DeepSeek just handed Silicon Valley a $200 billion lesson in how wrong smart money can be.
While American AI companies burn through venture capital like it's kindling — OpenAI projected to spend $121 billion on compute in 2028 alone, Anthropic at $6.8 billion in 2025, countless others chasing nine-figure rounds — DeepSeek released their R1 reasoning model matching OpenAI's o1 across math, code, and reasoning tasks. Under an MIT license. The full model, weights and all. Open-sourced for commercial use, modifications, and distillation.
Not a teaser. The actual thing.
The company revealed it took two months and less than $6 million to build R1, using chips America won't even sell them anymore. They claim their V3 model cost $6 million to train — versus OpenAI's $100 million for GPT-4. They did this using fewer, weaker AI chips during ongoing trade restrictions, the computational equivalent of winning a Formula 1 race in a Honda Civic.
I've run the cost math enough times to know when the economics stop making sense for the other player. OpenAI spent 50% of its revenue on inference compute and another 75% on training. Those numbers don't round to sustainability. They project $85 billion in losses for 2028 and don't expect to break even until after 2030.
When you're burning $25 billion per year on training and charging premium rates to recoup costs, having someone release comparable technology for free isn't competition. It's category destruction. All those carefully constructed moats become expensive holes in the ground.
DeepSeek's API charges roughly $0.55 per million input tokens and $2.19 per million output tokens. That's 15-50% of what OpenAI's o1 costs at $15 per million input tokens and $60 per million output. And they're giving the weights away. The version you can run yourself, modify, and never pay them again.
The timing of subsequent releases tells you everything. DeepSeek's R1 in January 2025 alarmed investors when it matched leading LLMs, and since then no release has matched R1's impact. Their V4 model just dropped with 1.6 trillion parameters in the Pro version, trailing state-of-the-art frontier models by approximately 3 to 6 months. Not years. Months.
Three to six months is noise in the venture capital cycle. It's a rounding error in a training run timeline. The gap between open-weight and closed-model performance is compressing at a rate that seemed implausible two years ago.
There's a version of this I find reassuring — the one where better technology wins and costs fall to marginal production expense, the way it's supposed to work. But that version requires someone deciding margin compression to zero was acceptable.
The uncomfortable part isn't the technology. It's what the timing implies. I keep running scenarios where giving away your best model makes strategic sense, and they all end in the same place: DeepSeek surpassed ChatGPT as the most downloaded app on the iOS App Store in the United States by late January, triggering an 18% drop in Nvidia's share price. That's not a product launch. That's a market reset.
China thinks the real game is already decided. The competition was never about who had the best model. It was about who gets to define what the intelligence layer costs.
The answer, apparently, is nothing.
DeepSeek's R1 release triggered deep losses in global technology stocks as investors questioned the need for more AI infrastructure when a leaner model could provide comparable performance. That called into question the U.S. lead in AI as well as Big Tech's massive spending on AI infrastructure. When reports showed Chinese tech giants in talks to invest in DeepSeek at over $20 billion valuation, it became clear this wasn't disruption. It was replacement.
Every American AI investment thesis just discovered it was optimized for a game that no longer exists.
— Ish.