A couple of days ago, Nvidia’s stock price took a stumble on the other side of the ocean. While the reasons are complex, I prefer to believe this was the capital market experiencing a “visceral shiver” in response to DeepSeek R1.
The situation is actually quite surreal. An AI laboratory not born of Silicon Valley released an open-source model that not only backed OpenAI’s “ace card” o1 into a corner regarding math and coding reasoning, but even more absurdly, its API output price is about 3% of o1’s (you read that right—it’s a fraction of a fraction).
If OpenAI is building AI with a rocket-launching budget, DeepSeek is like someone building a nuclear reactor in a garage using Legos. This isn’t just a victory for open source; it is practically a “dimensional strike” against “compute hegemony.”
Today, let’s skip the dry academic parameters and dig into how this “high-IQ wall,” guarded for three years, was punctured.
1. The End of Brute Force Aesthetics
This isn’t imitation; it’s changing lanes to overtake.
For the past three years, a term has been trending in the LLM circle: “Scaling Law.” The subtext is: no matter how smart you are, if you don’t have the money to buy tens of thousands of GPUs, you have to kneel. But R1 has slapped this logic hard in the face.
Its core breakthrough lies in DeepSeek-R1-Zero—a prodigy trained almost entirely without human annotated data (SFT), relying purely on “Reinforcement Learning” (RL) to spar with itself. It’s like locking a child who has never attended school in a room, giving them only math problems, giving candy for right answers and punishment for wrong ones, and eventually, they derive a set of problem-solving methods on their own.
The subversive nature of this is: It proves that advanced logical reasoning capabilities do not require feeding on massive amounts of human “standard answers”; machines can emerge with these skills through self-reflection (Chain of Thought).
This reminds me of a saying: “Truth stands within the range of a cannon.” But now, R1 tells us that sometimes, a sniper rifle works better than a cannon.
2. Distillation: Putting “Einstein” in a Smartphone
If the R1 large model is “flexing muscles,” then the Distilled models released alongside it are the true “poison” keeping competitors awake at night.
You might not have noticed, but R1-level reasoning capabilities have been successfully “condensed” into a 32B or even smaller model. What does this mean? It means that the gaming PC under your desk, or the flagship mobile phone coming soon, can run an AI with logical capabilities comparable to a PhD student.
This chart is evidence of “insubordination”: small parameter models, after distillation, are wiping the floor with many heavyweights on math benchmarks.
This shatters the myth that “intelligence is a privilege of the cloud nobility.” DeepSeek has dragged the technical route from “renting expensive cloud brains” straight toward “giving every terminal independent thinking capabilities.”
It reminds me of how MP3s killed CD players. When high-quality intelligence can be replicated at extremely low cost and run locally, shouldn’t those business models trying to “collect tolls” via API metering start recalculating their books?
3. When Price Wars Become “Dimensional Strikes”
Let’s do some vulgar accounting.
OpenAI’s o1 model costs tens of dollars to output one million tokens; this price destines it to be a “tool for aristocrats.” DeepSeek R1, however, has slashed the price through the floorboards—approximately $2.19 / 1M tokens.
| Model | Input Cost (Per 1M tokens) | Output Cost (Per 1M tokens) |
|---|---|---|
| DeepSeek R1 | $0.55 | $2.19 |
| OpenAI o1 | $15.00 | $60.00 |
This isn’t competition in the same dimension; this is a massacre.
It’s as if you are still selling gold bars by the ounce, and the guy next door suddenly invents a “philosopher’s stone” and starts selling gold by the ton. For developers, this cost difference means: full-process AI thinking and ultra-long text code refactoring, which we didn’t dare do before, can now be run with our eyes closed.
Some even jokingly call R1 the “Temu” of the AI world, but this metaphor is inappropriate because its quality benchmarks against “Hermès.” This combination of “top-tier performance + cabbage prices” has kicked AI from the “exploration phase” directly into the “adoption phase.”
4. Do We Really Need That Many GPUs?
This also leads me to a dangerous thought: If we no longer need such massive parameter counts to achieve top-tier reasoning capabilities, is the current GPU bubble a bit… big?
I mean, if the future direction is “small and precise” reasoning models coupled with efficient distillation technology, will the current frantic arms race of hoarding compute power suddenly hit the brakes one day?
Of course, training still requires compute, but the structure of compute demand on the inference side might undergo drastic changes. This might be why Wall Street is so sensitive to R1—it shakes the underlying narrative that “compute will always be in short supply.”
5. Final Thoughts
I particularly like the candid geek vibe in DeepSeek’s technical report—no flashy launch event, no Steve Jobs-style speeches, just lines of code and that hardcore paper.
It reminds me of the early internet spirit: Open, Shared, Breaking Monopolies.
DeepSeek R1’s appearance may not directly kill OpenAI, but it has injected a massive shot of adrenaline into this slightly dull, class-solidified AI circle. It tells all developers: Don’t worship myths; myths are meant to be broken.
At this moment, watching R1 spit out its thinking process line by line in the terminal window, I feel like I can hear the sound of the old world’s walls cracking.
References：
