The attached video is a Hashed Vibe Labs promotional video I made for fun with Seedance 2.0. It came out in one try. Character consistency, camera work, background music, and even keyboard sound effects—all of it was added by AI on its own from just a few lines of prompts. The video is not only better than Kling 3.0; the quality of its music and sound effects is in a different league. Many existing AI video tools were gacha machines. You had to keep running them until something usable came out, and you threw away eight out of ten results. Most of what Seedance 2.0 produces is quite usable. But the real shock is that it costs far less than half of what American video models cost.
Musk even acknowledged that "it's happening fast," while Beijing Daily declared, "From DeepSeek to Seedance, Chinese AI has done it." Right after the launch of Seedance 2.0, the combined market capitalization of America's big tech companies fell by more than 1,300 trillion won. If DeepSeek created a Sputnik moment in text AI a year ago, Seedance 2.0 is creating the same shockwave in video.
The same pattern is playing out with language models. As soon as Moonshot AI's Kimi K2.5 launched, it took a commanding first place on OpenRouter in token usage. When connected to OpenClaw, it costs less than one-hundredth the price of Claude Opus 4.5. With the same budget, while an American model can make ten attempts, a Chinese model can make fifty or more than a hundred. If agents ultimately become smarter through trial and error, it is obvious who has the advantage. When the RAND Corporation analyzed the operating cost of Chinese models as one-fourth to one-sixth that of American models, it was not simply saying that they are cheap. It was saying that you can build smarter agents with the same amount of money.
When model performance becomes similar and one side falls behind on price, the final battleground is scaling. The lifeline of scaling is electricity. What Musk said in Davos gets to the heart of it: "The bottleneck for AI is ultimately electricity. The US is short of it, and China is the exception." Since 2021, the generation capacity China has installed exceeds the total cumulative amount the US has built since its founding. Over the next five years, China is expected to add six times as much as the US. US power demand barely grew for 20 years, then suddenly surged because of AI, while building a single power plant takes years. Data centers in Texas alone have requested power on the scale of dozens of gigawatts, but only a little over 1 gigawatt was approved over the past year. Musk's move to merge SpaceX and xAI, saying that "the cheapest place to run AI within three years will be space," is, looked at another way, an admission that the fight on the ground is that blocked. One country says, "Let's go to space" for a solution, while another is laying solar panels across the land at a furious pace every year. The direction is clear just from looking at this picture.
Model performance, token costs, and electricity. China has secured two of the three pillars of AI competition and is rapidly closing the gap on the remaining one. What the US has left is its ecosystem. Trust with enterprise customers, regulation, and the existing order that stretches from Hollywood to Wall Street. But an ecosystem is ultimately a device for buying time. If the moat is eroding faster than it takes to cross it, it is not a moat. When DeepSeek appeared a year ago, the market absorbed it as a temporary shock. A year has passed. The shock has become structural. Price has already become China's weapon, and time is increasingly becoming America's enemy.
[Ad] Applications for Hashed Vibe Labs Seoul Edition are open until February 19.
