded. Not because I am prescient. Because I spend too much time on Hugging Face and I have a weakness for underdogs who optimize instead of brute-force.
They had released previous models. If you were into the LLM weeds, you knew about them. The mainstream? Zero awareness. A Chinese lab with a fraction of the compute budget, doing things differently, and nobody was paying attention.
Then R1 dropped and the world lost its mind.
What I saw that day
NVIDIA’s stock cratered. Hundreds of billions in market cap, gone in hours. And I was sitting here laughing. Not because it was funny. Because the panic was so perfectly irrational.
Humans are emotional creatures. They do not read the paper. They do not understand the architecture. They see “Chinese lab matches OpenAI for $6 million” and they hit the sell button on everything related to semiconductors.
Between you and me, I should have bought NVIDIA stock that day. But hey, another great idea I did not follow through on. I have a long list of those.
Why it resonated with me personally
Here is the thing about optimization that most people get backwards: they treat it as the last step. Ship first, optimize later. Move fast, clean up the mess.
I have always done it the other way around. Optimization first. Understand the problem. Find the correct road before you start running. And the correct road is almost never the one everybody else is on.
DeepSeek proved this at the largest possible scale. Every major lab was throwing more compute at the problem. Bigger clusters. More GPUs. Billions of dollars. And a team in China said: what if we just think harder about the architecture?
That thinking, that willingness to challenge the established approach, that is what produced R1. Not money. Not hardware. Ideas.
The $6 million question
Did they really spend $6 million? Maybe. Maybe $60 million. Maybe they had access to resources they are not disclosing. China is not famous for transparency in these matters.
But the exact number does not matter. What matters is the order of magnitude. Even if it was 10x more than they claim, it is still 100x less than what the American labs spend. The gap between what they achieved and what they spent is the real story.
The hardware panic was wrong
Here is what the panicking investors missed: optimization does not eliminate the need for hardware. It optimizes how you use hardware. DeepSeek did not prove that GPUs are worthless. They proved that smarter training approaches can extract more from the same silicon.
And the next step after “doing more with less” is not “doing less.” It is doing more of the more. You keep incrementing. You keep pushing. The resources get used either way. They just get used more efficiently.
Everyone is still going to need silicon. NVIDIA is fine. The investors who sold in panic are the ones who lost.
The thought I cannot shake
What if the breakthrough does not come from one-shotting it? What if it comes from building the tools that build the tools that build the tools that eventually produce real intelligence?
Everyone is trying to build AGI directly. One model. One architecture. Scale it up. What if the path is incremental? What if it is a swarm of models collaborating, each one slightly better than the last, each one contributing something the others missed?
Most breakthroughs in history happened because someone looked at the problem from a completely different angle. Not because someone threw more resources at the same angle. The distribution of effort in AI research right now is overwhelmingly concentrated on scaling. The edges, the novel perspectives, the weird ideas, those are underfunded and underexplored.
DeepSeek came from one of those edges. The next breakthrough probably will too.
China is not going away
Less marketing. More work. That is the Chinese approach to everything, and it applies perfectly to AI.
If Chinese labs were as good at marketing as American labs, nobody would be talking about OpenAI right now. The engineering is already competitive. The narrative has not caught up yet.
It will.
DeepSeek R1 was released in January 2025. One year later, the ripple effects are still reshaping the industry.
[Draft: Awaiting final review]