Nvidia says DeepSeek just did the easy part

Here’s a statement from Nvidia on the emergent Chinese competitor that catalyzed the single biggest case of value destruction in recorded history:

“DeepSeek is an excellent AI advancement and a perfect example of Test Time Scaling. DeepSeek’s work illustrates how new models can be created using that technique, leveraging widely-available models and compute that is fully export control compliant. Inference requires significant numbers of NVIDIA GPUs and high-performance networking. We now have three scaling laws: pre-training and post-training, which continue, and new test-time scaling.”

Basically, Nvidia’s framing is that DeepSeek did the easy part, using training wheels. But if you want to ride in the Tour de France, you need to pick up what Jensen Huang’s putting down.

Training costs (teaching the model how to do things) are lower than inference costs (having the model do things on an ongoing basis). To make a comparison, it’s more labor-intensive to construct buildings brick by brick, day after day, than it is to draw up the blueprints.

If DeepSeek’s outage today was merely the intersection of surging demand and limited supply, rather than the product of a cyberattack as was claimed, that would likely corroborate the broad thrust of Nvidia’s appeal to investors: remember inference costs!