A Chinese Model Erased 589 Billion Dollars of Nvidia in One Session
On January 27 a competitor claimed frontier level performance at a fraction of the training cost, and the largest single day market value loss in history followed.
The Session
On Monday January 27, 2025, Nvidia shares fell roughly 17 percent, erasing approximately 589 billion dollars of market capitalization. It was the largest single day loss of market value by any company in history.
The trigger was a Chinese startup releasing a reasoning model that performed comparably to leading Western systems. The company reported that one of its models had been trained for around 5.6 million dollars, against figures above 100 million dollars commonly associated with frontier training runs.
Why That Number Threatened the Trade
The investment case for Nvidia rested on a chain of assumptions. Artificial intelligence capability requires enormous compute. Compute requires accelerators. Nvidia sells the accelerators. Therefore capability improvements require continuously escalating hardware spending.
A credible demonstration that comparable capability could be achieved for a small fraction of the assumed cost attacked the first link. If capability scales with algorithmic efficiency rather than only with spending, the projected demand curve for hardware flattens.
The market was not repricing Nvidia's earnings. It was repricing the assumption that intelligence requires unlimited hardware, which is the assumption the entire sector was capitalized on.
The Counterarguments Worth Taking Seriously
Several objections surfaced immediately and they were not merely defensive. The reported training cost likely covered a final training run rather than total investment including research, failed experiments, and infrastructure. Comparing a final run to another company's all in figure is not comparing like with like.
A second argument invoked a well known economic pattern. When a resource becomes cheaper to use, total consumption frequently rises rather than falls, because cheaper access opens applications that were previously uneconomic. On that view, more efficient models expand the number of deployments and inference demand grows even as training becomes cheaper per unit of capability.
Both points have merit, and the following months saw substantial recovery in the affected names, which suggests the market eventually gave them weight.
Training Versus Inference
The distinction that clarified the debate is between training and inference. Training is the one time, enormously expensive process of building a model. Inference is running the finished model to answer queries, which is far cheaper per instance and happens continuously at scale.
Efficiency gains in training reduce one category of demand. If cheaper models drive far wider deployment, inference demand grows, and inference is the larger long run market. Whether the shift is net negative for hardware depends on which effect dominates, and that was genuinely unknown in January 2025.
What the Episode Demonstrated
The durable lesson is about concentration. A single company had grown large enough that a research paper from an unfamiliar competitor could remove more market value in one session than most companies are worth in total.
It also showed how narratives function as load bearing structures in valuation. The position was priced on a story about how capability scales. When a credible alternative story appeared, the repricing was instantaneous, because nothing in the fundamentals had changed that day. Only the assumption had.
The Bottom Line
The largest one day loss in market history was caused by a challenge to an assumption rather than by any change in results. When a valuation rests on a narrative, the narrative is the position you actually hold.