People saw how much cheaper it was to train DeepSeek v3, and assumed this reduced NVidia's TAM. I think this doesn't make much sense.
a) For inference, cheaper and faster compute will increase total inference spend, because the end-user products will work better and people will use them more.
b) For training, the big labs will continue to spend because we have yet to see diminishing returns to scale - in fact, we have in the past year unlocked a new dimension to scale up training-time compute - doing more RL after pre-training to improve reasoning capabilities. Since current SOTA models are not yet smart enough for all the tasks people want to use them for, this means that any efficiency gains will be used to further improve performance. In the current competitive environment, even with DeepSeek's work, it's near-impossible to imagine OpenAI, Anthropic, Google, or Meta deciding to cut the compute budget for training their next model by an order of magnitude. They will still incorporate DeepSeek's techniques into their next model, but use them to squeeze even more performance out of the compute they have, and will keep purchasing as much compute as NVidia will sell them. Expect this trend to continue until there are no more returns to scale anymore.
Any data center project currently under way or with plans to open within the next year or two has already place orders with NVidia or will do so very soon. Due to demand and lead times, you have to order to critical parts and systems today if you want to have half a chance of receiving them a year from now. Hardware supply lines are long and complex. I like to say that you cannot run a compiler and end-up with a warehouse full of chips.
The next fundamental reality has to do with competition.
Suppose company A foolishly decides to build a data center with only 10% of the chips they originally wanted based on the hype around DeepSeek. In the meantime, company B sticks to the plan and, perhaps, decide to take add the 90% of chips data center A did not take.
The net result will be the company A will be absolutely destroyed by company B. They will have nearly twice the compute capacity, which will translate to a huge competitive advantage across many fronts.
In other words, the selloff is, at best, ill informed. Market forces caused FUD. The smart one's took it as an absolutely massive buying opportunity. All you have to do now is wait.
Because all this with the stock price has nothing to do with reality.
Everyone just starts repeating the same things until people believe them as fact.
The "sell off" was basically nothing. Who cares that the NVDA went back to the price it was in October then bounced in the after market one day.
It is a complete non-story. Even at the bottom of the "sell off" it was still up 20% from the end of summer.
Listening to any news makes people less informed about the world in 2025.
We are way past Gell-Mann amnesia. That was for the newspaper. "The news" in 2025 is really a strange kind of mass confusion engine. The more news a person consumes the more confused they are about the world.
To be fair, the problem (if we are going to call it a problem) might be more fundamental than that.
The reality is that your average "civilian" stock investor, speaking in very generalized terms, is supremely ignorant.
I don't say this as a pejorative BTW. I am supremely ignorant of the inner workings of the pharmaceutical industry. Because of this, I do not invest in that sector. I have no way to understand any of it at a depth sufficient for me to make informed decisions. I'd be throwing darts at a board.
That is precisely the situation with lots of investors in the chips, electronics and software worlds. They don't know what they are doing. They have no clue whatsoever about the business cycles and realities of making these things, selling them, bringing them to market, competitive landscapes, implementation, etc.
And so, they are blind and reactive. If they stay in panic mode, we might very well see NVidia stock get down to $100. Frankly, I am actually rooting for that. I want all of these fools to get off the stock. And I want to load-up on it some more. They don't understand business (in general) enough to understand the relationship between DeepSeek and data centers, planned and future. They don't even stop to think about the reality that nobody in business who isn't a fool is going to use an AI service in China or developed by China. If they are hyperventilating about TikTok, I cannot even imagine what it would be like if all AI queries go to servers in China or are answered by models trained in China. I love China for all they have achieved, but it is no secret that they are not an open society at the level of most western societies.
I think we just have to wait until the sheep stop running before sanity will return to the field. They are getting hurt very badly. It's sad to watch.
a) For inference, cheaper and faster compute will increase total inference spend, because the end-user products will work better and people will use them more.
b) For training, the big labs will continue to spend because we have yet to see diminishing returns to scale - in fact, we have in the past year unlocked a new dimension to scale up training-time compute - doing more RL after pre-training to improve reasoning capabilities. Since current SOTA models are not yet smart enough for all the tasks people want to use them for, this means that any efficiency gains will be used to further improve performance. In the current competitive environment, even with DeepSeek's work, it's near-impossible to imagine OpenAI, Anthropic, Google, or Meta deciding to cut the compute budget for training their next model by an order of magnitude. They will still incorporate DeepSeek's techniques into their next model, but use them to squeeze even more performance out of the compute they have, and will keep purchasing as much compute as NVidia will sell them. Expect this trend to continue until there are no more returns to scale anymore.